xref: /freebsd/sys/kern/vfs_vnops.c (revision c6989859ae9388eeb46a24fe88f9b8d07101c710)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/sx.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/sysctl.h>
76 #include <sys/ttycom.h>
77 #include <sys/conf.h>
78 #include <sys/syslog.h>
79 #include <sys/unistd.h>
80 #include <sys/user.h>
81 
82 #include <security/audit/audit.h>
83 #include <security/mac/mac_framework.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_extern.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pager.h>
92 
93 #ifdef HWPMC_HOOKS
94 #include <sys/pmckern.h>
95 #endif
96 
97 static fo_rdwr_t	vn_read;
98 static fo_rdwr_t	vn_write;
99 static fo_rdwr_t	vn_io_fault;
100 static fo_truncate_t	vn_truncate;
101 static fo_ioctl_t	vn_ioctl;
102 static fo_poll_t	vn_poll;
103 static fo_kqfilter_t	vn_kqfilter;
104 static fo_stat_t	vn_statfile;
105 static fo_close_t	vn_closefile;
106 static fo_mmap_t	vn_mmap;
107 static fo_fallocate_t	vn_fallocate;
108 
109 struct 	fileops vnops = {
110 	.fo_read = vn_io_fault,
111 	.fo_write = vn_io_fault,
112 	.fo_truncate = vn_truncate,
113 	.fo_ioctl = vn_ioctl,
114 	.fo_poll = vn_poll,
115 	.fo_kqfilter = vn_kqfilter,
116 	.fo_stat = vn_statfile,
117 	.fo_close = vn_closefile,
118 	.fo_chmod = vn_chmod,
119 	.fo_chown = vn_chown,
120 	.fo_sendfile = vn_sendfile,
121 	.fo_seek = vn_seek,
122 	.fo_fill_kinfo = vn_fill_kinfo,
123 	.fo_mmap = vn_mmap,
124 	.fo_fallocate = vn_fallocate,
125 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
126 };
127 
128 const u_int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
131     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
132 static int vn_io_fault_prefault = 0;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
134     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static int vn_io_pgcache_read_enable = 1;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
137     &vn_io_pgcache_read_enable, 0,
138     "Enable copying from page cache for reads, avoiding fs");
139 static u_long vn_io_faults_cnt;
140 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
141     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
142 
143 static int vfs_allow_read_dir = 0;
144 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
145     &vfs_allow_read_dir, 0,
146     "Enable read(2) of directory by root for filesystems that support it");
147 
148 /*
149  * Returns true if vn_io_fault mode of handling the i/o request should
150  * be used.
151  */
152 static bool
153 do_vn_io_fault(struct vnode *vp, struct uio *uio)
154 {
155 	struct mount *mp;
156 
157 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
158 	    (mp = vp->v_mount) != NULL &&
159 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
160 }
161 
162 /*
163  * Structure used to pass arguments to vn_io_fault1(), to do either
164  * file- or vnode-based I/O calls.
165  */
166 struct vn_io_fault_args {
167 	enum {
168 		VN_IO_FAULT_FOP,
169 		VN_IO_FAULT_VOP
170 	} kind;
171 	struct ucred *cred;
172 	int flags;
173 	union {
174 		struct fop_args_tag {
175 			struct file *fp;
176 			fo_rdwr_t *doio;
177 		} fop_args;
178 		struct vop_args_tag {
179 			struct vnode *vp;
180 		} vop_args;
181 	} args;
182 };
183 
184 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
185     struct vn_io_fault_args *args, struct thread *td);
186 
187 int
188 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
189 {
190 	struct thread *td = ndp->ni_cnd.cn_thread;
191 
192 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
193 }
194 
195 /*
196  * Common code for vnode open operations via a name lookup.
197  * Lookup the vnode and invoke VOP_CREATE if needed.
198  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
199  *
200  * Note that this does NOT free nameidata for the successful case,
201  * due to the NDINIT being done elsewhere.
202  */
203 int
204 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
205     struct ucred *cred, struct file *fp)
206 {
207 	struct vnode *vp;
208 	struct mount *mp;
209 	struct thread *td = ndp->ni_cnd.cn_thread;
210 	struct vattr vat;
211 	struct vattr *vap = &vat;
212 	int fmode, error;
213 
214 restart:
215 	fmode = *flagp;
216 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
217 	    O_EXCL | O_DIRECTORY))
218 		return (EINVAL);
219 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
220 		ndp->ni_cnd.cn_nameiop = CREATE;
221 		/*
222 		 * Set NOCACHE to avoid flushing the cache when
223 		 * rolling in many files at once.
224 		*/
225 		ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
226 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
227 			ndp->ni_cnd.cn_flags |= FOLLOW;
228 		if ((fmode & O_BENEATH) != 0)
229 			ndp->ni_cnd.cn_flags |= BENEATH;
230 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
231 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
232 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
233 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
234 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
235 			bwillwrite();
236 		if ((error = namei(ndp)) != 0)
237 			return (error);
238 		if (ndp->ni_vp == NULL) {
239 			VATTR_NULL(vap);
240 			vap->va_type = VREG;
241 			vap->va_mode = cmode;
242 			if (fmode & O_EXCL)
243 				vap->va_vaflags |= VA_EXCLUSIVE;
244 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
245 				NDFREE(ndp, NDF_ONLY_PNBUF);
246 				vput(ndp->ni_dvp);
247 				if ((error = vn_start_write(NULL, &mp,
248 				    V_XSLEEP | PCATCH)) != 0)
249 					return (error);
250 				goto restart;
251 			}
252 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
253 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
254 #ifdef MAC
255 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
256 			    &ndp->ni_cnd, vap);
257 			if (error == 0)
258 #endif
259 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
260 						   &ndp->ni_cnd, vap);
261 			vput(ndp->ni_dvp);
262 			vn_finished_write(mp);
263 			if (error) {
264 				NDFREE(ndp, NDF_ONLY_PNBUF);
265 				return (error);
266 			}
267 			fmode &= ~O_TRUNC;
268 			vp = ndp->ni_vp;
269 		} else {
270 			if (ndp->ni_dvp == ndp->ni_vp)
271 				vrele(ndp->ni_dvp);
272 			else
273 				vput(ndp->ni_dvp);
274 			ndp->ni_dvp = NULL;
275 			vp = ndp->ni_vp;
276 			if (fmode & O_EXCL) {
277 				error = EEXIST;
278 				goto bad;
279 			}
280 			if (vp->v_type == VDIR) {
281 				error = EISDIR;
282 				goto bad;
283 			}
284 			fmode &= ~O_CREAT;
285 		}
286 	} else {
287 		ndp->ni_cnd.cn_nameiop = LOOKUP;
288 		ndp->ni_cnd.cn_flags = ISOPEN |
289 		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
290 		if (!(fmode & FWRITE))
291 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
292 		if ((fmode & O_BENEATH) != 0)
293 			ndp->ni_cnd.cn_flags |= BENEATH;
294 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
295 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
296 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
297 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
298 		if ((error = namei(ndp)) != 0)
299 			return (error);
300 		vp = ndp->ni_vp;
301 	}
302 	error = vn_open_vnode(vp, fmode, cred, td, fp);
303 	if (error)
304 		goto bad;
305 	*flagp = fmode;
306 	return (0);
307 bad:
308 	NDFREE(ndp, NDF_ONLY_PNBUF);
309 	vput(vp);
310 	*flagp = fmode;
311 	ndp->ni_vp = NULL;
312 	return (error);
313 }
314 
315 static int
316 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
317 {
318 	struct flock lf;
319 	int error, lock_flags, type;
320 
321 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
322 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
323 		return (0);
324 	KASSERT(fp != NULL, ("open with flock requires fp"));
325 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
326 		return (EOPNOTSUPP);
327 
328 	lock_flags = VOP_ISLOCKED(vp);
329 	VOP_UNLOCK(vp);
330 
331 	lf.l_whence = SEEK_SET;
332 	lf.l_start = 0;
333 	lf.l_len = 0;
334 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
335 	type = F_FLOCK;
336 	if ((fmode & FNONBLOCK) == 0)
337 		type |= F_WAIT;
338 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
339 	if (error == 0)
340 		fp->f_flag |= FHASLOCK;
341 
342 	vn_lock(vp, lock_flags | LK_RETRY);
343 	if (error == 0 && VN_IS_DOOMED(vp))
344 		error = ENOENT;
345 	return (error);
346 }
347 
348 /*
349  * Common code for vnode open operations once a vnode is located.
350  * Check permissions, and call the VOP_OPEN routine.
351  */
352 int
353 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
354     struct thread *td, struct file *fp)
355 {
356 	accmode_t accmode;
357 	int error;
358 
359 	if (vp->v_type == VLNK)
360 		return (EMLINK);
361 	if (vp->v_type == VSOCK)
362 		return (EOPNOTSUPP);
363 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
364 		return (ENOTDIR);
365 	accmode = 0;
366 	if (fmode & (FWRITE | O_TRUNC)) {
367 		if (vp->v_type == VDIR)
368 			return (EISDIR);
369 		accmode |= VWRITE;
370 	}
371 	if (fmode & FREAD)
372 		accmode |= VREAD;
373 	if (fmode & FEXEC)
374 		accmode |= VEXEC;
375 	if ((fmode & O_APPEND) && (fmode & FWRITE))
376 		accmode |= VAPPEND;
377 #ifdef MAC
378 	if (fmode & O_CREAT)
379 		accmode |= VCREAT;
380 	if (fmode & O_VERIFY)
381 		accmode |= VVERIFY;
382 	error = mac_vnode_check_open(cred, vp, accmode);
383 	if (error)
384 		return (error);
385 
386 	accmode &= ~(VCREAT | VVERIFY);
387 #endif
388 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
389 		error = VOP_ACCESS(vp, accmode, cred, td);
390 		if (error != 0)
391 			return (error);
392 	}
393 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
394 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
395 	error = VOP_OPEN(vp, fmode, cred, td, fp);
396 	if (error != 0)
397 		return (error);
398 
399 	error = vn_open_vnode_advlock(vp, fmode, fp);
400 	if (error == 0 && (fmode & FWRITE) != 0) {
401 		error = VOP_ADD_WRITECOUNT(vp, 1);
402 		if (error == 0) {
403 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
404 			     __func__, vp, vp->v_writecount);
405 		}
406 	}
407 
408 	/*
409 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
410 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
411 	 * Arrange for that by having fdrop() to use vn_closefile().
412 	 */
413 	if (error != 0) {
414 		fp->f_flag |= FOPENFAILED;
415 		fp->f_vnode = vp;
416 		if (fp->f_ops == &badfileops) {
417 			fp->f_type = DTYPE_VNODE;
418 			fp->f_ops = &vnops;
419 		}
420 		vref(vp);
421 	}
422 
423 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
424 	return (error);
425 
426 }
427 
428 /*
429  * Check for write permissions on the specified vnode.
430  * Prototype text segments cannot be written.
431  * It is racy.
432  */
433 int
434 vn_writechk(struct vnode *vp)
435 {
436 
437 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
438 	/*
439 	 * If there's shared text associated with
440 	 * the vnode, try to free it up once.  If
441 	 * we fail, we can't allow writing.
442 	 */
443 	if (VOP_IS_TEXT(vp))
444 		return (ETXTBSY);
445 
446 	return (0);
447 }
448 
449 /*
450  * Vnode close call
451  */
452 static int
453 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
454     struct thread *td, bool keep_ref)
455 {
456 	struct mount *mp;
457 	int error, lock_flags;
458 
459 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
460 	    MNT_EXTENDED_SHARED(vp->v_mount))
461 		lock_flags = LK_SHARED;
462 	else
463 		lock_flags = LK_EXCLUSIVE;
464 
465 	vn_start_write(vp, &mp, V_WAIT);
466 	vn_lock(vp, lock_flags | LK_RETRY);
467 	AUDIT_ARG_VNODE1(vp);
468 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
469 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
470 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
471 		    __func__, vp, vp->v_writecount);
472 	}
473 	error = VOP_CLOSE(vp, flags, file_cred, td);
474 	if (keep_ref)
475 		VOP_UNLOCK(vp);
476 	else
477 		vput(vp);
478 	vn_finished_write(mp);
479 	return (error);
480 }
481 
482 int
483 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
484     struct thread *td)
485 {
486 
487 	return (vn_close1(vp, flags, file_cred, td, false));
488 }
489 
490 /*
491  * Heuristic to detect sequential operation.
