xref: /freebsd/sys/kern/vfs_vnops.c (revision d13def78ccef6dbc25c2e197089ee5fc4d7b82c3)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/sx.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/sysctl.h>
76 #include <sys/ttycom.h>
77 #include <sys/conf.h>
78 #include <sys/syslog.h>
79 #include <sys/unistd.h>
80 #include <sys/user.h>
81 
82 #include <security/audit/audit.h>
83 #include <security/mac/mac_framework.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_extern.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pager.h>
92 
93 #ifdef HWPMC_HOOKS
94 #include <sys/pmckern.h>
95 #endif
96 
97 static fo_rdwr_t	vn_read;
98 static fo_rdwr_t	vn_write;
99 static fo_rdwr_t	vn_io_fault;
100 static fo_truncate_t	vn_truncate;
101 static fo_ioctl_t	vn_ioctl;
102 static fo_poll_t	vn_poll;
103 static fo_kqfilter_t	vn_kqfilter;
104 static fo_stat_t	vn_statfile;
105 static fo_close_t	vn_closefile;
106 static fo_mmap_t	vn_mmap;
107 static fo_fallocate_t	vn_fallocate;
108 
109 struct 	fileops vnops = {
110 	.fo_read = vn_io_fault,
111 	.fo_write = vn_io_fault,
112 	.fo_truncate = vn_truncate,
113 	.fo_ioctl = vn_ioctl,
114 	.fo_poll = vn_poll,
115 	.fo_kqfilter = vn_kqfilter,
116 	.fo_stat = vn_statfile,
117 	.fo_close = vn_closefile,
118 	.fo_chmod = vn_chmod,
119 	.fo_chown = vn_chown,
120 	.fo_sendfile = vn_sendfile,
121 	.fo_seek = vn_seek,
122 	.fo_fill_kinfo = vn_fill_kinfo,
123 	.fo_mmap = vn_mmap,
124 	.fo_fallocate = vn_fallocate,
125 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
126 };
127 
128 static const int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
131     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
132 static int vn_io_fault_prefault = 0;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
134     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static u_long vn_io_faults_cnt;
136 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
137     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
138 
139 static int vfs_allow_read_dir = 0;
140 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
141     &vfs_allow_read_dir, 0,
142     "Enable read(2) of directory by root for filesystems that support it");
143 
144 /*
145  * Returns true if vn_io_fault mode of handling the i/o request should
146  * be used.
147  */
148 static bool
149 do_vn_io_fault(struct vnode *vp, struct uio *uio)
150 {
151 	struct mount *mp;
152 
153 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
154 	    (mp = vp->v_mount) != NULL &&
155 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
156 }
157 
158 /*
159  * Structure used to pass arguments to vn_io_fault1(), to do either
160  * file- or vnode-based I/O calls.
161  */
162 struct vn_io_fault_args {
163 	enum {
164 		VN_IO_FAULT_FOP,
165 		VN_IO_FAULT_VOP
166 	} kind;
167 	struct ucred *cred;
168 	int flags;
169 	union {
170 		struct fop_args_tag {
171 			struct file *fp;
172 			fo_rdwr_t *doio;
173 		} fop_args;
174 		struct vop_args_tag {
175 			struct vnode *vp;
176 		} vop_args;
177 	} args;
178 };
179 
180 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
181     struct vn_io_fault_args *args, struct thread *td);
182 
183 int
184 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
185 {
186 	struct thread *td = ndp->ni_cnd.cn_thread;
187 
188 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
189 }
190 
191 /*
192  * Common code for vnode open operations via a name lookup.
193  * Lookup the vnode and invoke VOP_CREATE if needed.
194  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
195  *
196  * Note that this does NOT free nameidata for the successful case,
197  * due to the NDINIT being done elsewhere.
198  */
199 int
200 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
201     struct ucred *cred, struct file *fp)
202 {
203 	struct vnode *vp;
204 	struct mount *mp;
205 	struct thread *td = ndp->ni_cnd.cn_thread;
206 	struct vattr vat;
207 	struct vattr *vap = &vat;
208 	int fmode, error;
209 
210 restart:
211 	fmode = *flagp;
212 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
213 	    O_EXCL | O_DIRECTORY))
214 		return (EINVAL);
215 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
216 		ndp->ni_cnd.cn_nameiop = CREATE;
217 		/*
218 		 * Set NOCACHE to avoid flushing the cache when
219 		 * rolling in many files at once.
220 		*/
221 		ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
222 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
223 			ndp->ni_cnd.cn_flags |= FOLLOW;
224 		if ((fmode & O_BENEATH) != 0)
225 			ndp->ni_cnd.cn_flags |= BENEATH;
226 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
227 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
228 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
229 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
230 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
231 			bwillwrite();
232 		if ((error = namei(ndp)) != 0)
233 			return (error);
234 		if (ndp->ni_vp == NULL) {
235 			VATTR_NULL(vap);
236 			vap->va_type = VREG;
237 			vap->va_mode = cmode;
238 			if (fmode & O_EXCL)
239 				vap->va_vaflags |= VA_EXCLUSIVE;
240 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
241 				NDFREE(ndp, NDF_ONLY_PNBUF);
242 				vput(ndp->ni_dvp);
243 				if ((error = vn_start_write(NULL, &mp,
244 				    V_XSLEEP | PCATCH)) != 0)
245 					return (error);
246 				goto restart;
247 			}
248 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
249 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
250 #ifdef MAC
251 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
252 			    &ndp->ni_cnd, vap);
253 			if (error == 0)
254 #endif
255 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
256 						   &ndp->ni_cnd, vap);
257 			vput(ndp->ni_dvp);
258 			vn_finished_write(mp);
259 			if (error) {
260 				NDFREE(ndp, NDF_ONLY_PNBUF);
261 				return (error);
262 			}
263 			fmode &= ~O_TRUNC;
264 			vp = ndp->ni_vp;
265 		} else {
266 			if (ndp->ni_dvp == ndp->ni_vp)
267 				vrele(ndp->ni_dvp);
268 			else
269 				vput(ndp->ni_dvp);
270 			ndp->ni_dvp = NULL;
271 			vp = ndp->ni_vp;
272 			if (fmode & O_EXCL) {
273 				error = EEXIST;
274 				goto bad;
275 			}
276 			if (vp->v_type == VDIR) {
277 				error = EISDIR;
278 				goto bad;
279 			}
280 			fmode &= ~O_CREAT;
281 		}
282 	} else {
283 		ndp->ni_cnd.cn_nameiop = LOOKUP;
284 		ndp->ni_cnd.cn_flags = ISOPEN |
285 		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
286 		if (!(fmode & FWRITE))
287 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
288 		if ((fmode & O_BENEATH) != 0)
289 			ndp->ni_cnd.cn_flags |= BENEATH;
290 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
291 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
292 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
293 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
294 		if ((error = namei(ndp)) != 0)
295 			return (error);
296 		vp = ndp->ni_vp;
297 	}
298 	error = vn_open_vnode(vp, fmode, cred, td, fp);
299 	if (error)
300 		goto bad;
301 	*flagp = fmode;
302 	return (0);
303 bad:
304 	NDFREE(ndp, NDF_ONLY_PNBUF);
305 	vput(vp);
306 	*flagp = fmode;
307 	ndp->ni_vp = NULL;
308 	return (error);
309 }
310 
311 static int
312 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
313 {
314 	struct flock lf;
315 	int error, lock_flags, type;
316 
317 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
318 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
319 		return (0);
320 	KASSERT(fp != NULL, ("open with flock requires fp"));
321 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
322 		return (EOPNOTSUPP);
323 
324 	lock_flags = VOP_ISLOCKED(vp);
325 	VOP_UNLOCK(vp);
326 
327 	lf.l_whence = SEEK_SET;
328 	lf.l_start = 0;
329 	lf.l_len = 0;
330 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
331 	type = F_FLOCK;
332 	if ((fmode & FNONBLOCK) == 0)
333 		type |= F_WAIT;
334 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
335 	if (error == 0)
336 		fp->f_flag |= FHASLOCK;
337 
338 	vn_lock(vp, lock_flags | LK_RETRY);
339 	if (error == 0 && VN_IS_DOOMED(vp))
340 		error = ENOENT;
341 	return (error);
342 }
343 
344 /*
345  * Common code for vnode open operations once a vnode is located.