492  */
493 static int
494 sequential_heuristic(struct uio *uio, struct file *fp)
495 {
496 	enum uio_rw rw;
497 
498 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
499 
500 	rw = uio->uio_rw;
501 	if (fp->f_flag & FRDAHEAD)
502 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
503 
504 	/*
505 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
506 	 * that the first I/O is normally considered to be slightly
507 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
508 	 * unless previous seeks have reduced f_seqcount to 0, in which
509 	 * case offset 0 is not special.
510 	 */
511 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
512 	    uio->uio_offset == fp->f_nextoff[rw]) {
513 		/*
514 		 * f_seqcount is in units of fixed-size blocks so that it
515 		 * depends mainly on the amount of sequential I/O and not
516 		 * much on the number of sequential I/O's.  The fixed size
517 		 * of 16384 is hard-coded here since it is (not quite) just
518 		 * a magic size that works well here.  This size is more
519 		 * closely related to the best I/O size for real disks than
520 		 * to any block size used by software.
521 		 */
522 		if (uio->uio_resid >= IO_SEQMAX * 16384)
523 			fp->f_seqcount[rw] = IO_SEQMAX;
524 		else {
525 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
526 			if (fp->f_seqcount[rw] > IO_SEQMAX)
527 				fp->f_seqcount[rw] = IO_SEQMAX;
528 		}
529 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
530 	}
531 
532 	/* Not sequential.  Quickly draw-down sequentiality. */
533 	if (fp->f_seqcount[rw] > 1)
534 		fp->f_seqcount[rw] = 1;
535 	else
536 		fp->f_seqcount[rw] = 0;
537 	return (0);
538 }
539 
540 /*
541  * Package up an I/O request on a vnode into a uio and do it.
542  */
543 int
544 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
545     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
546     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
547 {
548 	struct uio auio;
549 	struct iovec aiov;
550 	struct mount *mp;
551 	struct ucred *cred;
552 	void *rl_cookie;
553 	struct vn_io_fault_args args;
554 	int error, lock_flags;
555 
556 	if (offset < 0 && vp->v_type != VCHR)
557 		return (EINVAL);
558 	auio.uio_iov = &aiov;
559 	auio.uio_iovcnt = 1;
560 	aiov.iov_base = base;
561 	aiov.iov_len = len;
562 	auio.uio_resid = len;
563 	auio.uio_offset = offset;
564 	auio.uio_segflg = segflg;
565 	auio.uio_rw = rw;
566 	auio.uio_td = td;
567 	error = 0;
568 
569 	if ((ioflg & IO_NODELOCKED) == 0) {
570 		if ((ioflg & IO_RANGELOCKED) == 0) {
571 			if (rw == UIO_READ) {
572 				rl_cookie = vn_rangelock_rlock(vp, offset,
573 				    offset + len);
574 			} else {
575 				rl_cookie = vn_rangelock_wlock(vp, offset,
576 				    offset + len);
577 			}
578 		} else
579 			rl_cookie = NULL;
580 		mp = NULL;
581 		if (rw == UIO_WRITE) {
582 			if (vp->v_type != VCHR &&
583 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
584 			    != 0)
585 				goto out;
586 			if (MNT_SHARED_WRITES(mp) ||
587 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
588 				lock_flags = LK_SHARED;
589 			else
590 				lock_flags = LK_EXCLUSIVE;
591 		} else
592 			lock_flags = LK_SHARED;
593 		vn_lock(vp, lock_flags | LK_RETRY);
594 	} else
595 		rl_cookie = NULL;
596 
597 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
598 #ifdef MAC
599 	if ((ioflg & IO_NOMACCHECK) == 0) {
600 		if (rw == UIO_READ)
601 			error = mac_vnode_check_read(active_cred, file_cred,
602 			    vp);
603 		else
604 			error = mac_vnode_check_write(active_cred, file_cred,
605 			    vp);
606 	}
607 #endif
608 	if (error == 0) {
609 		if (file_cred != NULL)
610 			cred = file_cred;
611 		else
612 			cred = active_cred;
613 		if (do_vn_io_fault(vp, &auio)) {
614 			args.kind = VN_IO_FAULT_VOP;
615 			args.cred = cred;
616 			args.flags = ioflg;
617 			args.args.vop_args.vp = vp;
618 			error = vn_io_fault1(vp, &auio, &args, td);
619 		} else if (rw == UIO_READ) {
620 			error = VOP_READ(vp, &auio, ioflg, cred);
621 		} else /* if (rw == UIO_WRITE) */ {
622 			error = VOP_WRITE(vp, &auio, ioflg, cred);
623 		}
624 	}
625 	if (aresid)
626 		*aresid = auio.uio_resid;
627 	else
628 		if (auio.uio_resid && error == 0)
629 			error = EIO;
630 	if ((ioflg & IO_NODELOCKED) == 0) {
631 		VOP_UNLOCK(vp);
632 		if (mp != NULL)
633 			vn_finished_write(mp);
634 	}
635  out:
636 	if (rl_cookie != NULL)
637 		vn_rangelock_unlock(vp, rl_cookie);
638 	return (error);
639 }
640 
641 /*
642  * Package up an I/O request on a vnode into a uio and do it.  The I/O
643  * request is split up into smaller chunks and we try to avoid saturating
644  * the buffer cache while potentially holding a vnode locked, so we
645  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
646  * to give other processes a chance to lock the vnode (either other processes
647  * core'ing the same binary, or unrelated processes scanning the directory).
648  */
649 int
650 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
651     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
652     struct ucred *file_cred, size_t *aresid, struct thread *td)
653 {
654 	int error = 0;
655 	ssize_t iaresid;
656 
657 	do {
658 		int chunk;
659 
660 		/*
661 		 * Force `offset' to a multiple of MAXBSIZE except possibly
662 		 * for the first chunk, so that filesystems only need to
663 		 * write full blocks except possibly for the first and last
664 		 * chunks.
665 		 */
666 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
667 
668 		if (chunk > len)
669 			chunk = len;
670 		if (rw != UIO_READ && vp->v_type == VREG)
671 			bwillwrite();
672 		iaresid = 0;
673 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
674 		    ioflg, active_cred, file_cred, &iaresid, td);
675 		len -= chunk;	/* aresid calc already includes length */
676 		if (error)
677 			break;
678 		offset += chunk;
679 		base = (char *)base + chunk;
680 		kern_yield(PRI_USER);
681 	} while (len);
682 	if (aresid)
683 		*aresid = len + iaresid;
684 	return (error);
685 }
686 
687 #if OFF_MAX <= LONG_MAX
688 off_t
689 foffset_lock(struct file *fp, int flags)
690 {
691 	volatile short *flagsp;
692 	off_t res;
693 	short state;
694 
695 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
696 
697 	if ((flags & FOF_NOLOCK) != 0)
698 		return (atomic_load_long(&fp->f_offset));
699 
700 	/*
701 	 * According to McKusick the vn lock was protecting f_offset here.
702 	 * It is now protected by the FOFFSET_LOCKED flag.
703 	 */
704 	flagsp = &fp->f_vnread_flags;
705 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
706 		return (atomic_load_long(&fp->f_offset));
707 
708 	sleepq_lock(&fp->f_vnread_flags);
709 	state = atomic_load_16(flagsp);
710 	for (;;) {
711 		if ((state & FOFFSET_LOCKED) == 0) {
712 			if (!atomic_fcmpset_acq_16(flagsp, &state,
713 			    FOFFSET_LOCKED))
714 				continue;
715 			break;
716 		}
717 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
718 			if (!atomic_fcmpset_acq_16(flagsp, &state,
719 			    state | FOFFSET_LOCK_WAITING))
720 				continue;
721 		}
722 		DROP_GIANT();
723 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
724 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
725 		PICKUP_GIANT();
726 		sleepq_lock(&fp->f_vnread_flags);
727 		state = atomic_load_16(flagsp);
728 	}
729 	res = atomic_load_long(&fp->f_offset);
730 	sleepq_release(&fp->f_vnread_flags);
731 	return (res);
732 }
733 
734 void
735 foffset_unlock(struct file *fp, off_t val, int flags)
736 {
737 	volatile short *flagsp;
738 	short state;
739 
740 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
741 
742 	if ((flags & FOF_NOUPDATE) == 0)
743 		atomic_store_long(&fp->f_offset, val);
744 	if ((flags & FOF_NEXTOFF_R) != 0)
745 		fp->f_nextoff[UIO_READ] = val;
746 	if ((flags & FOF_NEXTOFF_W) != 0)
747 		fp->f_nextoff[UIO_WRITE] = val;
748 
749 	if ((flags & FOF_NOLOCK) != 0)
750 		return;
751 
752 	flagsp = &fp->f_vnread_flags;
753 	state = atomic_load_16(flagsp);
754 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
755 	    atomic_cmpset_rel_16(flagsp, state, 0))
756 		return;
757 
758 	sleepq_lock(&fp->f_vnread_flags);
759 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
760 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
761 	fp->f_vnread_flags = 0;
762 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
763 	sleepq_release(&fp->f_vnread_flags);
764 }
765 #else
766 off_t
767 foffset_lock(struct file *fp, int flags)
768 {
769 	struct mtx *mtxp;
770 	off_t res;
771 
772 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
773 
774 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
775 	mtx_lock(mtxp);
776 	if ((flags & FOF_NOLOCK) == 0) {
777 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
778 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
779 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
780 			    "vofflock", 0);
781 		}
782 		fp->f_vnread_flags |= FOFFSET_LOCKED;
783 	}
784 	res = fp->f_offset;
785 	mtx_unlock(mtxp);
786 	return (res);
787 }
788 
789 void
790 foffset_unlock(struct file *fp, off_t val, int flags)
791 {
792 	struct mtx *mtxp;
793 
794 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
795 
796 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
797 	mtx_lock(mtxp);
798 	if ((flags & FOF_NOUPDATE) == 0)
799 		fp->f_offset = val;
800 	if ((flags & FOF_NEXTOFF_R) != 0)
801 		fp->f_nextoff[UIO_READ] = val;
802 	if ((flags & FOF_NEXTOFF_W) != 0)
803 		fp->f_nextoff[UIO_WRITE] = val;
804 	if ((flags & FOF_NOLOCK) == 0) {
805 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
806 		    ("Lost FOFFSET_LOCKED"));
807 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
808 			wakeup(&fp->f_vnread_flags);
809 		fp->f_vnread_flags = 0;
810 	}
811 	mtx_unlock(mtxp);
812 }
813 #endif
814 
815 void
816 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
817 {
818 
819 	if ((flags & FOF_OFFSET) == 0)
820 		uio->uio_offset = foffset_lock(fp, flags);
821 }
822 
823 void
824 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
825 {
826 
827 	if ((flags & FOF_OFFSET) == 0)
828 		foffset_unlock(fp, uio->uio_offset, flags);
829 }
830 
831 static int
832 get_advice(struct file *fp, struct uio *uio)
833 {
834 	struct mtx *mtxp;
835 	int ret;
836 
837 	ret = POSIX_FADV_NORMAL;
838 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
839 		return (ret);
840 
841 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
842 	mtx_lock(mtxp);
843 	if (fp->f_advice != NULL &&
844 	    uio->uio_offset >= fp->f_advice->fa_start &&
845 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
846 		ret = fp->f_advice->fa_advice;
847 	mtx_unlock(mtxp);
848 	return (ret);
849 }
850 
851 int
852 vn_read_from_obj(struct vnode *vp, struct uio *uio)
853 {
854 	vm_object_t obj;
855 	vm_page_t ma[io_hold_cnt + 2];
856 	off_t off, vsz;
857 	ssize_t resid;
858 	int error, i, j;
859 
860 	obj = vp->v_object;
861 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
862 	MPASS(obj != NULL);
863 	MPASS(obj->type == OBJT_VNODE);
864 
865 	/*
866 	 * Depends on type stability of vm_objects.
867 	 */
868 	vm_object_pip_add(obj, 1);
869 	if ((obj->flags & OBJ_DEAD) != 0) {
870 		/*
871 		 * Note that object might be already reused from the
872 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
873 		 * we recheck for DOOMED vnode state after all pages
874 		 * are busied, and retract then.
875 		 *
876 		 * But we check for OBJ_DEAD to ensure that we do not
877 		 * busy pages while vm_object_terminate_pages()
878 		 * processes the queue.