346  * Check permissions, and call the VOP_OPEN routine.
347  */
348 int
349 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
350     struct thread *td, struct file *fp)
351 {
352 	accmode_t accmode;
353 	int error;
354 
355 	if (vp->v_type == VLNK)
356 		return (EMLINK);
357 	if (vp->v_type == VSOCK)
358 		return (EOPNOTSUPP);
359 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
360 		return (ENOTDIR);
361 	accmode = 0;
362 	if (fmode & (FWRITE | O_TRUNC)) {
363 		if (vp->v_type == VDIR)
364 			return (EISDIR);
365 		accmode |= VWRITE;
366 	}
367 	if (fmode & FREAD)
368 		accmode |= VREAD;
369 	if (fmode & FEXEC)
370 		accmode |= VEXEC;
371 	if ((fmode & O_APPEND) && (fmode & FWRITE))
372 		accmode |= VAPPEND;
373 #ifdef MAC
374 	if (fmode & O_CREAT)
375 		accmode |= VCREAT;
376 	if (fmode & O_VERIFY)
377 		accmode |= VVERIFY;
378 	error = mac_vnode_check_open(cred, vp, accmode);
379 	if (error)
380 		return (error);
381 
382 	accmode &= ~(VCREAT | VVERIFY);
383 #endif
384 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
385 		error = VOP_ACCESS(vp, accmode, cred, td);
386 		if (error != 0)
387 			return (error);
388 	}
389 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
390 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
391 	error = VOP_OPEN(vp, fmode, cred, td, fp);
392 	if (error != 0)
393 		return (error);
394 
395 	error = vn_open_vnode_advlock(vp, fmode, fp);
396 	if (error == 0 && (fmode & FWRITE) != 0) {
397 		error = VOP_ADD_WRITECOUNT(vp, 1);
398 		if (error == 0) {
399 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
400 			     __func__, vp, vp->v_writecount);
401 		}
402 	}
403 
404 	/*
405 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
406 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
407 	 * Arrange for that by having fdrop() to use vn_closefile().
408 	 */
409 	if (error != 0) {
410 		fp->f_flag |= FOPENFAILED;
411 		fp->f_vnode = vp;
412 		if (fp->f_ops == &badfileops) {
413 			fp->f_type = DTYPE_VNODE;
414 			fp->f_ops = &vnops;
415 		}
416 		vref(vp);
417 	}
418 
419 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
420 	return (error);
421 
422 }
423 
424 /*
425  * Check for write permissions on the specified vnode.
426  * Prototype text segments cannot be written.
427  * It is racy.
428  */
429 int
430 vn_writechk(struct vnode *vp)
431 {
432 
433 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
434 	/*
435 	 * If there's shared text associated with
436 	 * the vnode, try to free it up once.  If
437 	 * we fail, we can't allow writing.
438 	 */
439 	if (VOP_IS_TEXT(vp))
440 		return (ETXTBSY);
441 
442 	return (0);
443 }
444 
445 /*
446  * Vnode close call
447  */
448 static int
449 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
450     struct thread *td, bool keep_ref)
451 {
452 	struct mount *mp;
453 	int error, lock_flags;
454 
455 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
456 	    MNT_EXTENDED_SHARED(vp->v_mount))
457 		lock_flags = LK_SHARED;
458 	else
459 		lock_flags = LK_EXCLUSIVE;
460 
461 	vn_start_write(vp, &mp, V_WAIT);
462 	vn_lock(vp, lock_flags | LK_RETRY);
463 	AUDIT_ARG_VNODE1(vp);
464 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
465 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
466 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
467 		    __func__, vp, vp->v_writecount);
468 	}
469 	error = VOP_CLOSE(vp, flags, file_cred, td);
470 	if (keep_ref)
471 		VOP_UNLOCK(vp);
472 	else
473 		vput(vp);
474 	vn_finished_write(mp);
475 	return (error);
476 }
477 
478 int
479 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
480     struct thread *td)
481 {
482 
483 	return (vn_close1(vp, flags, file_cred, td, false));
484 }
485 
486 /*
487  * Heuristic to detect sequential operation.
488  */
489 static int
490 sequential_heuristic(struct uio *uio, struct file *fp)
491 {
492 	enum uio_rw rw;
493 
494 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
495 
496 	rw = uio->uio_rw;
497 	if (fp->f_flag & FRDAHEAD)
498 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
499 
500 	/*
501 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
502 	 * that the first I/O is normally considered to be slightly
503 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
504 	 * unless previous seeks have reduced f_seqcount to 0, in which
505 	 * case offset 0 is not special.
506 	 */
507 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
508 	    uio->uio_offset == fp->f_nextoff[rw]) {
509 		/*
510 		 * f_seqcount is in units of fixed-size blocks so that it
511 		 * depends mainly on the amount of sequential I/O and not
512 		 * much on the number of sequential I/O's.  The fixed size
513 		 * of 16384 is hard-coded here since it is (not quite) just
514 		 * a magic size that works well here.  This size is more
515 		 * closely related to the best I/O size for real disks than
516 		 * to any block size used by software.
517 		 */
518 		if (uio->uio_resid >= IO_SEQMAX * 16384)
519 			fp->f_seqcount[rw] = IO_SEQMAX;
520 		else {
521 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
522 			if (fp->f_seqcount[rw] > IO_SEQMAX)
523 				fp->f_seqcount[rw] = IO_SEQMAX;
524 		}
525 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
526 	}
527 
528 	/* Not sequential.  Quickly draw-down sequentiality. */
529 	if (fp->f_seqcount[rw] > 1)
530 		fp->f_seqcount[rw] = 1;
531 	else
532 		fp->f_seqcount[rw] = 0;
533 	return (0);
534 }
535 
536 /*
537  * Package up an I/O request on a vnode into a uio and do it.