879 		 */
880 		error = EJUSTRETURN;
881 		goto out_pip;
882 	}
883 
884 	resid = uio->uio_resid;
885 	off = uio->uio_offset;
886 	for (i = 0; resid > 0; i++) {
887 		MPASS(i < io_hold_cnt + 2);
888 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
889 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
890 		    VM_ALLOC_NOWAIT);
891 		if (ma[i] == NULL)
892 			break;
893 
894 		/*
895 		 * Skip invalid pages.  Valid mask can be partial only
896 		 * at EOF, and we clip later.
897 		 */
898 		if (vm_page_none_valid(ma[i])) {
899 			vm_page_sunbusy(ma[i]);
900 			break;
901 		}
902 
903 		resid -= PAGE_SIZE;
904 		off += PAGE_SIZE;
905 	}
906 	if (i == 0) {
907 		error = EJUSTRETURN;
908 		goto out_pip;
909 	}
910 
911 	/*
912 	 * Check VIRF_DOOMED after we busied our pages.  Since
913 	 * vgonel() terminates the vnode' vm_object, it cannot
914 	 * process past pages busied by us.
915 	 */
916 	if (VN_IS_DOOMED(vp)) {
917 		error = EJUSTRETURN;
918 		goto out;
919 	}
920 
921 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
922 	if (resid > uio->uio_resid)
923 		resid = uio->uio_resid;
924 
925 	/*
926 	 * Unlocked read of vnp_size is safe because truncation cannot
927 	 * pass busied page.  But we load vnp_size into a local
928 	 * variable so that possible concurrent extension does not
929 	 * break calculation.
930 	 */
931 #if defined(__powerpc__) && !defined(__powerpc64__)
932 	vsz = obj->un_pager.vnp.vnp_size;
933 #else
934 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
935 #endif
936 	if (uio->uio_offset + resid > vsz)
937 		resid = vsz - uio->uio_offset;
938 
939 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
940 
941 out:
942 	for (j = 0; j < i; j++) {
943 		if (error == 0)
944 			vm_page_reference(ma[j]);
945 		vm_page_sunbusy(ma[j]);
946 	}
947 out_pip:
948 	vm_object_pip_wakeup(obj);
949 	if (error != 0)
950 		return (error);
951 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
952 }
953 
954 /*
955  * File table vnode read routine.
956  */
957 static int
958 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
959     struct thread *td)
960 {
961 	struct vnode *vp;
962 	off_t orig_offset;
963 	int error, ioflag;
964 	int advice;
965 
966 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
967 	    uio->uio_td, td));
968 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
969 	vp = fp->f_vnode;
970 	ioflag = 0;
971 	if (fp->f_flag & FNONBLOCK)
972 		ioflag |= IO_NDELAY;
973 	if (fp->f_flag & O_DIRECT)
974 		ioflag |= IO_DIRECT;
975 
976 	/*
977 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
978 	 * allows us to avoid unneeded work outright.
979 	 */
980 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
981 	    (vp->v_irflag & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
982 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
983 		if (error == 0) {
984 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
985 			return (0);
986 		}
987 		if (error != EJUSTRETURN)
988 			return (error);
989 	}
990 
991 	advice = get_advice(fp, uio);
992 	vn_lock(vp, LK_SHARED | LK_RETRY);
993 
994 	switch (advice) {
995 	case POSIX_FADV_NORMAL:
996 	case POSIX_FADV_SEQUENTIAL:
997 	case POSIX_FADV_NOREUSE:
998 		ioflag |= sequential_heuristic(uio, fp);
999 		break;
1000 	case POSIX_FADV_RANDOM:
1001 		/* Disable read-ahead for random I/O. */
1002 		break;
1003 	}
1004 	orig_offset = uio->uio_offset;
1005 
1006 #ifdef MAC
1007 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1008 	if (error == 0)
1009 #endif
1010 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1011 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1012 	VOP_UNLOCK(vp);
1013 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1014 	    orig_offset != uio->uio_offset)
1015 		/*
1016 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1017 		 * for the backing file after a POSIX_FADV_NOREUSE
1018 		 * read(2).
1019 		 */
1020 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1021 		    POSIX_FADV_DONTNEED);
1022 	return (error);
1023 }
1024 
1025 /*
1026  * File table vnode write routine.
1027  */
1028 static int
1029 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1030     struct thread *td)
1031 {
1032 	struct vnode *vp;
1033 	struct mount *mp;
1034 	off_t orig_offset;
1035 	int error, ioflag, lock_flags;
1036 	int advice;
1037 
1038 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1039 	    uio->uio_td, td));
1040 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1041 	vp = fp->f_vnode;
1042 	if (vp->v_type == VREG)
1043 		bwillwrite();
1044 	ioflag = IO_UNIT;
1045 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1046 		ioflag |= IO_APPEND;
1047 	if (fp->f_flag & FNONBLOCK)
1048 		ioflag |= IO_NDELAY;
1049 	if (fp->f_flag & O_DIRECT)
1050 		ioflag |= IO_DIRECT;
1051 	if ((fp->f_flag & O_FSYNC) ||
1052 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1053 		ioflag |= IO_SYNC;
1054 	mp = NULL;
1055 	if (vp->v_type != VCHR &&
1056 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1057 		goto unlock;
1058 
1059 	advice = get_advice(fp, uio);
1060 
1061 	if (MNT_SHARED_WRITES(mp) ||
1062 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1063 		lock_flags = LK_SHARED;
1064 	} else {
1065 		lock_flags = LK_EXCLUSIVE;
1066 	}
1067 
1068 	vn_lock(vp, lock_flags | LK_RETRY);
1069 	switch (advice) {
1070 	case POSIX_FADV_NORMAL:
1071 	case POSIX_FADV_SEQUENTIAL:
1072 	case POSIX_FADV_NOREUSE:
1073 		ioflag |= sequential_heuristic(uio, fp);
1074 		break;
1075 	case POSIX_FADV_RANDOM:
1076 		/* XXX: Is this correct? */
1077 		break;
1078 	}
1079 	orig_offset = uio->uio_offset;
1080 
1081 #ifdef MAC
1082 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1083 	if (error == 0)
1084 #endif
1085 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1086 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1087 	VOP_UNLOCK(vp);
1088 	if (vp->v_type != VCHR)
1089 		vn_finished_write(mp);
1090 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1091 	    orig_offset != uio->uio_offset)
1092 		/*
1093 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1094 		 * for the backing file after a POSIX_FADV_NOREUSE
1095 		 * write(2).
1096 		 */
1097 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1098 		    POSIX_FADV_DONTNEED);
1099 unlock:
1100 	return (error);
1101 }
1102 
1103 /*
1104  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1105  * prevent the following deadlock:
1106  *
1107  * Assume that the thread A reads from the vnode vp1 into userspace
1108  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1109  * currently not resident, then system ends up with the call chain
1110  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1111  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1112  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1113  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1114  * backed by the pages of vnode vp1, and some page in buf2 is not
1115  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1116  *
1117  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1118  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1119  * Instead, it first tries to do the whole range i/o with pagefaults
1120  * disabled. If all pages in the i/o buffer are resident and mapped,
1121  * VOP will succeed (ignoring the genuine filesystem errors).
1122  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1123  * i/o in chunks, with all pages in the chunk prefaulted and held
1124  * using vm_fault_quick_hold_pages().
1125  *
1126  * Filesystems using this deadlock avoidance scheme should use the
1127  * array of the held pages from uio, saved in the curthread->td_ma,
1128  * instead of doing uiomove().  A helper function
1129  * vn_io_fault_uiomove() converts uiomove request into
1130  * uiomove_fromphys() over td_ma array.
1131  *
1132  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1133  * make the current i/o request atomic with respect to other i/os and
1134  * truncations.
1135  */
1136 
1137 /*
1138  * Decode vn_io_fault_args and perform the corresponding i/o.
1139  */
1140 static int
1141 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1142     struct thread *td)
1143 {
1144 	int error, save;
1145 
1146 	error = 0;
1147 	save = vm_fault_disable_pagefaults();
1148 	switch (args->kind) {
1149 	case VN_IO_FAULT_FOP:
1150 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1151 		    uio, args->cred, args->flags, td);
1152 		break;
1153 	case VN_IO_FAULT_VOP:
1154 		if (uio->uio_rw == UIO_READ) {
1155 			error = VOP_READ(args->args.vop_args.vp, uio,
1156 			    args->flags, args->cred);
1157 		} else if (uio->uio_rw == UIO_WRITE) {
1158 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1159 			    args->flags, args->cred);
1160 		}
1161 		break;
1162 	default:
1163 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1164 		    args->kind, uio->uio_rw);
1165 	}
1166 	vm_fault_enable_pagefaults(save);
1167 	return (error);
1168 }
1169 
1170 static int
1171 vn_io_fault_touch(char *base, const struct uio *uio)
1172 {
1173 	int r;
1174 
1175 	r = fubyte(base);
1176 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1177 		return (EFAULT);
1178 	return (0);
1179 }
1180 
1181 static int
1182 vn_io_fault_prefault_user(const struct uio *uio)
1183 {
1184 	char *base;
1185 	const struct iovec *iov;
1186 	size_t len;
1187 	ssize_t resid;
1188 	int error, i;
1189 
1190 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1191 	    ("vn_io_fault_prefault userspace"));
1192 
1193 	error = i = 0;
1194 	iov = uio->uio_iov;
1195 	resid = uio->uio_resid;
1196 	base = iov->iov_base;
1197 	len = iov->iov_len;
1198 	while (resid > 0) {
1199 		error = vn_io_fault_touch(base, uio);
1200 		if (error != 0)
1201 			break;
1202 		if (len < PAGE_SIZE) {
1203 			if (len != 0) {
1204 				error = vn_io_fault_touch(base + len - 1, uio);
1205 				if (error != 0)
1206 					break;
1207 				resid -= len;
1208 			}
1209 			if (++i >= uio->uio_iovcnt)
1210 				break;
1211 			iov = uio->uio_iov + i;
1212 			base = iov->iov_base;
1213 			len = iov->iov_len;
1214 		} else {
1215 			len -= PAGE_SIZE;
1216 			base += PAGE_SIZE;
1217 			resid -= PAGE_SIZE;
1218 		}
1219 	}
1220 	return (error);
1221 }
1222 
1223 /*
1224  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1225  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1226  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1227  * into args and call vn_io_fault1() to handle faults during the user
1228  * mode buffer accesses.
1229  */
1230 static int
1231 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1232     struct thread *td)
1233 {
1234 	vm_page_t ma[io_hold_cnt + 2];
1235 	struct uio *uio_clone, short_uio;
1236 	struct iovec short_iovec[1];
1237 	vm_page_t *prev_td_ma;
1238 	vm_prot_t prot;
1239 	vm_offset_t addr, end;
1240 	size_t len, resid;
1241 	ssize_t adv;
1242 	int error, cnt, saveheld, prev_td_ma_cnt;
1243 
1244 	if (vn_io_fault_prefault) {
1245 		error = vn_io_fault_prefault_user(uio);
1246 		if (error != 0)
1247 			return (error); /* Or ignore ? */
1248 	}
1249 
1250 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1251 
1252 	/*
1253 	 * The UFS follows IO_UNIT directive and replays back both
1254 	 * uio_offset and uio_resid if an error is encountered during the
1255 	 * operation.  But, since the iovec may be already advanced,
1256 	 * uio is still in an inconsistent state.
1257 	 *
1258 	 * Cache a copy of the original uio, which is advanced to the redo
1259 	 * point using UIO_NOCOPY below.
1260 	 */
1261 	uio_clone = cloneuio(uio);
1262 	resid = uio->uio_resid;
1263 
1264 	short_uio.uio_segflg = UIO_USERSPACE;
1265 	short_uio.uio_rw = uio->uio_rw;
1266 	short_uio.uio_td = uio->uio_td;
1267 
1268 	error = vn_io_fault_doio(args, uio, td);
1269 	if (error != EFAULT)
1270 		goto out;
1271 
1272 	atomic_add_long(&vn_io_faults_cnt, 1);
1273 	uio_clone->uio_segflg = UIO_NOCOPY;
1274 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1275 	uio_clone->uio_segflg = uio->uio_segflg;
1276 
1277 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1278 	prev_td_ma = td->td_ma;
1279 	prev_td_ma_cnt = td->td_ma_cnt;
1280 
1281 	while (uio_clone->uio_resid != 0) {
1282 		len = uio_clone->uio_iov->iov_len;
1283 		if (len == 0) {
1284 			KASSERT(uio_clone->uio_iovcnt >= 1,
1285 			    ("iovcnt underflow"));
1286 			uio_clone->uio_iov++;
1287 			uio_clone->uio_iovcnt--;
1288 			continue;
1289 		}
1290 		if (len > ptoa(io_hold_cnt))
1291 			len = ptoa(io_hold_cnt);
1292 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1293 		end = round_page(addr + len);
1294 		if (end < addr) {
1295 			error = EFAULT;
1296 			break;
1297 		}
1298 		cnt = atop(end - trunc_page(addr));
1299 		/*
1300 		 * A perfectly misaligned address and length could cause
1301 		 * both the start and the end of the chunk to use partial
1302 		 * page.  +2 accounts for such a situation.
1303 		 */
1304 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1305 		    addr, len, prot, ma, io_hold_cnt + 2);
1306 		if (cnt == -1) {
1307 			error = EFAULT;
1308 			break;
1309 		}
1310 		short_uio.uio_iov = &short_iovec[0];
1311 		short_iovec[0].iov_base = (void *)addr;
1312 		short_uio.uio_iovcnt = 1;
1313 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1314 		short_uio.uio_offset = uio_clone->uio_offset;
1315 		td->td_ma = ma;
1316 		td->td_ma_cnt = cnt;
1317 
1318 		error = vn_io_fault_doio(args, &short_uio, td);
1319 		vm_page_unhold_pages(ma, cnt);
1320 		adv = len - short_uio.uio_resid;
1321 
1322 		uio_clone->uio_iov->iov_base =
1323 		    (char *)uio_clone->uio_iov->iov_base + adv;
1324 		uio_clone->uio_iov->iov_len -= adv;
1325 		uio_clone->uio_resid -= adv;
1326 		uio_clone->uio_offset += adv;
1327 
1328 		uio->uio_resid -= adv;
1329 		uio->uio_offset += adv;
1330 
1331 		if (error != 0 || adv == 0)
1332 			break;
1333 	}
1334 	td->td_ma = prev_td_ma;
1335 	td->td_ma_cnt = prev_td_ma_cnt;
1336 	curthread_pflags_restore(saveheld);
1337 out:
1338 	free(uio_clone, M_IOV);
1339 	return (error);
1340 }
1341 
1342 static int
1343 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1344     int flags, struct thread *td)
1345 {
1346 	fo_rdwr_t *doio;
1347 	struct vnode *vp;
1348 	void *rl_cookie;
1349 	struct vn_io_fault_args args;
1350 	int error;
1351 
1352 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1353 	vp = fp->f_vnode;
1354 
1355 	/*
1356 	 * The ability to read(2) on a directory has historically been
1357 	 * allowed for all users, but this can and has been the source of
1358 	 * at least one security issue in the past.  As such, it is now hidden
1359 	 * away behind a sysctl for those that actually need it to use it, and
1360 	 * restricted to root when it's turned on to make it relatively safe to
1361 	 * leave on for longer sessions of need.
1362 	 */
1363 	if (vp->v_type == VDIR) {
1364 		KASSERT(uio->uio_rw == UIO_READ,
1365 		    ("illegal write attempted on a directory"));
1366 		if (!vfs_allow_read_dir)
1367 			return (EISDIR);
1368 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1369 			return (EISDIR);
1370 	}
1371 
1372 	foffset_lock_uio(fp, uio, flags);
1373 	if (do_vn_io_fault(vp, uio)) {
1374 		args.kind = VN_IO_FAULT_FOP;
1375 		args.args.fop_args.fp = fp;
1376 		args.args.fop_args.doio = doio;
1377 		args.cred = active_cred;
1378 		args.flags = flags | FOF_OFFSET;
1379 		if (uio->uio_rw == UIO_READ) {
1380 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1381 			    uio->uio_offset + uio->uio_resid);
1382 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1383 		    (flags & FOF_OFFSET) == 0) {
1384 			/* For appenders, punt and lock the whole range. */
1385 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1386 		} else {
1387 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1388 			    uio->uio_offset + uio->uio_resid);
1389 		}
1390 		error = vn_io_fault1(vp, uio, &args, td);
1391 		vn_rangelock_unlock(vp, rl_cookie);
1392 	} else {
1393 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1394 	}
1395 	foffset_unlock_uio(fp, uio, flags);
1396 	return (error);
1397 }
1398 
1399 /*
1400  * Helper function to perform the requested uiomove operation using
1401  * the held pages for io->uio_iov[0].iov_base buffer instead of
1402  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1403  * instead of iov_base prevents page faults that could occur due to
1404  * pmap_collect() invalidating the mapping created by
1405  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1406  * object cleanup revoking the write access from page mappings.
1407  *
1408  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1409  * instead of plain uiomove().
1410  */
1411 int
1412 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1413 {
1414 	struct uio transp_uio;
1415 	struct iovec transp_iov[1];
1416 	struct thread *td;
1417 	size_t adv;
1418 	int error, pgadv;
1419 
1420 	td = curthread;
1421 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1422 	    uio->uio_segflg != UIO_USERSPACE)
1423 		return (uiomove(data, xfersize, uio));
1424 
1425 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1426 	transp_iov[0].iov_base = data;
1427 	transp_uio.uio_iov = &transp_iov[0];
1428 	transp_uio.uio_iovcnt = 1;
1429 	if (xfersize > uio->uio_resid)
1430 		xfersize = uio->uio_resid;
1431 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1432 	transp_uio.uio_offset = 0;
1433 	transp_uio.uio_segflg = UIO_SYSSPACE;
1434 	/*
1435 	 * Since transp_iov points to data, and td_ma page array
1436 	 * corresponds to original uio->uio_iov, we need to invert the
1437 	 * direction of the i/o operation as passed to
1438 	 * uiomove_fromphys().
1439 	 */
1440 	switch (uio->uio_rw) {
1441 	case UIO_WRITE:
1442 		transp_uio.uio_rw = UIO_READ;
1443 		break;
1444 	case UIO_READ:
1445 		transp_uio.uio_rw = UIO_WRITE;
1446 		break;
1447 	}
1448 	transp_uio.uio_td = uio->uio_td;
1449 	error = uiomove_fromphys(td->td_ma,
1450 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1451 	    xfersize, &transp_uio);
1452 	adv = xfersize - transp_uio.uio_resid;
1453 	pgadv =
1454 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1455 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1456 	td->td_ma += pgadv;
1457 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1458 	    pgadv));
1459 	td->td_ma_cnt -= pgadv;
1460 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1461 	uio->uio_iov->iov_len -= adv;
1462 	uio->uio_resid -= adv;
1463 	uio->uio_offset += adv;
1464 	return (error);
1465 }
1466 
1467 int
1468 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1469     struct uio *uio)
1470 {
1471 	struct thread *td;
1472 	vm_offset_t iov_base;
1473 	int cnt, pgadv;
1474 
1475 	td = curthread;
1476 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1477 	    uio->uio_segflg != UIO_USERSPACE)
1478 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1479 
1480 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1481 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1482 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1483 	switch (uio->uio_rw) {
1484 	case UIO_WRITE:
1485 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1486 		    offset, cnt);
1487 		break;
1488 	case UIO_READ:
1489 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1490 		    cnt);
1491 		break;
1492 	}
1493 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1494 	td->td_ma += pgadv;
1495 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1496 	    pgadv));
1497 	td->td_ma_cnt -= pgadv;
1498 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1499 	uio->uio_iov->iov_len -= cnt;
1500 	uio->uio_resid -= cnt;
1501 	uio->uio_offset += cnt;
1502 	return (0);
1503 }
1504 
1505 /*
1506  * File table truncate routine.
1507  */
1508 static int
1509 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1510     struct thread *td)
1511 {
1512 	struct mount *mp;
1513 	struct vnode *vp;
1514 	void *rl_cookie;
1515 	int error;
1516 
1517 	vp = fp->f_vnode;
1518 
1519 	/*
1520 	 * Lock the whole range for truncation.  Otherwise split i/o
1521 	 * might happen partly before and partly after the truncation.
1522 	 */
1523 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1524 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1525 	if (error)
1526 		goto out1;
1527 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1528 	AUDIT_ARG_VNODE1(vp);
1529 	if (vp->v_type == VDIR) {
1530 		error = EISDIR;
1531 		goto out;
1532 	}
1533 #ifdef MAC
1534 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1535 	if (error)
1536 		goto out;
1537 #endif
1538 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1539 	    fp->f_cred);
1540 out:
1541 	VOP_UNLOCK(vp);
1542 	vn_finished_write(mp);
1543 out1:
1544 	vn_rangelock_unlock(vp, rl_cookie);
1545 	return (error);
1546 }
1547 
1548 /*
1549  * Truncate a file that is already locked.
1550  */
1551 int
1552 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1553     struct ucred *cred)
1554 {
1555 	struct vattr vattr;
1556 	int error;
1557 
1558 	error = VOP_ADD_WRITECOUNT(vp, 1);
1559 	if (error == 0) {
1560 		VATTR_NULL(&vattr);
1561 		vattr.va_size = length;
1562 		if (sync)
1563 			vattr.va_vaflags |= VA_SYNC;
1564 		error = VOP_SETATTR(vp, &vattr, cred);
1565 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1566 	}
1567 	return (error);
1568 }
1569 
1570 /*
1571  * File table vnode stat routine.
1572  */
1573 static int
1574 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1575     struct thread *td)
1576 {
1577 	struct vnode *vp = fp->f_vnode;
1578 	int error;
1579 
1580 	vn_lock(vp, LK_SHARED | LK_RETRY);
1581 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1582 	VOP_UNLOCK(vp);
1583 
1584 	return (error);
1585 }
1586 
1587 /*
1588  * File table vnode ioctl routine.
1589  */
1590 static int
1591 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1592     struct thread *td)
1593 {
1594 	struct vattr vattr;
1595 	struct vnode *vp;
1596 	struct fiobmap2_arg *bmarg;
1597 	int error;
1598 
1599 	vp = fp->f_vnode;
1600 	switch (vp->v_type) {
1601 	case VDIR:
1602 	case VREG:
1603 		switch (com) {
1604 		case FIONREAD:
1605 			vn_lock(vp, LK_SHARED | LK_RETRY);
1606 			error = VOP_GETATTR(vp, &vattr, active_cred);
1607 			VOP_UNLOCK(vp);
1608 			if (error == 0)
1609 				*(int *)data = vattr.va_size - fp->f_offset;
1610 			return (error);
1611 		case FIOBMAP2:
1612 			bmarg = (struct fiobmap2_arg *)data;
1613 			vn_lock(vp, LK_SHARED | LK_RETRY);
1614 #ifdef MAC
1615 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1616 			    vp);
1617 			if (error == 0)
1618 #endif
1619 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1620 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1621 			VOP_UNLOCK(vp);
1622 			return (error);
1623 		case FIONBIO:
1624 		case FIOASYNC:
1625 			return (0);
1626 		default:
1627 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1628 			    active_cred, td));
1629 		}
1630 		break;
1631 	case VCHR:
1632 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1633 		    active_cred, td));
1634 	default:
1635 		return (ENOTTY);
1636 	}
1637 }
1638 
1639 /*
1640  * File table vnode poll routine.
1641  */
1642 static int
1643 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1644     struct thread *td)
1645 {
1646 	struct vnode *vp;
1647 	int error;
1648 
1649 	vp = fp->f_vnode;
1650 #if defined(MAC) || defined(AUDIT)
1651 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1652 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1653 		AUDIT_ARG_VNODE1(vp);
1654 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1655 		VOP_UNLOCK(vp);
1656 		if (error != 0)
1657 			return (error);
1658 	}
1659 #endif
1660 	error = VOP_POLL(vp, events, fp->f_cred, td);
1661 	return (error);
1662 }
1663 
1664 /*
1665  * Acquire the requested lock and then check for validity.  LK_RETRY
1666  * permits vn_lock to return doomed vnodes.