538  */
539 int
540 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
541     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
542     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
543 {
544 	struct uio auio;
545 	struct iovec aiov;
546 	struct mount *mp;
547 	struct ucred *cred;
548 	void *rl_cookie;
549 	struct vn_io_fault_args args;
550 	int error, lock_flags;
551 
552 	if (offset < 0 && vp->v_type != VCHR)
553 		return (EINVAL);
554 	auio.uio_iov = &aiov;
555 	auio.uio_iovcnt = 1;
556 	aiov.iov_base = base;
557 	aiov.iov_len = len;
558 	auio.uio_resid = len;
559 	auio.uio_offset = offset;
560 	auio.uio_segflg = segflg;
561 	auio.uio_rw = rw;
562 	auio.uio_td = td;
563 	error = 0;
564 
565 	if ((ioflg & IO_NODELOCKED) == 0) {
566 		if ((ioflg & IO_RANGELOCKED) == 0) {
567 			if (rw == UIO_READ) {
568 				rl_cookie = vn_rangelock_rlock(vp, offset,
569 				    offset + len);
570 			} else {
571 				rl_cookie = vn_rangelock_wlock(vp, offset,
572 				    offset + len);
573 			}
574 		} else
575 			rl_cookie = NULL;
576 		mp = NULL;
577 		if (rw == UIO_WRITE) {
578 			if (vp->v_type != VCHR &&
579 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
580 			    != 0)
581 				goto out;
582 			if (MNT_SHARED_WRITES(mp) ||
583 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
584 				lock_flags = LK_SHARED;
585 			else
586 				lock_flags = LK_EXCLUSIVE;
587 		} else
588 			lock_flags = LK_SHARED;
589 		vn_lock(vp, lock_flags | LK_RETRY);
590 	} else
591 		rl_cookie = NULL;
592 
593 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
594 #ifdef MAC
595 	if ((ioflg & IO_NOMACCHECK) == 0) {
596 		if (rw == UIO_READ)
597 			error = mac_vnode_check_read(active_cred, file_cred,
598 			    vp);
599 		else
600 			error = mac_vnode_check_write(active_cred, file_cred,
601 			    vp);
602 	}
603 #endif
604 	if (error == 0) {
605 		if (file_cred != NULL)
606 			cred = file_cred;
607 		else
608 			cred = active_cred;
609 		if (do_vn_io_fault(vp, &auio)) {
610 			args.kind = VN_IO_FAULT_VOP;
611 			args.cred = cred;
612 			args.flags = ioflg;
613 			args.args.vop_args.vp = vp;
614 			error = vn_io_fault1(vp, &auio, &args, td);
615 		} else if (rw == UIO_READ) {
616 			error = VOP_READ(vp, &auio, ioflg, cred);
617 		} else /* if (rw == UIO_WRITE) */ {
618 			error = VOP_WRITE(vp, &auio, ioflg, cred);
619 		}
620 	}
621 	if (aresid)
622 		*aresid = auio.uio_resid;
623 	else
624 		if (auio.uio_resid && error == 0)
625 			error = EIO;
626 	if ((ioflg & IO_NODELOCKED) == 0) {
627 		VOP_UNLOCK(vp);
628 		if (mp != NULL)
629 			vn_finished_write(mp);
630 	}
631  out:
632 	if (rl_cookie != NULL)
633 		vn_rangelock_unlock(vp, rl_cookie);
634 	return (error);
635 }
636 
637 /*
638  * Package up an I/O request on a vnode into a uio and do it.  The I/O
639  * request is split up into smaller chunks and we try to avoid saturating
640  * the buffer cache while potentially holding a vnode locked, so we
641  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
642  * to give other processes a chance to lock the vnode (either other processes
643  * core'ing the same binary, or unrelated processes scanning the directory).
644  */
645 int
646 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
647     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
648     struct ucred *file_cred, size_t *aresid, struct thread *td)
649 {
650 	int error = 0;
651 	ssize_t iaresid;
652 
653 	do {
654 		int chunk;
655 
656 		/*
657 		 * Force `offset' to a multiple of MAXBSIZE except possibly
658 		 * for the first chunk, so that filesystems only need to
659 		 * write full blocks except possibly for the first and last
660 		 * chunks.
661 		 */
662 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
663 
664 		if (chunk > len)
665 			chunk = len;
666 		if (rw != UIO_READ && vp->v_type == VREG)
667 			bwillwrite();
668 		iaresid = 0;
669 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
670 		    ioflg, active_cred, file_cred, &iaresid, td);
671 		len -= chunk;	/* aresid calc already includes length */
672 		if (error)
673 			break;
674 		offset += chunk;
675 		base = (char *)base + chunk;
676 		kern_yield(PRI_USER);
677 	} while (len);
678 	if (aresid)
679 		*aresid = len + iaresid;
680 	return (error);
681 }
682 
683 #if OFF_MAX <= LONG_MAX
684 off_t
685 foffset_lock(struct file *fp, int flags)
686 {
687 	volatile short *flagsp;
688 	off_t res;
689 	short state;
690 
691 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
692 
693 	if ((flags & FOF_NOLOCK) != 0)
694 		return (atomic_load_long(&fp->f_offset));
695 
696 	/*
697 	 * According to McKusick the vn lock was protecting f_offset here.
698 	 * It is now protected by the FOFFSET_LOCKED flag.
699 	 */
700 	flagsp = &fp->f_vnread_flags;
701 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
702 		return (atomic_load_long(&fp->f_offset));
703 
704 	sleepq_lock(&fp->f_vnread_flags);
705 	state = atomic_load_16(flagsp);
706 	for (;;) {
707 		if ((state & FOFFSET_LOCKED) == 0) {
708 			if (!atomic_fcmpset_acq_16(flagsp, &state,
709 			    FOFFSET_LOCKED))
710 				continue;
711 			break;
712 		}
713 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
714 			if (!atomic_fcmpset_acq_16(flagsp, &state,
715 			    state | FOFFSET_LOCK_WAITING))
716 				continue;
717 		}
718 		DROP_GIANT();
719 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
720 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
721 		PICKUP_GIANT();
722 		sleepq_lock(&fp->f_vnread_flags);
723 		state = atomic_load_16(flagsp);
724 	}
725 	res = atomic_load_long(&fp->f_offset);
726 	sleepq_release(&fp->f_vnread_flags);
727 	return (res);
728 }
729 
730 void
731 foffset_unlock(struct file *fp, off_t val, int flags)
732 {
733 	volatile short *flagsp;
734 	short state;
735 
736 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
737 
738 	if ((flags & FOF_NOUPDATE) == 0)
739 		atomic_store_long(&fp->f_offset, val);
740 	if ((flags & FOF_NEXTOFF_R) != 0)
741 		fp->f_nextoff[UIO_READ] = val;
742 	if ((flags & FOF_NEXTOFF_W) != 0)
743 		fp->f_nextoff[UIO_WRITE] = val;
744 
745 	if ((flags & FOF_NOLOCK) != 0)
746 		return;
747 
748 	flagsp = &fp->f_vnread_flags;
749 	state = atomic_load_16(flagsp);
750 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
751 	    atomic_cmpset_rel_16(flagsp, state, 0))
752 		return;
753 
754 	sleepq_lock(&fp->f_vnread_flags);
755 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
756 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
757 	fp->f_vnread_flags = 0;
758 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
759 	sleepq_release(&fp->f_vnread_flags);
760 }
761 #else
762 off_t
763 foffset_lock(struct file *fp, int flags)
764 {
765 	struct mtx *mtxp;
766 	off_t res;
767 
768 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
769 
770 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
771 	mtx_lock(mtxp);
772 	if ((flags & FOF_NOLOCK) == 0) {
773 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
774 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
775 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
776 			    "vofflock", 0);
777 		}
778 		fp->f_vnread_flags |= FOFFSET_LOCKED;
779 	}
780 	res = fp->f_offset;
781 	mtx_unlock(mtxp);
782 	return (res);
783 }
784 
785 void
786 foffset_unlock(struct file *fp, off_t val, int flags)
787 {
788 	struct mtx *mtxp;
789 
790 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
791 
792 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
793 	mtx_lock(mtxp);
794 	if ((flags & FOF_NOUPDATE) == 0)
795 		fp->f_offset = val;
796 	if ((flags & FOF_NEXTOFF_R) != 0)
797 		fp->f_nextoff[UIO_READ] = val;
798 	if ((flags & FOF_NEXTOFF_W) != 0)
799 		fp->f_nextoff[UIO_WRITE] = val;
800 	if ((flags & FOF_NOLOCK) == 0) {
801 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
802 		    ("Lost FOFFSET_LOCKED"));
803 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
804 			wakeup(&fp->f_vnread_flags);
805 		fp->f_vnread_flags = 0;
806 	}
807 	mtx_unlock(mtxp);
808 }
809 #endif
810 
811 void
812 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
813 {
814 
815 	if ((flags & FOF_OFFSET) == 0)
816 		uio->uio_offset = foffset_lock(fp, flags);
817 }
818 
819 void
820 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
821 {
822 
823 	if ((flags & FOF_OFFSET) == 0)
824 		foffset_unlock(fp, uio->uio_offset, flags);
825 }
826 
827 static int
828 get_advice(struct file *fp, struct uio *uio)
829 {
830 	struct mtx *mtxp;
831 	int ret;
832 
833 	ret = POSIX_FADV_NORMAL;
834 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
835 		return (ret);
836 
837 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
838 	mtx_lock(mtxp);
839 	if (fp->f_advice != NULL &&
840 	    uio->uio_offset >= fp->f_advice->fa_start &&
841 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
842 		ret = fp->f_advice->fa_advice;
843 	mtx_unlock(mtxp);
844 	return (ret);
845 }
846 
847 /*
848  * File table vnode read routine.
849  */
850 static int
851 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
852     struct thread *td)
853 {
854 	struct vnode *vp;
855 	off_t orig_offset;
856 	int error, ioflag;
857 	int advice;
858 
859 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
860 	    uio->uio_td, td));
861 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
862 	vp = fp->f_vnode;
863 	ioflag = 0;
864 	if (fp->f_flag & FNONBLOCK)
865 		ioflag |= IO_NDELAY;
866 	if (fp->f_flag & O_DIRECT)
867 		ioflag |= IO_DIRECT;
868 	advice = get_advice(fp, uio);
869 	vn_lock(vp, LK_SHARED | LK_RETRY);
870 
871 	switch (advice) {
872 	case POSIX_FADV_NORMAL:
873 	case POSIX_FADV_SEQUENTIAL:
874 	case POSIX_FADV_NOREUSE:
875 		ioflag |= sequential_heuristic(uio, fp);
876 		break;
877 	case POSIX_FADV_RANDOM:
878 		/* Disable read-ahead for random I/O. */
879 		break;
880 	}
881 	orig_offset = uio->uio_offset;
882 
883 #ifdef MAC
884 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
885 	if (error == 0)
886 #endif
887 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
888 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
889 	VOP_UNLOCK(vp);
890 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
891 	    orig_offset != uio->uio_offset)
892 		/*
893 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
894 		 * for the backing file after a POSIX_FADV_NOREUSE
895 		 * read(2).
896 		 */
897 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
898 		    POSIX_FADV_DONTNEED);
899 	return (error);
900 }
901 
902 /*
903  * File table vnode write routine.
904  */
905 static int
906 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
907     struct thread *td)
908 {
909 	struct vnode *vp;
910 	struct mount *mp;
911 	off_t orig_offset;
912 	int error, ioflag, lock_flags;
913 	int advice;
914 
915 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
916 	    uio->uio_td, td));
917 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
918 	vp = fp->f_vnode;
919 	if (vp->v_type == VREG)
920 		bwillwrite();
921 	ioflag = IO_UNIT;
922 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
923 		ioflag |= IO_APPEND;
924 	if (fp->f_flag & FNONBLOCK)
925 		ioflag |= IO_NDELAY;
926 	if (fp->f_flag & O_DIRECT)
927 		ioflag |= IO_DIRECT;
928 	if ((fp->f_flag & O_FSYNC) ||
929 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
930 		ioflag |= IO_SYNC;
931 	mp = NULL;
932 	if (vp->v_type != VCHR &&
933 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
934 		goto unlock;
935 
936 	advice = get_advice(fp, uio);
937 
938 	if (MNT_SHARED_WRITES(mp) ||
939 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
940 		lock_flags = LK_SHARED;
941 	} else {
942 		lock_flags = LK_EXCLUSIVE;
943 	}
944 
945 	vn_lock(vp, lock_flags | LK_RETRY);
946 	switch (advice) {
947 	case POSIX_FADV_NORMAL:
948 	case POSIX_FADV_SEQUENTIAL:
949 	case POSIX_FADV_NOREUSE:
950 		ioflag |= sequential_heuristic(uio, fp);
951 		break;
952 	case POSIX_FADV_RANDOM:
953 		/* XXX: Is this correct? */
954 		break;
955 	}
956 	orig_offset = uio->uio_offset;
957 
958 #ifdef MAC
959 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
960 	if (error == 0)
961 #endif
962 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
963 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
964 	VOP_UNLOCK(vp);
965 	if (vp->v_type != VCHR)
966 		vn_finished_write(mp);
967 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
968 	    orig_offset != uio->uio_offset)
969 		/*
970 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
971 		 * for the backing file after a POSIX_FADV_NOREUSE
972 		 * write(2).