1667  */
1668 static int __noinline
1669 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1670     int error)
1671 {
1672 
1673 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1674 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1675 
1676 	if (error == 0)
1677 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1678 
1679 	if ((flags & LK_RETRY) == 0) {
1680 		if (error == 0) {
1681 			VOP_UNLOCK(vp);
1682 			error = ENOENT;
1683 		}
1684 		return (error);
1685 	}
1686 
1687 	/*
1688 	 * LK_RETRY case.
1689 	 *
1690 	 * Nothing to do if we got the lock.
1691 	 */
1692 	if (error == 0)
1693 		return (0);
1694 
1695 	/*
1696 	 * Interlock was dropped by the call in _vn_lock.
1697 	 */
1698 	flags &= ~LK_INTERLOCK;
1699 	do {
1700 		error = VOP_LOCK1(vp, flags, file, line);
1701 	} while (error != 0);
1702 	return (0);
1703 }
1704 
1705 int
1706 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1707 {
1708 	int error;
1709 
1710 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1711 	    ("vn_lock: no locktype (%d passed)", flags));
1712 	VNPASS(vp->v_holdcnt > 0, vp);
1713 	error = VOP_LOCK1(vp, flags, file, line);
1714 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1715 		return (_vn_lock_fallback(vp, flags, file, line, error));
1716 	return (0);
1717 }
1718 
1719 /*
1720  * File table vnode close routine.
1721  */
1722 static int
1723 vn_closefile(struct file *fp, struct thread *td)
1724 {
1725 	struct vnode *vp;
1726 	struct flock lf;
1727 	int error;
1728 	bool ref;
1729 
1730 	vp = fp->f_vnode;
1731 	fp->f_ops = &badfileops;
1732 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1733 
1734 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1735 
1736 	if (__predict_false(ref)) {
1737 		lf.l_whence = SEEK_SET;
1738 		lf.l_start = 0;
1739 		lf.l_len = 0;
1740 		lf.l_type = F_UNLCK;
1741 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1742 		vrele(vp);
1743 	}
1744 	return (error);
1745 }
1746 
1747 /*
1748  * Preparing to start a filesystem write operation. If the operation is
1749  * permitted, then we bump the count of operations in progress and
1750  * proceed. If a suspend request is in progress, we wait until the
1751  * suspension is over, and then proceed.
1752  */
1753 static int
1754 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1755 {
1756 	int error, mflags;
1757 
1758 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1759 	    vfs_op_thread_enter(mp)) {
1760 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1761 		vfs_mp_count_add_pcpu(mp, writeopcount, 1);
1762 		vfs_op_thread_exit(mp);
1763 		return (0);
1764 	}
1765 
1766 	if (mplocked)
1767 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1768 	else
1769 		MNT_ILOCK(mp);
1770 
1771 	error = 0;
1772 
1773 	/*
1774 	 * Check on status of suspension.
1775 	 */
1776 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1777 	    mp->mnt_susp_owner != curthread) {
1778 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1779 		    (flags & PCATCH) : 0) | (PUSER - 1);
1780 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1781 			if (flags & V_NOWAIT) {
1782 				error = EWOULDBLOCK;
1783 				goto unlock;
1784 			}
1785 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1786 			    "suspfs", 0);
1787 			if (error)
1788 				goto unlock;
1789 		}
1790 	}
1791 	if (flags & V_XSLEEP)
1792 		goto unlock;
1793 	mp->mnt_writeopcount++;
1794 unlock:
1795 	if (error != 0 || (flags & V_XSLEEP) != 0)
1796 		MNT_REL(mp);
1797 	MNT_IUNLOCK(mp);
1798 	return (error);
1799 }
1800 
1801 int
1802 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1803 {
1804 	struct mount *mp;
1805 	int error;
1806 
1807 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1808 	    ("V_MNTREF requires mp"));
1809 
1810 	error = 0;
1811 	/*
1812 	 * If a vnode is provided, get and return the mount point that
1813 	 * to which it will write.
1814 	 */
1815 	if (vp != NULL) {
1816 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1817 			*mpp = NULL;
1818 			if (error != EOPNOTSUPP)
1819 				return (error);
1820 			return (0);
1821 		}
1822 	}
1823 	if ((mp = *mpp) == NULL)
1824 		return (0);
1825 
1826 	/*
1827 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1828 	 * a vfs_ref().
1829 	 * As long as a vnode is not provided we need to acquire a
1830 	 * refcount for the provided mountpoint too, in order to
1831 	 * emulate a vfs_ref().
1832 	 */
1833 	if (vp == NULL && (flags & V_MNTREF) == 0)
1834 		vfs_ref(mp);
1835 
1836 	return (vn_start_write_refed(mp, flags, false));
1837 }
1838 
1839 /*
1840  * Secondary suspension. Used by operations such as vop_inactive
1841  * routines that are needed by the higher level functions. These
1842  * are allowed to proceed until all the higher level functions have
1843  * completed (indicated by mnt_writeopcount dropping to zero). At that
1844  * time, these operations are halted until the suspension is over.
1845  */
1846 int
1847 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1848 {
1849 	struct mount *mp;
1850 	int error;
1851 
1852 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1853 	    ("V_MNTREF requires mp"));
1854 
1855  retry:
1856 	if (vp != NULL) {
1857 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1858 			*mpp = NULL;
1859 			if (error != EOPNOTSUPP)
1860 				return (error);
1861 			return (0);
1862 		}
1863 	}
1864 	/*
1865 	 * If we are not suspended or have not yet reached suspended
1866 	 * mode, then let the operation proceed.
1867 	 */
1868 	if ((mp = *mpp) == NULL)
1869 		return (0);
1870 
1871 	/*
1872 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1873 	 * a vfs_ref().
1874 	 * As long as a vnode is not provided we need to acquire a
1875 	 * refcount for the provided mountpoint too, in order to
1876 	 * emulate a vfs_ref().
1877 	 */
1878 	MNT_ILOCK(mp);
1879 	if (vp == NULL && (flags & V_MNTREF) == 0)
1880 		MNT_REF(mp);
1881 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1882 		mp->mnt_secondary_writes++;
1883 		mp->mnt_secondary_accwrites++;
1884 		MNT_IUNLOCK(mp);
1885 		return (0);
1886 	}
1887 	if (flags & V_NOWAIT) {
1888 		MNT_REL(mp);
1889 		MNT_IUNLOCK(mp);
1890 		return (EWOULDBLOCK);
1891 	}
1892 	/*
1893 	 * Wait for the suspension to finish.
1894 	 */
1895 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1896 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1897 	    "suspfs", 0);
1898 	vfs_rel(mp);
1899 	if (error == 0)
1900 		goto retry;
1901 	return (error);
1902 }
1903 
1904 /*
1905  * Filesystem write operation has completed. If we are suspending and this
1906  * operation is the last one, notify the suspender that the suspension is
1907  * now in effect.
1908  */
1909 void
1910 vn_finished_write(struct mount *mp)
1911 {
1912 	int c;
1913 
1914 	if (mp == NULL)
1915 		return;
1916 
1917 	if (vfs_op_thread_enter(mp)) {
1918 		vfs_mp_count_sub_pcpu(mp, writeopcount, 1);
1919 		vfs_mp_count_sub_pcpu(mp, ref, 1);
1920 		vfs_op_thread_exit(mp);
1921 		return;
1922 	}
1923 
1924 	MNT_ILOCK(mp);
1925 	vfs_assert_mount_counters(mp);
1926 	MNT_REL(mp);
1927 	c = --mp->mnt_writeopcount;
1928 	if (mp->mnt_vfs_ops == 0) {
1929 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1930 		MNT_IUNLOCK(mp);
1931 		return;
1932 	}
1933 	if (c < 0)
1934 		vfs_dump_mount_counters(mp);
1935 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1936 		wakeup(&mp->mnt_writeopcount);
1937 	MNT_IUNLOCK(mp);
1938 }
1939 
1940 /*
1941  * Filesystem secondary write operation has completed. If we are
1942  * suspending and this operation is the last one, notify the suspender
1943  * that the suspension is now in effect.
1944  */
1945 void
1946 vn_finished_secondary_write(struct mount *mp)
1947 {
1948 	if (mp == NULL)
1949 		return;
1950 	MNT_ILOCK(mp);
1951 	MNT_REL(mp);
1952 	mp->mnt_secondary_writes--;
1953 	if (mp->mnt_secondary_writes < 0)
1954 		panic("vn_finished_secondary_write: neg cnt");
1955 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1956 	    mp->mnt_secondary_writes <= 0)
1957 		wakeup(&mp->mnt_secondary_writes);
1958 	MNT_IUNLOCK(mp);
1959 }
1960 
1961 /*
1962  * Request a filesystem to suspend write operations.
1963  */
1964 int
1965 vfs_write_suspend(struct mount *mp, int flags)
1966 {
1967 	int error;
1968 
1969 	vfs_op_enter(mp);
1970 
1971 	MNT_ILOCK(mp);
1972 	vfs_assert_mount_counters(mp);
1973 	if (mp->mnt_susp_owner == curthread) {
1974 		vfs_op_exit_locked(mp);
1975 		MNT_IUNLOCK(mp);
1976 		return (EALREADY);
1977 	}
1978 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
1979 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1980 
1981 	/*
1982 	 * Unmount holds a write reference on the mount point.  If we
1983 	 * own busy reference and drain for writers, we deadlock with
1984 	 * the reference draining in the unmount path.  Callers of
1985 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1986 	 * vfs_busy() reference is owned and caller is not in the
1987 	 * unmount context.
1988 	 */
1989 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1990 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1991 		vfs_op_exit_locked(mp);
1992 		MNT_IUNLOCK(mp);
1993 		return (EBUSY);
1994 	}
1995 
1996 	mp->mnt_kern_flag |= MNTK_SUSPEND;
1997 	mp->mnt_susp_owner = curthread;
1998 	if (mp->mnt_writeopcount > 0)
1999 		(void) msleep(&mp->mnt_writeopcount,
2000 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2001 	else
2002 		MNT_IUNLOCK(mp);
2003 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2004 		vfs_write_resume(mp, 0);
2005 		/* vfs_write_resume does vfs_op_exit() for us */
2006 	}
2007 	return (error);
2008 }
2009 
2010 /*
2011  * Request a filesystem to resume write operations.
2012  */
2013 void
2014 vfs_write_resume(struct mount *mp, int flags)
2015 {
2016 
2017 	MNT_ILOCK(mp);
2018 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2019 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2020 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2021 				       MNTK_SUSPENDED);
2022 		mp->mnt_susp_owner = NULL;
2023 		wakeup(&mp->mnt_writeopcount);
2024 		wakeup(&mp->mnt_flag);
2025 		curthread->td_pflags &= ~TDP_IGNSUSP;
2026 		if ((flags & VR_START_WRITE) != 0) {
2027 			MNT_REF(mp);
2028 			mp->mnt_writeopcount++;
2029 		}
2030 		MNT_IUNLOCK(mp);
2031 		if ((flags & VR_NO_SUSPCLR) == 0)
2032 			VFS_SUSP_CLEAN(mp);
2033 		vfs_op_exit(mp);
2034 	} else if ((flags & VR_START_WRITE) != 0) {
2035 		MNT_REF(mp);
2036 		vn_start_write_refed(mp, 0, true);
2037 	} else {
2038 		MNT_IUNLOCK(mp);
2039 	}
2040 }
2041 
2042 /*
2043  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2044  * methods.
2045  */
2046 int
2047 vfs_write_suspend_umnt(struct mount *mp)
2048 {
2049 	int error;
2050 
2051 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2052 	    ("vfs_write_suspend_umnt: recursed"));
2053 
2054 	/* dounmount() already called vn_start_write(). */
2055 	for (;;) {
2056 		vn_finished_write(mp);
2057 		error = vfs_write_suspend(mp, 0);
2058 		if (error != 0) {
2059 			vn_start_write(NULL, &mp, V_WAIT);
2060 			return (error);
2061 		}
2062 		MNT_ILOCK(mp);
2063 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2064 			break;
2065 		MNT_IUNLOCK(mp);
2066 		vn_start_write(NULL, &mp, V_WAIT);
2067 	}
2068 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2069 	wakeup(&mp->mnt_flag);
2070 	MNT_IUNLOCK(mp);
2071 	curthread->td_pflags |= TDP_IGNSUSP;
2072 	return (0);
2073 }
2074 
2075 /*
2076  * Implement kqueues for files by translating it to vnode operation.
2077  */
2078 static int
2079 vn_kqfilter(struct file *fp, struct knote *kn)
2080 {
2081 
2082 	return (VOP_KQFILTER(fp->f_vnode, kn));
2083 }
2084 
2085 /*
2086  * Simplified in-kernel wrapper calls for extended attribute access.
2087  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2088  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2089  */
2090 int
2091 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2092     const char *attrname, int *buflen, char *buf, struct thread *td)
2093 {
2094 	struct uio	auio;
2095 	struct iovec	iov;
2096 	int	error;
2097 
2098 	iov.iov_len = *buflen;
2099 	iov.iov_base = buf;
2100 
2101 	auio.uio_iov = &iov;
2102 	auio.uio_iovcnt = 1;
2103 	auio.uio_rw = UIO_READ;
2104 	auio.uio_segflg = UIO_SYSSPACE;
2105 	auio.uio_td = td;
2106 	auio.uio_offset = 0;
2107 	auio.uio_resid = *buflen;
2108 
2109 	if ((ioflg & IO_NODELOCKED) == 0)
2110 		vn_lock(vp, LK_SHARED | LK_RETRY);
2111 
2112 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2113 
2114 	/* authorize attribute retrieval as kernel */
2115 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2116 	    td);
2117 
2118 	if ((ioflg & IO_NODELOCKED) == 0)
2119 		VOP_UNLOCK(vp);
2120 
2121 	if (error == 0) {
2122 		*buflen = *buflen - auio.uio_resid;
2123 	}
2124 
2125 	return (error);
2126 }
2127 
2128 /*
2129  * XXX failure mode if partially written?
2130  */
2131 int
2132 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2133     const char *attrname, int buflen, char *buf, struct thread *td)
2134 {
2135 	struct uio	auio;
2136 	struct iovec	iov;
2137 	struct mount	*mp;
2138 	int	error;
2139 
2140 	iov.iov_len = buflen;
2141 	iov.iov_base = buf;
2142 
2143 	auio.uio_iov = &iov;
2144 	auio.uio_iovcnt = 1;
2145 	auio.uio_rw = UIO_WRITE;
2146 	auio.uio_segflg = UIO_SYSSPACE;
2147 	auio.uio_td = td;
2148 	auio.uio_offset = 0;
2149 	auio.uio_resid = buflen;
2150 
2151 	if ((ioflg & IO_NODELOCKED) == 0) {
2152 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2153 			return (error);
2154 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2155 	}
2156 
2157 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2158 
2159 	/* authorize attribute setting as kernel */
2160 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2161 
2162 	if ((ioflg & IO_NODELOCKED) == 0) {
2163 		vn_finished_write(mp);
2164 		VOP_UNLOCK(vp);
2165 	}
2166 
2167 	return (error);
2168 }
2169 
2170 int
2171 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2172     const char *attrname, struct thread *td)
2173 {
2174 	struct mount	*mp;
2175 	int	error;
2176 
2177 	if ((ioflg & IO_NODELOCKED) == 0) {
2178 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2179 			return (error);
2180 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2181 	}
2182 
2183 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2184 
2185 	/* authorize attribute removal as kernel */
2186 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2187 	if (error == EOPNOTSUPP)
2188 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2189 		    NULL, td);
2190 
2191 	if ((ioflg & IO_NODELOCKED) == 0) {
2192 		vn_finished_write(mp);
2193 		VOP_UNLOCK(vp);
2194 	}
2195 
2196 	return (error);
2197 }
2198 
2199 static int
2200 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2201     struct vnode **rvp)
2202 {
2203 
2204 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2205 }
2206 
2207 int
2208 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2209 {
2210 
2211 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2212 	    lkflags, rvp));
2213 }
2214 
2215 int
2216 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2217     int lkflags, struct vnode **rvp)
2218 {
2219 	struct mount *mp;
2220 	int ltype, error;
2221 
2222 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2223 	mp = vp->v_mount;
2224 	ltype = VOP_ISLOCKED(vp);
2225 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2226 	    ("vn_vget_ino: vp not locked"));
2227 	error = vfs_busy(mp, MBF_NOWAIT);
2228 	if (error != 0) {
2229 		vfs_ref(mp);
2230 		VOP_UNLOCK(vp);
2231 		error = vfs_busy(mp, 0);
2232 		vn_lock(vp, ltype | LK_RETRY);
2233 		vfs_rel(mp);
2234 		if (error != 0)
2235 			return (ENOENT);
2236 		if (VN_IS_DOOMED(vp)) {
2237 			vfs_unbusy(mp);
2238 			return (ENOENT);
2239 		}
2240 	}
2241 	VOP_UNLOCK(vp);
2242 	error = alloc(mp, alloc_arg, lkflags, rvp);
2243 	vfs_unbusy(mp);
2244 	if (error != 0 || *rvp != vp)
2245 		vn_lock(vp, ltype | LK_RETRY);
2246 	if (VN_IS_DOOMED(vp)) {
2247 		if (error == 0) {
2248 			if (*rvp == vp)
2249 				vunref(vp);
2250 			else
2251 				vput(*rvp);
2252 		}
2253 		error = ENOENT;
2254 	}
2255 	return (error);
2256 }
2257 
2258 int
2259 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2260     struct thread *td)
2261 {
2262 
2263 	if (vp->v_type != VREG || td == NULL)
2264 		return (0);
2265 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2266 	    lim_cur(td, RLIMIT_FSIZE)) {
2267 		PROC_LOCK(td->td_proc);
2268 		kern_psignal(td->td_proc, SIGXFSZ);
2269 		PROC_UNLOCK(td->td_proc);
2270 		return (EFBIG);
2271 	}
2272 	return (0);
2273 }
2274 
2275 int
2276 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2277     struct thread *td)
2278 {
2279 	struct vnode *vp;
2280 
2281 	vp = fp->f_vnode;
2282 #ifdef AUDIT
2283 	vn_lock(vp, LK_SHARED | LK_RETRY);
2284 	AUDIT_ARG_VNODE1(vp);
2285 	VOP_UNLOCK(vp);
2286 #endif
2287 	return (setfmode(td, active_cred, vp, mode));
2288 }
2289 
2290 int
2291 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2292     struct thread *td)
2293 {
2294 	struct vnode *vp;
2295 
2296 	vp = fp->f_vnode;
2297 #ifdef AUDIT
2298 	vn_lock(vp, LK_SHARED | LK_RETRY);
2299 	AUDIT_ARG_VNODE1(vp);
2300 	VOP_UNLOCK(vp);
2301 #endif
2302 	return (setfown(td, active_cred, vp, uid, gid));
2303 }
2304 
2305 void
2306 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2307 {
2308 	vm_object_t object;
2309 
2310 	if ((object = vp->v_object) == NULL)
2311 		return;
2312 	VM_OBJECT_WLOCK(object);
2313 	vm_object_page_remove(object, start, end, 0);
2314 	VM_OBJECT_WUNLOCK(object);
2315 }
2316 
2317 int
2318 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2319 {
2320 	struct vattr va;
2321 	daddr_t bn, bnp;
2322 	uint64_t bsize;
2323 	off_t noff;
2324 	int error;
2325 
2326 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2327 	    ("Wrong command %lu", cmd));
2328 
2329 	if (vn_lock(vp, LK_SHARED) != 0)
2330 		return (EBADF);
2331 	if (vp->v_type != VREG) {
2332 		error = ENOTTY;
2333 		goto unlock;
2334 	}
2335 	error = VOP_GETATTR(vp, &va, cred);
2336 	if (error != 0)
2337 		goto unlock;
2338 	noff = *off;
2339 	if (noff >= va.va_size) {
2340 		error = ENXIO;
2341 		goto unlock;
2342 	}
2343 	bsize = vp->v_mount->mnt_stat.f_iosize;
2344 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2345 	    noff % bsize) {
2346 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2347 		if (error == EOPNOTSUPP) {
2348 			error = ENOTTY;
2349 			goto unlock;
2350 		}
2351 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2352 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2353 			noff = bn * bsize;
2354 			if (noff < *off)
2355 				noff = *off;
2356 			goto unlock;
2357 		}
2358 	}
2359 	if (noff > va.va_size)
2360 		noff = va.va_size;
2361 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2362 	if (cmd == FIOSEEKDATA)
2363 		error = ENXIO;
2364 unlock:
2365 	VOP_UNLOCK(vp);
2366 	if (error == 0)
2367 		*off = noff;
2368 	return (error);
2369 }
2370 
2371 int
2372 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2373 {
2374 	struct ucred *cred;
2375 	struct vnode *vp;
2376 	struct vattr vattr;
2377 	off_t foffset, size;
2378 	int error, noneg;
2379 
2380 	cred = td->td_ucred;
2381 	vp = fp->f_vnode;
2382 	foffset = foffset_lock(fp, 0);
2383 	noneg = (vp->v_type != VCHR);
2384 	error = 0;
2385 	switch (whence) {
2386 	case L_INCR:
2387 		if (noneg &&
2388 		    (foffset < 0 ||
2389 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2390 			error = EOVERFLOW;
2391 			break;
2392 		}
2393 		offset += foffset;
2394 		break;
2395 	case L_XTND:
2396 		vn_lock(vp, LK_SHARED | LK_RETRY);
2397 		error = VOP_GETATTR(vp, &vattr, cred);
2398 		VOP_UNLOCK(vp);
2399 		if (error)
2400 			break;
2401 
2402 		/*
2403 		 * If the file references a disk device, then fetch
2404 		 * the media size and use that to determine the ending
2405 		 * offset.
2406 		 */
2407 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2408 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2409 			vattr.va_size = size;
2410 		if (noneg &&
2411 		    (vattr.va_size > OFF_MAX ||
2412 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2413 			error = EOVERFLOW;
2414 			break;
2415 		}
2416 		offset += vattr.va_size;
2417 		break;
2418 	case L_SET:
2419 		break;
2420 	case SEEK_DATA:
2421 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2422 		if (error == ENOTTY)
2423 			error = EINVAL;
2424 		break;
2425 	case SEEK_HOLE:
2426 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2427 		if (error == ENOTTY)
2428 			error = EINVAL;
2429 		break;
2430 	default:
2431 		error = EINVAL;
2432 	}
2433 	if (error == 0 && noneg && offset < 0)
2434 		error = EINVAL;
2435 	if (error != 0)
2436 		goto drop;
2437 	VFS_KNOTE_UNLOCKED(vp, 0);
2438 	td->td_uretoff.tdu_off = offset;
2439 drop:
2440 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2441 	return (error);
2442 }
2443 
2444 int
2445 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2446     struct thread *td)
2447 {
2448 	int error;
2449 
2450 	/*
2451 	 * Grant permission if the caller is the owner of the file, or
2452 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2453 	 * on the file.  If the time pointer is null, then write
2454 	 * permission on the file is also sufficient.
2455 	 *
2456 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2457 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2458 	 * will be allowed to set the times [..] to the current
2459 	 * server time.
2460 	 */
2461 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2462 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2463 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2464 	return (error);
2465 }
2466 
2467 int
2468 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2469 {
2470 	struct vnode *vp;
2471 	int error;
2472 
2473 	if (fp->f_type == DTYPE_FIFO)
2474 		kif->kf_type = KF_TYPE_FIFO;
2475 	else
2476 		kif->kf_type = KF_TYPE_VNODE;
2477 	vp = fp->f_vnode;
2478 	vref(vp);
2479 	FILEDESC_SUNLOCK(fdp);
2480 	error = vn_fill_kinfo_vnode(vp, kif);
2481 	vrele(vp);
2482 	FILEDESC_SLOCK(fdp);
2483 	return (error);
2484 }
2485 
2486 static inline void
2487 vn_fill_junk(struct kinfo_file *kif)
2488 {
2489 	size_t len, olen;
2490 
2491 	/*
2492 	 * Simulate vn_fullpath returning changing values for a given
2493 	 * vp during e.g. coredump.
2494 	 */
2495 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2496 	olen = strlen(kif->kf_path);
2497 	if (len < olen)
2498 		strcpy(&kif->kf_path[len - 1], "$");
2499 	else
2500 		for (; olen < len; olen++)
2501 			strcpy(&kif->kf_path[olen], "A");
2502 }
2503 
2504 int
2505 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2506 {
2507 	struct vattr va;
2508 	char *fullpath, *freepath;
2509 	int error;
2510 
2511 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2512 	freepath = NULL;
2513 	fullpath = "-";
2514 	error = vn_fullpath(vp, &fullpath, &freepath);
2515 	if (error == 0) {
2516 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2517 	}
2518 	if (freepath != NULL)
2519 		free(freepath, M_TEMP);
2520 
2521 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2522 		vn_fill_junk(kif);
2523 	);
2524 
2525 	/*
2526 	 * Retrieve vnode attributes.