973 		 */
974 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
975 		    POSIX_FADV_DONTNEED);
976 unlock:
977 	return (error);
978 }
979 
980 /*
981  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
982  * prevent the following deadlock:
983  *
984  * Assume that the thread A reads from the vnode vp1 into userspace
985  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
986  * currently not resident, then system ends up with the call chain
987  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
988  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
989  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
990  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
991  * backed by the pages of vnode vp1, and some page in buf2 is not
992  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
993  *
994  * To prevent the lock order reversal and deadlock, vn_io_fault() does
995  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
996  * Instead, it first tries to do the whole range i/o with pagefaults
997  * disabled. If all pages in the i/o buffer are resident and mapped,
998  * VOP will succeed (ignoring the genuine filesystem errors).
999  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1000  * i/o in chunks, with all pages in the chunk prefaulted and held
1001  * using vm_fault_quick_hold_pages().
1002  *
1003  * Filesystems using this deadlock avoidance scheme should use the
1004  * array of the held pages from uio, saved in the curthread->td_ma,
1005  * instead of doing uiomove().  A helper function
1006  * vn_io_fault_uiomove() converts uiomove request into
1007  * uiomove_fromphys() over td_ma array.
1008  *
1009  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1010  * make the current i/o request atomic with respect to other i/os and
1011  * truncations.
1012  */
1013 
1014 /*
1015  * Decode vn_io_fault_args and perform the corresponding i/o.
1016  */
1017 static int
1018 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1019     struct thread *td)
1020 {
1021 	int error, save;
1022 
1023 	error = 0;
1024 	save = vm_fault_disable_pagefaults();
1025 	switch (args->kind) {
1026 	case VN_IO_FAULT_FOP:
1027 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1028 		    uio, args->cred, args->flags, td);
1029 		break;
1030 	case VN_IO_FAULT_VOP:
1031 		if (uio->uio_rw == UIO_READ) {
1032 			error = VOP_READ(args->args.vop_args.vp, uio,
1033 			    args->flags, args->cred);
1034 		} else if (uio->uio_rw == UIO_WRITE) {
1035 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1036 			    args->flags, args->cred);
1037 		}
1038 		break;
1039 	default:
1040 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1041 		    args->kind, uio->uio_rw);
1042 	}
1043 	vm_fault_enable_pagefaults(save);
1044 	return (error);
1045 }
1046 
1047 static int
1048 vn_io_fault_touch(char *base, const struct uio *uio)
1049 {
1050 	int r;
1051 
1052 	r = fubyte(base);
1053 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1054 		return (EFAULT);
1055 	return (0);
1056 }
1057 
1058 static int
1059 vn_io_fault_prefault_user(const struct uio *uio)
1060 {
1061 	char *base;
1062 	const struct iovec *iov;
1063 	size_t len;
1064 	ssize_t resid;
1065 	int error, i;
1066 
1067 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1068 	    ("vn_io_fault_prefault userspace"));
1069 
1070 	error = i = 0;
1071 	iov = uio->uio_iov;
1072 	resid = uio->uio_resid;
1073 	base = iov->iov_base;
1074 	len = iov->iov_len;
1075 	while (resid > 0) {
1076 		error = vn_io_fault_touch(base, uio);
1077 		if (error != 0)
1078 			break;
1079 		if (len < PAGE_SIZE) {
1080 			if (len != 0) {
1081 				error = vn_io_fault_touch(base + len - 1, uio);
1082 				if (error != 0)
1083 					break;
1084 				resid -= len;
1085 			}
1086 			if (++i >= uio->uio_iovcnt)
1087 				break;
1088 			iov = uio->uio_iov + i;
1089 			base = iov->iov_base;
1090 			len = iov->iov_len;
1091 		} else {
1092 			len -= PAGE_SIZE;
1093 			base += PAGE_SIZE;
1094 			resid -= PAGE_SIZE;
1095 		}
1096 	}
1097 	return (error);
1098 }
1099 
1100 /*
1101  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1102  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1103  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1104  * into args and call vn_io_fault1() to handle faults during the user
1105  * mode buffer accesses.
1106  */
1107 static int
1108 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1109     struct thread *td)
1110 {
1111 	vm_page_t ma[io_hold_cnt + 2];
1112 	struct uio *uio_clone, short_uio;
1113 	struct iovec short_iovec[1];
1114 	vm_page_t *prev_td_ma;
1115 	vm_prot_t prot;
1116 	vm_offset_t addr, end;
1117 	size_t len, resid;
1118 	ssize_t adv;
1119 	int error, cnt, saveheld, prev_td_ma_cnt;
1120 
1121 	if (vn_io_fault_prefault) {
1122 		error = vn_io_fault_prefault_user(uio);
1123 		if (error != 0)
1124 			return (error); /* Or ignore ? */
1125 	}
1126 
1127 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1128 
1129 	/*
1130 	 * The UFS follows IO_UNIT directive and replays back both
1131 	 * uio_offset and uio_resid if an error is encountered during the
1132 	 * operation.  But, since the iovec may be already advanced,
1133 	 * uio is still in an inconsistent state.
1134 	 *
1135 	 * Cache a copy of the original uio, which is advanced to the redo
1136 	 * point using UIO_NOCOPY below.
1137 	 */
1138 	uio_clone = cloneuio(uio);
1139 	resid = uio->uio_resid;
1140 
1141 	short_uio.uio_segflg = UIO_USERSPACE;
1142 	short_uio.uio_rw = uio->uio_rw;
1143 	short_uio.uio_td = uio->uio_td;
1144 
1145 	error = vn_io_fault_doio(args, uio, td);
1146 	if (error != EFAULT)
1147 		goto out;
1148 
1149 	atomic_add_long(&vn_io_faults_cnt, 1);
1150 	uio_clone->uio_segflg = UIO_NOCOPY;
1151 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1152 	uio_clone->uio_segflg = uio->uio_segflg;
1153 
1154 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1155 	prev_td_ma = td->td_ma;
1156 	prev_td_ma_cnt = td->td_ma_cnt;
1157 
1158 	while (uio_clone->uio_resid != 0) {
1159 		len = uio_clone->uio_iov->iov_len;
1160 		if (len == 0) {
1161 			KASSERT(uio_clone->uio_iovcnt >= 1,
1162 			    ("iovcnt underflow"));
1163 			uio_clone->uio_iov++;
1164 			uio_clone->uio_iovcnt--;
1165 			continue;
1166 		}
1167 		if (len > io_hold_cnt * PAGE_SIZE)
1168 			len = io_hold_cnt * PAGE_SIZE;
1169 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1170 		end = round_page(addr + len);
1171 		if (end < addr) {
1172 			error = EFAULT;
1173 			break;
1174 		}
1175 		cnt = atop(end - trunc_page(addr));
1176 		/*
1177 		 * A perfectly misaligned address and length could cause
1178 		 * both the start and the end of the chunk to use partial
1179 		 * page.  +2 accounts for such a situation.
1180 		 */
1181 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1182 		    addr, len, prot, ma, io_hold_cnt + 2);
1183 		if (cnt == -1) {
1184 			error = EFAULT;
1185 			break;
1186 		}
1187 		short_uio.uio_iov = &short_iovec[0];
1188 		short_iovec[0].iov_base = (void *)addr;
1189 		short_uio.uio_iovcnt = 1;
1190 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1191 		short_uio.uio_offset = uio_clone->uio_offset;
1192 		td->td_ma = ma;
1193 		td->td_ma_cnt = cnt;
1194 
1195 		error = vn_io_fault_doio(args, &short_uio, td);
1196 		vm_page_unhold_pages(ma, cnt);
1197 		adv = len - short_uio.uio_resid;
1198 
1199 		uio_clone->uio_iov->iov_base =
1200 		    (char *)uio_clone->uio_iov->iov_base + adv;
1201 		uio_clone->uio_iov->iov_len -= adv;
1202 		uio_clone->uio_resid -= adv;
1203 		uio_clone->uio_offset += adv;
1204 
1205 		uio->uio_resid -= adv;
1206 		uio->uio_offset += adv;
1207 
1208 		if (error != 0 || adv == 0)
1209 			break;
1210 	}
1211 	td->td_ma = prev_td_ma;
1212 	td->td_ma_cnt = prev_td_ma_cnt;
1213 	curthread_pflags_restore(saveheld);
1214 out:
1215 	free(uio_clone, M_IOV);
1216 	return (error);
1217 }
1218 
1219 static int
1220 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1221     int flags, struct thread *td)
1222 {
1223 	fo_rdwr_t *doio;
1224 	struct vnode *vp;
1225 	void *rl_cookie;
1226 	struct vn_io_fault_args args;
1227 	int error;
1228 
1229 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1230 	vp = fp->f_vnode;
1231 
1232 	/*
1233 	 * The ability to read(2) on a directory has historically been
1234 	 * allowed for all users, but this can and has been the source of
1235 	 * at least one security issue in the past.  As such, it is now hidden
1236 	 * away behind a sysctl for those that actually need it to use it, and
1237 	 * restricted to root when it's turned on to make it relatively safe to
1238 	 * leave on for longer sessions of need.
1239 	 */
1240 	if (vp->v_type == VDIR) {
1241 		KASSERT(uio->uio_rw == UIO_READ,
1242 		    ("illegal write attempted on a directory"));
1243 		if (!vfs_allow_read_dir)
1244 			return (EISDIR);
1245 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1246 			return (EISDIR);
1247 	}
1248 
1249 	foffset_lock_uio(fp, uio, flags);
1250 	if (do_vn_io_fault(vp, uio)) {
1251 		args.kind = VN_IO_FAULT_FOP;
1252 		args.args.fop_args.fp = fp;
1253 		args.args.fop_args.doio = doio;
1254 		args.cred = active_cred;
1255 		args.flags = flags | FOF_OFFSET;
1256 		if (uio->uio_rw == UIO_READ) {
1257 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1258 			    uio->uio_offset + uio->uio_resid);
1259 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1260 		    (flags & FOF_OFFSET) == 0) {
1261 			/* For appenders, punt and lock the whole range. */
1262 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1263 		} else {
1264 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1265 			    uio->uio_offset + uio->uio_resid);
1266 		}
1267 		error = vn_io_fault1(vp, uio, &args, td);
1268 		vn_rangelock_unlock(vp, rl_cookie);
1269 	} else {
1270 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1271 	}
1272 	foffset_unlock_uio(fp, uio, flags);
1273 	return (error);
1274 }
1275 
1276 /*
1277  * Helper function to perform the requested uiomove operation using
1278  * the held pages for io->uio_iov[0].iov_base buffer instead of
1279  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1280  * instead of iov_base prevents page faults that could occur due to
1281  * pmap_collect() invalidating the mapping created by
1282  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1283  * object cleanup revoking the write access from page mappings.
1284  *
1285  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1286  * instead of plain uiomove().
1287  */
1288 int
1289 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1290 {
1291 	struct uio transp_uio;
1292 	struct iovec transp_iov[1];
1293 	struct thread *td;
1294 	size_t adv;
1295 	int error, pgadv;
1296 
1297 	td = curthread;
1298 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1299 	    uio->uio_segflg != UIO_USERSPACE)
1300 		return (uiomove(data, xfersize, uio));
1301 
1302 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1303 	transp_iov[0].iov_base = data;
1304 	transp_uio.uio_iov = &transp_iov[0];
1305 	transp_uio.uio_iovcnt = 1;
1306 	if (xfersize > uio->uio_resid)
1307 		xfersize = uio->uio_resid;
1308 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1309 	transp_uio.uio_offset = 0;
1310 	transp_uio.uio_segflg = UIO_SYSSPACE;
1311 	/*
1312 	 * Since transp_iov points to data, and td_ma page array
1313 	 * corresponds to original uio->uio_iov, we need to invert the
1314 	 * direction of the i/o operation as passed to
1315 	 * uiomove_fromphys().
1316 	 */
1317 	switch (uio->uio_rw) {
1318 	case UIO_WRITE:
1319 		transp_uio.uio_rw = UIO_READ;
1320 		break;
1321 	case UIO_READ:
1322 		transp_uio.uio_rw = UIO_WRITE;
1323 		break;
1324 	}
1325 	transp_uio.uio_td = uio->uio_td;
1326 	error = uiomove_fromphys(td->td_ma,
1327 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1328 	    xfersize, &transp_uio);
1329 	adv = xfersize - transp_uio.uio_resid;
1330 	pgadv =
1331 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1332 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1333 	td->td_ma += pgadv;
1334 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1335 	    pgadv));
1336 	td->td_ma_cnt -= pgadv;
1337 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1338 	uio->uio_iov->iov_len -= adv;
1339 	uio->uio_resid -= adv;
1340 	uio->uio_offset += adv;
1341 	return (error);
1342 }
1343 
1344 int
1345 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1346     struct uio *uio)
1347 {
1348 	struct thread *td;
1349 	vm_offset_t iov_base;
1350 	int cnt, pgadv;
1351 
1352 	td = curthread;
1353 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1354 	    uio->uio_segflg != UIO_USERSPACE)
1355 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1356 
1357 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1358 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1359 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1360 	switch (uio->uio_rw) {
1361 	case UIO_WRITE:
1362 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1363 		    offset, cnt);
1364 		break;
1365 	case UIO_READ:
1366 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1367 		    cnt);
1368 		break;
1369 	}
1370 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1371 	td->td_ma += pgadv;
1372 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1373 	    pgadv));
1374 	td->td_ma_cnt -= pgadv;
1375 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1376 	uio->uio_iov->iov_len -= cnt;
1377 	uio->uio_resid -= cnt;
1378 	uio->uio_offset += cnt;
1379 	return (0);
1380 }
1381 
1382 /*
1383  * File table truncate routine.