2527 	 */
2528 	va.va_fsid = VNOVAL;
2529 	va.va_rdev = NODEV;
2530 	vn_lock(vp, LK_SHARED | LK_RETRY);
2531 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2532 	VOP_UNLOCK(vp);
2533 	if (error != 0)
2534 		return (error);
2535 	if (va.va_fsid != VNOVAL)
2536 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2537 	else
2538 		kif->kf_un.kf_file.kf_file_fsid =
2539 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2540 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2541 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2542 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2543 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2544 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2545 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2546 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2547 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2548 	return (0);
2549 }
2550 
2551 int
2552 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2553     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2554     struct thread *td)
2555 {
2556 #ifdef HWPMC_HOOKS
2557 	struct pmckern_map_in pkm;
2558 #endif
2559 	struct mount *mp;
2560 	struct vnode *vp;
2561 	vm_object_t object;
2562 	vm_prot_t maxprot;
2563 	boolean_t writecounted;
2564 	int error;
2565 
2566 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2567     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2568 	/*
2569 	 * POSIX shared-memory objects are defined to have
2570 	 * kernel persistence, and are not defined to support
2571 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2572 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2573 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2574 	 * flag to request this behavior.
2575 	 */
2576 	if ((fp->f_flag & FPOSIXSHM) != 0)
2577 		flags |= MAP_NOSYNC;
2578 #endif
2579 	vp = fp->f_vnode;
2580 
2581 	/*
2582 	 * Ensure that file and memory protections are
2583 	 * compatible.  Note that we only worry about
2584 	 * writability if mapping is shared; in this case,
2585 	 * current and max prot are dictated by the open file.
2586 	 * XXX use the vnode instead?  Problem is: what
2587 	 * credentials do we use for determination? What if
2588 	 * proc does a setuid?
2589 	 */
2590 	mp = vp->v_mount;
2591 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2592 		maxprot = VM_PROT_NONE;
2593 		if ((prot & VM_PROT_EXECUTE) != 0)
2594 			return (EACCES);
2595 	} else
2596 		maxprot = VM_PROT_EXECUTE;
2597 	if ((fp->f_flag & FREAD) != 0)
2598 		maxprot |= VM_PROT_READ;
2599 	else if ((prot & VM_PROT_READ) != 0)
2600 		return (EACCES);
2601 
2602 	/*
2603 	 * If we are sharing potential changes via MAP_SHARED and we
2604 	 * are trying to get write permission although we opened it
2605 	 * without asking for it, bail out.
2606 	 */
2607 	if ((flags & MAP_SHARED) != 0) {
2608 		if ((fp->f_flag & FWRITE) != 0)
2609 			maxprot |= VM_PROT_WRITE;
2610 		else if ((prot & VM_PROT_WRITE) != 0)
2611 			return (EACCES);
2612 	} else {
2613 		maxprot |= VM_PROT_WRITE;
2614 		cap_maxprot |= VM_PROT_WRITE;
2615 	}
2616 	maxprot &= cap_maxprot;
2617 
2618 	/*
2619 	 * For regular files and shared memory, POSIX requires that
2620 	 * the value of foff be a legitimate offset within the data
2621 	 * object.  In particular, negative offsets are invalid.
2622 	 * Blocking negative offsets and overflows here avoids
2623 	 * possible wraparound or user-level access into reserved
2624 	 * ranges of the data object later.  In contrast, POSIX does
2625 	 * not dictate how offsets are used by device drivers, so in
2626 	 * the case of a device mapping a negative offset is passed
2627 	 * on.
2628 	 */
2629 	if (
2630 #ifdef _LP64
2631 	    size > OFF_MAX ||
2632 #endif
2633 	    foff > OFF_MAX - size)
2634 		return (EINVAL);
2635 
2636 	writecounted = FALSE;
2637 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2638 	    &foff, &object, &writecounted);
2639 	if (error != 0)
2640 		return (error);
2641 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2642 	    foff, writecounted, td);
2643 	if (error != 0) {
2644 		/*
2645 		 * If this mapping was accounted for in the vnode's
2646 		 * writecount, then undo that now.
2647 		 */
2648 		if (writecounted)
2649 			vm_pager_release_writecount(object, 0, size);
2650 		vm_object_deallocate(object);
2651 	}
2652 #ifdef HWPMC_HOOKS
2653 	/* Inform hwpmc(4) if an executable is being mapped. */
2654 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2655 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2656 			pkm.pm_file = vp;
2657 			pkm.pm_address = (uintptr_t) *addr;
2658 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2659 		}
2660 	}
2661 #endif
2662 	return (error);
2663 }
2664 
2665 void
2666 vn_fsid(struct vnode *vp, struct vattr *va)
2667 {
2668 	fsid_t *f;
2669 
2670 	f = &vp->v_mount->mnt_stat.f_fsid;
2671 	va->va_fsid = (uint32_t)f->val[1];
2672 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2673 	va->va_fsid += (uint32_t)f->val[0];
2674 }
2675 
2676 int
2677 vn_fsync_buf(struct vnode *vp, int waitfor)
2678 {
2679 	struct buf *bp, *nbp;
2680 	struct bufobj *bo;
2681 	struct mount *mp;
2682 	int error, maxretry;
2683 
2684 	error = 0;
2685 	maxretry = 10000;     /* large, arbitrarily chosen */
2686 	mp = NULL;
2687 	if (vp->v_type == VCHR) {
2688 		VI_LOCK(vp);
2689 		mp = vp->v_rdev->si_mountpt;
2690 		VI_UNLOCK(vp);
2691 	}
2692 	bo = &vp->v_bufobj;
2693 	BO_LOCK(bo);
2694 loop1:
2695 	/*
2696 	 * MARK/SCAN initialization to avoid infinite loops.
2697 	 */
2698         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2699 		bp->b_vflags &= ~BV_SCANNED;
2700 		bp->b_error = 0;
2701 	}
2702 
2703 	/*
2704 	 * Flush all dirty buffers associated with a vnode.
2705 	 */
2706 loop2:
2707 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2708 		if ((bp->b_vflags & BV_SCANNED) != 0)
2709 			continue;
2710 		bp->b_vflags |= BV_SCANNED;
2711 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2712 			if (waitfor != MNT_WAIT)
2713 				continue;
2714 			if (BUF_LOCK(bp,
2715 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2716 			    BO_LOCKPTR(bo)) != 0) {
2717 				BO_LOCK(bo);
2718 				goto loop1;
2719 			}
2720 			BO_LOCK(bo);
2721 		}
2722 		BO_UNLOCK(bo);
2723 		KASSERT(bp->b_bufobj == bo,
2724 		    ("bp %p wrong b_bufobj %p should be %p",
2725 		    bp, bp->b_bufobj, bo));
2726 		if ((bp->b_flags & B_DELWRI) == 0)
2727 			panic("fsync: not dirty");
2728 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2729 			vfs_bio_awrite(bp);
2730 		} else {
2731 			bremfree(bp);
2732 			bawrite(bp);
2733 		}
2734 		if (maxretry < 1000)
2735 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2736 		BO_LOCK(bo);
2737 		goto loop2;
2738 	}
2739 
2740 	/*
2741 	 * If synchronous the caller expects us to completely resolve all
2742 	 * dirty buffers in the system.  Wait for in-progress I/O to
2743 	 * complete (which could include background bitmap writes), then
2744 	 * retry if dirty blocks still exist.
2745 	 */
2746 	if (waitfor == MNT_WAIT) {
2747 		bufobj_wwait(bo, 0, 0);
2748 		if (bo->bo_dirty.bv_cnt > 0) {
2749 			/*
2750 			 * If we are unable to write any of these buffers
2751 			 * then we fail now rather than trying endlessly
2752 			 * to write them out.
2753 			 */
2754 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2755 				if ((error = bp->b_error) != 0)
2756 					break;
2757 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2758 			    (error == 0 && --maxretry >= 0))
2759 				goto loop1;
2760 			if (error == 0)
2761 				error = EAGAIN;
2762 		}
2763 	}
2764 	BO_UNLOCK(bo);
2765 	if (error != 0)
2766 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2767 
2768 	return (error);
2769 }
2770 
2771 /*
2772  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2773  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2774  * to do the actual copy.
2775  * vn_generic_copy_file_range() is factored out, so it can be called
2776  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2777  * different file systems.
2778  */
2779 int
2780 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2781     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2782     struct ucred *outcred, struct thread *fsize_td)
2783 {
2784 	int error;
2785 	size_t len;
2786 	uint64_t uvalin, uvalout;
2787 
2788 	len = *lenp;
2789 	*lenp = 0;		/* For error returns. */
2790 	error = 0;
2791 
2792 	/* Do some sanity checks on the arguments. */
2793 	uvalin = *inoffp;
2794 	uvalin += len;
2795 	uvalout = *outoffp;
2796 	uvalout += len;
2797 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2798 		error = EISDIR;
2799 	else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2800 	    (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2801 	    uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2802 	    outvp->v_type != VREG)
2803 		error = EINVAL;
2804 	if (error != 0)
2805 		goto out;
2806 
2807 	/*
2808 	 * If the two vnode are for the same file system, call
2809 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2810 	 * which can handle copies across multiple file systems.
2811 	 */
2812 	*lenp = len;
2813 	if (invp->v_mount == outvp->v_mount)
2814 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2815 		    lenp, flags, incred, outcred, fsize_td);
2816 	else
2817 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2818 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2819 out:
2820 	return (error);
2821 }
2822 
2823 /*
2824  * Test len bytes of data starting at dat for all bytes == 0.
2825  * Return true if all bytes are zero, false otherwise.
2826  * Expects dat to be well aligned.
2827  */
2828 static bool
2829 mem_iszero(void *dat, int len)
2830 {
2831 	int i;
2832 	const u_int *p;
2833 	const char *cp;
2834 
2835 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2836 		if (len >= sizeof(*p)) {
2837 			if (*p != 0)
2838 				return (false);
2839 		} else {
2840 			cp = (const char *)p;
2841 			for (i = 0; i < len; i++, cp++)
2842 				if (*cp != '\0')
2843 					return (false);
2844 		}
2845 	}
2846 	return (true);
2847 }
2848 
2849 /*
2850  * Look for a hole in the output file and, if found, adjust *outoffp
2851  * and *xferp to skip past the hole.
2852  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2853  * to be written as 0's upon return.
2854  */
2855 static off_t
2856 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2857     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2858 {
2859 	int error;
2860 	off_t delta;
2861 
2862 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2863 		*dataoffp = *outoffp;
2864 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2865 		    curthread);
2866 		if (error == 0) {
2867 			*holeoffp = *dataoffp;
2868 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2869 			    curthread);
2870 		}
2871 		if (error != 0 || *holeoffp == *dataoffp) {
2872 			/*
2873 			 * Since outvp is unlocked, it may be possible for
2874 			 * another thread to do a truncate(), lseek(), write()
2875 			 * creating a hole at startoff between the above
2876 			 * VOP_IOCTL() calls, if the other thread does not do
2877 			 * rangelocking.
2878 			 * If that happens, *holeoffp == *dataoffp and finding
2879 			 * the hole has failed, so disable vn_skip_hole().
2880 			 */
2881 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2882 			return (xfer2);
2883 		}
2884 		KASSERT(*dataoffp >= *outoffp,
2885 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2886 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2887 		KASSERT(*holeoffp > *dataoffp,
2888 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2889 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2890 	}
2891 
2892 	/*
2893 	 * If there is a hole before the data starts, advance *outoffp and
2894 	 * *xferp past the hole.
2895 	 */
2896 	if (*dataoffp > *outoffp) {
2897 		delta = *dataoffp - *outoffp;
2898 		if (delta >= *xferp) {
2899 			/* Entire *xferp is a hole. */
2900 			*outoffp += *xferp;
2901 			*xferp = 0;
2902 			return (0);
2903 		}
2904 		*xferp -= delta;
2905 		*outoffp += delta;
2906 		xfer2 = MIN(xfer2, *xferp);
2907 	}
2908 
2909 	/*
2910 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2911 	 * that the write ends at the start of the hole.