1384  */
1385 static int
1386 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1387     struct thread *td)
1388 {
1389 	struct mount *mp;
1390 	struct vnode *vp;
1391 	void *rl_cookie;
1392 	int error;
1393 
1394 	vp = fp->f_vnode;
1395 
1396 	/*
1397 	 * Lock the whole range for truncation.  Otherwise split i/o
1398 	 * might happen partly before and partly after the truncation.
1399 	 */
1400 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1401 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1402 	if (error)
1403 		goto out1;
1404 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1405 	AUDIT_ARG_VNODE1(vp);
1406 	if (vp->v_type == VDIR) {
1407 		error = EISDIR;
1408 		goto out;
1409 	}
1410 #ifdef MAC
1411 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1412 	if (error)
1413 		goto out;
1414 #endif
1415 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1416 	    fp->f_cred);
1417 out:
1418 	VOP_UNLOCK(vp);
1419 	vn_finished_write(mp);
1420 out1:
1421 	vn_rangelock_unlock(vp, rl_cookie);
1422 	return (error);
1423 }
1424 
1425 /*
1426  * Truncate a file that is already locked.
1427  */
1428 int
1429 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1430     struct ucred *cred)
1431 {
1432 	struct vattr vattr;
1433 	int error;
1434 
1435 	error = VOP_ADD_WRITECOUNT(vp, 1);
1436 	if (error == 0) {
1437 		VATTR_NULL(&vattr);
1438 		vattr.va_size = length;
1439 		if (sync)
1440 			vattr.va_vaflags |= VA_SYNC;
1441 		error = VOP_SETATTR(vp, &vattr, cred);
1442 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1443 	}
1444 	return (error);
1445 }
1446 
1447 /*
1448  * File table vnode stat routine.
1449  */
1450 static int
1451 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1452     struct thread *td)
1453 {
1454 	struct vnode *vp = fp->f_vnode;
1455 	int error;
1456 
1457 	vn_lock(vp, LK_SHARED | LK_RETRY);
1458 	error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1459 	VOP_UNLOCK(vp);
1460 
1461 	return (error);
1462 }
1463 
1464 /*
1465  * Stat a vnode; implementation for the stat syscall
1466  */
1467 int
1468 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1469     struct ucred *file_cred, struct thread *td)
1470 {
1471 	struct vattr vattr;
1472 	struct vattr *vap;
1473 	int error;
1474 	u_short mode;
1475 
1476 	AUDIT_ARG_VNODE1(vp);
1477 #ifdef MAC
1478 	error = mac_vnode_check_stat(active_cred, file_cred, vp);
1479 	if (error)
1480 		return (error);
1481 #endif
1482 
1483 	vap = &vattr;
1484 
1485 	/*
1486 	 * Initialize defaults for new and unusual fields, so that file
1487 	 * systems which don't support these fields don't need to know
1488 	 * about them.
1489 	 */
1490 	vap->va_birthtime.tv_sec = -1;
1491 	vap->va_birthtime.tv_nsec = 0;
1492 	vap->va_fsid = VNOVAL;
1493 	vap->va_rdev = NODEV;
1494 
1495 	error = VOP_GETATTR(vp, vap, active_cred);
1496 	if (error)
1497 		return (error);
1498 
1499 	/*
1500 	 * Zero the spare stat fields
1501 	 */
1502 	bzero(sb, sizeof *sb);
1503 
1504 	/*
1505 	 * Copy from vattr table
1506 	 */
1507 	if (vap->va_fsid != VNOVAL)
1508 		sb->st_dev = vap->va_fsid;
1509 	else
1510 		sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1511 	sb->st_ino = vap->va_fileid;
1512 	mode = vap->va_mode;
1513 	switch (vap->va_type) {
1514 	case VREG:
1515 		mode |= S_IFREG;
1516 		break;
1517 	case VDIR:
1518 		mode |= S_IFDIR;
1519 		break;
1520 	case VBLK:
1521 		mode |= S_IFBLK;
1522 		break;
1523 	case VCHR:
1524 		mode |= S_IFCHR;
1525 		break;
1526 	case VLNK:
1527 		mode |= S_IFLNK;
1528 		break;
1529 	case VSOCK:
1530 		mode |= S_IFSOCK;
1531 		break;
1532 	case VFIFO:
1533 		mode |= S_IFIFO;
1534 		break;
1535 	default:
1536 		return (EBADF);
1537 	}
1538 	sb->st_mode = mode;
1539 	sb->st_nlink = vap->va_nlink;
1540 	sb->st_uid = vap->va_uid;
1541 	sb->st_gid = vap->va_gid;
1542 	sb->st_rdev = vap->va_rdev;
1543 	if (vap->va_size > OFF_MAX)
1544 		return (EOVERFLOW);
1545 	sb->st_size = vap->va_size;
1546 	sb->st_atim.tv_sec = vap->va_atime.tv_sec;
1547 	sb->st_atim.tv_nsec = vap->va_atime.tv_nsec;
1548 	sb->st_mtim.tv_sec = vap->va_mtime.tv_sec;
1549 	sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec;
1550 	sb->st_ctim.tv_sec = vap->va_ctime.tv_sec;
1551 	sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec;
1552 	sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec;
1553 	sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec;
1554 
1555 	/*
1556 	 * According to www.opengroup.org, the meaning of st_blksize is
1557 	 *   "a filesystem-specific preferred I/O block size for this
1558 	 *    object.  In some filesystem types, this may vary from file
1559 	 *    to file"
1560 	 * Use minimum/default of PAGE_SIZE (e.g. for VCHR).
1561 	 */
1562 
1563 	sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1564 
1565 	sb->st_flags = vap->va_flags;
1566 	if (priv_check_cred_vfs_generation(td->td_ucred))
1567 		sb->st_gen = 0;
1568 	else
1569 		sb->st_gen = vap->va_gen;
1570 
1571 	sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1572 	return (0);
1573 }
1574 
1575 /*
1576  * File table vnode ioctl routine.
1577  */
1578 static int
1579 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1580     struct thread *td)
1581 {
1582 	struct vattr vattr;
1583 	struct vnode *vp;
1584 	struct fiobmap2_arg *bmarg;
1585 	int error;
1586 
1587 	vp = fp->f_vnode;
1588 	switch (vp->v_type) {
1589 	case VDIR:
1590 	case VREG:
1591 		switch (com) {
1592 		case FIONREAD:
1593 			vn_lock(vp, LK_SHARED | LK_RETRY);
1594 			error = VOP_GETATTR(vp, &vattr, active_cred);
1595 			VOP_UNLOCK(vp);
1596 			if (error == 0)
1597 				*(int *)data = vattr.va_size - fp->f_offset;
1598 			return (error);
1599 		case FIOBMAP2:
1600 			bmarg = (struct fiobmap2_arg *)data;
1601 			vn_lock(vp, LK_SHARED | LK_RETRY);
1602 #ifdef MAC
1603 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1604 			    vp);
1605 			if (error == 0)
1606 #endif
1607 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1608 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1609 			VOP_UNLOCK(vp);
1610 			return (error);
1611 		case FIONBIO:
1612 		case FIOASYNC:
1613 			return (0);
1614 		default:
1615 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1616 			    active_cred, td));
1617 		}
1618 		break;
1619 	case VCHR:
1620 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1621 		    active_cred, td));
1622 	default:
1623 		return (ENOTTY);
1624 	}
1625 }
1626 
1627 /*
1628  * File table vnode poll routine.