2912 	 * *holeoffp should always be greater than *outoffp, but for the
2913 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2914 	 * value.
2915 	 */
2916 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2917 		xfer2 = *holeoffp - *outoffp;
2918 	return (xfer2);
2919 }
2920 
2921 /*
2922  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2923  * dat is a maximum of blksize in length and can be written repeatedly in
2924  * the chunk.
2925  * If growfile == true, just grow the file via vn_truncate_locked() instead
2926  * of doing actual writes.
2927  * If checkhole == true, a hole is being punched, so skip over any hole
2928  * already in the output file.
2929  */
2930 static int
2931 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2932     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2933 {
2934 	struct mount *mp;
2935 	off_t dataoff, holeoff, xfer2;
2936 	int error, lckf;
2937 
2938 	/*
2939 	 * Loop around doing writes of blksize until write has been completed.
2940 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2941 	 * done for each iteration, since the xfer argument can be very
2942 	 * large if there is a large hole to punch in the output file.
2943 	 */
2944 	error = 0;
2945 	holeoff = 0;
2946 	do {
2947 		xfer2 = MIN(xfer, blksize);
2948 		if (checkhole) {
2949 			/*
2950 			 * Punching a hole.  Skip writing if there is
2951 			 * already a hole in the output file.
2952 			 */
2953 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2954 			    &dataoff, &holeoff, cred);
2955 			if (xfer == 0)
2956 				break;
2957 			if (holeoff < 0)
2958 				checkhole = false;
2959 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2960 			    (intmax_t)xfer2));
2961 		}
2962 		bwillwrite();
2963 		mp = NULL;
2964 		error = vn_start_write(outvp, &mp, V_WAIT);
2965 		if (error == 0) {
2966 			if (MNT_SHARED_WRITES(mp))
2967 				lckf = LK_SHARED;
2968 			else
2969 				lckf = LK_EXCLUSIVE;
2970 			error = vn_lock(outvp, lckf);
2971 		}
2972 		if (error == 0) {
2973 			if (growfile)
2974 				error = vn_truncate_locked(outvp, outoff + xfer,
2975 				    false, cred);
2976 			else {
2977 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2978 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
2979 				    curthread->td_ucred, cred, NULL, curthread);
2980 				outoff += xfer2;
2981 				xfer -= xfer2;
2982 			}
2983 			VOP_UNLOCK(outvp);
2984 		}
2985 		if (mp != NULL)
2986 			vn_finished_write(mp);
2987 	} while (!growfile && xfer > 0 && error == 0);
2988 	return (error);
2989 }
2990 
2991 /*
2992  * Copy a byte range of one file to another.  This function can handle the
2993  * case where invp and outvp are on different file systems.
2994  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2995  * is no better file system specific way to do it.
2996  */
2997 int
2998 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2999     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3000     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3001 {
3002 	struct vattr va;
3003 	struct mount *mp;
3004 	struct uio io;
3005 	off_t startoff, endoff, xfer, xfer2;
3006 	u_long blksize;
3007 	int error;
3008 	bool cantseek, readzeros, eof, lastblock;
3009 	ssize_t aresid;
3010 	size_t copylen, len, savlen;
3011 	char *dat;
3012 	long holein, holeout;
3013 
3014 	holein = holeout = 0;
3015 	savlen = len = *lenp;
3016 	error = 0;
3017 	dat = NULL;
3018 
3019 	error = vn_lock(invp, LK_SHARED);
3020 	if (error != 0)
3021 		goto out;
3022 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3023 		holein = 0;
3024 	VOP_UNLOCK(invp);
3025 
3026 	mp = NULL;
3027 	error = vn_start_write(outvp, &mp, V_WAIT);
3028 	if (error == 0)
3029 		error = vn_lock(outvp, LK_EXCLUSIVE);
3030 	if (error == 0) {
3031 		/*
3032 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3033 		 * now that outvp is locked.
3034 		 */
3035 		if (fsize_td != NULL) {
3036 			io.uio_offset = *outoffp;
3037 			io.uio_resid = len;
3038 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3039 			if (error != 0)
3040 				error = EFBIG;
3041 		}
3042 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3043 			holeout = 0;
3044 		/*
3045 		 * Holes that are past EOF do not need to be written as a block
3046 		 * of zero bytes.  So, truncate the output file as far as
3047 		 * possible and then use va.va_size to decide if writing 0
3048 		 * bytes is necessary in the loop below.
3049 		 */
3050 		if (error == 0)
3051 			error = VOP_GETATTR(outvp, &va, outcred);
3052 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3053 		    *outoffp + len) {
3054 #ifdef MAC
3055 			error = mac_vnode_check_write(curthread->td_ucred,
3056 			    outcred, outvp);
3057 			if (error == 0)
3058 #endif
3059 				error = vn_truncate_locked(outvp, *outoffp,
3060 				    false, outcred);
3061 			if (error == 0)
3062 				va.va_size = *outoffp;
3063 		}
3064 		VOP_UNLOCK(outvp);
3065 	}
3066 	if (mp != NULL)
3067 		vn_finished_write(mp);
3068 	if (error != 0)
3069 		goto out;
3070 
3071 	/*
3072 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3073 	 * If hole sizes aren't available, set the blksize to the larger
3074 	 * f_iosize of invp and outvp.
3075 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3076 	 * This value is clipped at 4Kbytes and 1Mbyte.
3077 	 */
3078 	blksize = MAX(holein, holeout);
3079 	if (blksize == 0)
3080 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3081 		    outvp->v_mount->mnt_stat.f_iosize);
3082 	if (blksize < 4096)
3083 		blksize = 4096;
3084 	else if (blksize > 1024 * 1024)
3085 		blksize = 1024 * 1024;
3086 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3087 
3088 	/*
3089 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3090 	 * to find holes.  Otherwise, just scan the read block for all 0s
3091 	 * in the inner loop where the data copying is done.
3092 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3093 	 * support holes on the server, but do not support FIOSEEKHOLE.
3094 	 */
3095 	eof = false;
3096 	while (len > 0 && error == 0 && !eof) {
3097 		endoff = 0;			/* To shut up compilers. */
3098 		cantseek = true;
3099 		startoff = *inoffp;
3100 		copylen = len;
3101 
3102 		/*
3103 		 * Find the next data area.  If there is just a hole to EOF,
3104 		 * FIOSEEKDATA should fail and then we drop down into the
3105 		 * inner loop and create the hole on the outvp file.
3106 		 * (I do not know if any file system will report a hole to
3107 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3108 		 *  will fail for those file systems.)
3109 		 *
3110 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3111 		 * the code just falls through to the inner copy loop.
3112 		 */
3113 		error = EINVAL;
3114 		if (holein > 0)
3115 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3116 			    incred, curthread);
3117 		if (error == 0) {
3118 			endoff = startoff;
3119 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3120 			    incred, curthread);
3121 			/*
3122 			 * Since invp is unlocked, it may be possible for
3123 			 * another thread to do a truncate(), lseek(), write()
3124 			 * creating a hole at startoff between the above
3125 			 * VOP_IOCTL() calls, if the other thread does not do
3126 			 * rangelocking.
3127 			 * If that happens, startoff == endoff and finding
3128 			 * the hole has failed, so set an error.
3129 			 */
3130 			if (error == 0 && startoff == endoff)
3131 				error = EINVAL; /* Any error. Reset to 0. */
3132 		}
3133 		if (error == 0) {
3134 			if (startoff > *inoffp) {
3135 				/* Found hole before data block. */
3136 				xfer = MIN(startoff - *inoffp, len);
3137 				if (*outoffp < va.va_size) {
3138 					/* Must write 0s to punch hole. */
3139 					xfer2 = MIN(va.va_size - *outoffp,
3140 					    xfer);
3141 					memset(dat, 0, MIN(xfer2, blksize));
3142 					error = vn_write_outvp(outvp, dat,
3143 					    *outoffp, xfer2, blksize, false,
3144 					    holeout > 0, outcred);
3145 				}
3146 
3147 				if (error == 0 && *outoffp + xfer >
3148 				    va.va_size && xfer == len)
3149 					/* Grow last block. */
3150 					error = vn_write_outvp(outvp, dat,
3151 					    *outoffp, xfer, blksize, true,
3152 					    false, outcred);
3153 				if (error == 0) {
3154 					*inoffp += xfer;
3155 					*outoffp += xfer;
3156 					len -= xfer;
3157 				}
3158 			}
3159 			copylen = MIN(len, endoff - startoff);
3160 			cantseek = false;
3161 		} else {
3162 			cantseek = true;
3163 			startoff = *inoffp;
3164 			copylen = len;
3165 			error = 0;
3166 		}
3167 
3168 		xfer = blksize;
3169 		if (cantseek) {
3170 			/*
3171 			 * Set first xfer to end at a block boundary, so that
3172 			 * holes are more likely detected in the loop below via
3173 			 * the for all bytes 0 method.
3174 			 */
3175 			xfer -= (*inoffp % blksize);
3176 		}
3177 		/* Loop copying the data block. */
3178 		while (copylen > 0 && error == 0 && !eof) {
3179 			if (copylen < xfer)
3180 				xfer = copylen;
3181 			error = vn_lock(invp, LK_SHARED);
3182 			if (error != 0)
3183 				goto out;
3184 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3185 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3186 			    curthread->td_ucred, incred, &aresid,
3187 			    curthread);
3188 			VOP_UNLOCK(invp);
3189 			lastblock = false;
3190 			if (error == 0 && aresid > 0) {
3191 				/* Stop the copy at EOF on the input file. */
3192 				xfer -= aresid;
3193 				eof = true;
3194 				lastblock = true;
3195 			}
3196 			if (error == 0) {
3197 				/*
3198 				 * Skip the write for holes past the initial EOF
3199 				 * of the output file, unless this is the last
3200 				 * write of the output file at EOF.
3201 				 */
3202 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3203 				    false;
3204 				if (xfer == len)
3205 					lastblock = true;
3206 				if (!cantseek || *outoffp < va.va_size ||
3207 				    lastblock || !readzeros)
3208 					error = vn_write_outvp(outvp, dat,
3209 					    *outoffp, xfer, blksize,
3210 					    readzeros && lastblock &&
3211 					    *outoffp >= va.va_size, false,
3212 					    outcred);
3213 				if (error == 0) {
3214 					*inoffp += xfer;
3215 					startoff += xfer;
3216 					*outoffp += xfer;
3217 					copylen -= xfer;
3218 					len -= xfer;
3219 				}
3220 			}
3221 			xfer = blksize;
3222 		}
3223 	}
3224 out:
3225 	*lenp = savlen - len;
3226 	free(dat, M_TEMP);
3227 	return (error);
3228 }
3229 
3230 static int
3231 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3232 {
3233 	struct mount *mp;
3234 	struct vnode *vp;
3235 	off_t olen, ooffset;
3236 	int error;
3237 #ifdef AUDIT
3238 	int audited_vnode1 = 0;
3239 #endif
3240 
3241 	vp = fp->f_vnode;
3242 	if (vp->v_type != VREG)
3243 		return (ENODEV);
3244 
3245 	/* Allocating blocks may take a long time, so iterate. */
3246 	for (;;) {
3247 		olen = len;
3248 		ooffset = offset;
3249 
3250 		bwillwrite();
3251 		mp = NULL;
3252 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3253 		if (error != 0)
3254 			break;
3255 		error = vn_lock(vp, LK_EXCLUSIVE);
3256 		if (error != 0) {
3257 			vn_finished_write(mp);
3258 			break;
3259 		}
3260 #ifdef AUDIT
3261 		if (!audited_vnode1) {
3262 			AUDIT_ARG_VNODE1(vp);
3263 			audited_vnode1 = 1;
3264 		}
3265 #endif
3266 #ifdef MAC
3267 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3268 		if (error == 0)
3269 #endif
3270 			error = VOP_ALLOCATE(vp, &offset, &len);
3271 		VOP_UNLOCK(vp);
3272 		vn_finished_write(mp);
3273 
3274 		if (olen + ooffset != offset + len) {
3275 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3276 			    ooffset, olen, offset, len);
3277 		}
3278 		if (error != 0 || len == 0)
3279 			break;
3280 		KASSERT(olen > len, ("Iteration did not make progress?"));
3281 		maybe_yield();
3282 	}
3283 
3284 	return (error);
3285 }
3286