1629  */
1630 static int
1631 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1632     struct thread *td)
1633 {
1634 	struct vnode *vp;
1635 	int error;
1636 
1637 	vp = fp->f_vnode;
1638 #ifdef MAC
1639 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1640 	AUDIT_ARG_VNODE1(vp);
1641 	error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1642 	VOP_UNLOCK(vp);
1643 	if (!error)
1644 #endif
1645 
1646 	error = VOP_POLL(vp, events, fp->f_cred, td);
1647 	return (error);
1648 }
1649 
1650 /*
1651  * Acquire the requested lock and then check for validity.  LK_RETRY
1652  * permits vn_lock to return doomed vnodes.
1653  */
1654 static int __noinline
1655 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1656     int error)
1657 {
1658 
1659 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1660 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1661 
1662 	if (error == 0)
1663 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1664 
1665 	if ((flags & LK_RETRY) == 0) {
1666 		if (error == 0) {
1667 			VOP_UNLOCK(vp);
1668 			error = ENOENT;
1669 		}
1670 		return (error);
1671 	}
1672 
1673 	/*
1674 	 * LK_RETRY case.
1675 	 *
1676 	 * Nothing to do if we got the lock.
1677 	 */
1678 	if (error == 0)
1679 		return (0);
1680 
1681 	/*
1682 	 * Interlock was dropped by the call in _vn_lock.
1683 	 */
1684 	flags &= ~LK_INTERLOCK;
1685 	do {
1686 		error = VOP_LOCK1(vp, flags, file, line);
1687 	} while (error != 0);
1688 	return (0);
1689 }
1690 
1691 int
1692 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1693 {
1694 	int error;
1695 
1696 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1697 	    ("vn_lock: no locktype (%d passed)", flags));
1698 	VNPASS(vp->v_holdcnt > 0, vp);
1699 	error = VOP_LOCK1(vp, flags, file, line);
1700 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1701 		return (_vn_lock_fallback(vp, flags, file, line, error));
1702 	return (0);
1703 }
1704 
1705 /*
1706  * File table vnode close routine.
1707  */
1708 static int
1709 vn_closefile(struct file *fp, struct thread *td)
1710 {
1711 	struct vnode *vp;
1712 	struct flock lf;
1713 	int error;
1714 	bool ref;
1715 
1716 	vp = fp->f_vnode;
1717 	fp->f_ops = &badfileops;
1718 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1719 
1720 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1721 
1722 	if (__predict_false(ref)) {
1723 		lf.l_whence = SEEK_SET;
1724 		lf.l_start = 0;
1725 		lf.l_len = 0;
1726 		lf.l_type = F_UNLCK;
1727 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1728 		vrele(vp);
1729 	}
1730 	return (error);
1731 }
1732 
1733 /*
1734  * Preparing to start a filesystem write operation. If the operation is
1735  * permitted, then we bump the count of operations in progress and
1736  * proceed. If a suspend request is in progress, we wait until the
1737  * suspension is over, and then proceed.
1738  */
1739 static int
1740 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1741 {
1742 	int error, mflags;
1743 
1744 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1745 	    vfs_op_thread_enter(mp)) {
1746 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1747 		vfs_mp_count_add_pcpu(mp, writeopcount, 1);
1748 		vfs_op_thread_exit(mp);
1749 		return (0);
1750 	}
1751 
1752 	if (mplocked)
1753 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1754 	else
1755 		MNT_ILOCK(mp);
1756 
1757 	error = 0;
1758 
1759 	/*
1760 	 * Check on status of suspension.
1761 	 */
1762 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1763 	    mp->mnt_susp_owner != curthread) {
1764 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1765 		    (flags & PCATCH) : 0) | (PUSER - 1);
1766 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1767 			if (flags & V_NOWAIT) {
1768 				error = EWOULDBLOCK;
1769 				goto unlock;
1770 			}
1771 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1772 			    "suspfs", 0);
1773 			if (error)
1774 				goto unlock;
1775 		}
1776 	}
1777 	if (flags & V_XSLEEP)
1778 		goto unlock;
1779 	mp->mnt_writeopcount++;
1780 unlock:
1781 	if (error != 0 || (flags & V_XSLEEP) != 0)
1782 		MNT_REL(mp);
1783 	MNT_IUNLOCK(mp);
1784 	return (error);
1785 }
1786 
1787 int
1788 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1789 {
1790 	struct mount *mp;
1791 	int error;
1792 
1793 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1794 	    ("V_MNTREF requires mp"));
1795 
1796 	error = 0;
1797 	/*
1798 	 * If a vnode is provided, get and return the mount point that
1799 	 * to which it will write.
1800 	 */
1801 	if (vp != NULL) {
1802 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1803 			*mpp = NULL;
1804 			if (error != EOPNOTSUPP)
1805 				return (error);
1806 			return (0);
1807 		}
1808 	}
1809 	if ((mp = *mpp) == NULL)
1810 		return (0);
1811 
1812 	/*
1813 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1814 	 * a vfs_ref().
1815 	 * As long as a vnode is not provided we need to acquire a
1816 	 * refcount for the provided mountpoint too, in order to
1817 	 * emulate a vfs_ref().
1818 	 */
1819 	if (vp == NULL && (flags & V_MNTREF) == 0)
1820 		vfs_ref(mp);
1821 
1822 	return (vn_start_write_refed(mp, flags, false));
1823 }
1824 
1825 /*
1826  * Secondary suspension. Used by operations such as vop_inactive
1827  * routines that are needed by the higher level functions. These
1828  * are allowed to proceed until all the higher level functions have
1829  * completed (indicated by mnt_writeopcount dropping to zero). At that
1830  * time, these operations are halted until the suspension is over.
1831  */
1832 int
1833 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1834 {
1835 	struct mount *mp;
1836 	int error;
1837 
1838 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1839 	    ("V_MNTREF requires mp"));
1840 
1841  retry:
1842 	if (vp != NULL) {
1843 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1844 			*mpp = NULL;
1845 			if (error != EOPNOTSUPP)
1846 				return (error);
1847 			return (0);
1848 		}
1849 	}
1850 	/*
1851 	 * If we are not suspended or have not yet reached suspended
1852 	 * mode, then let the operation proceed.
1853 	 */
1854 	if ((mp = *mpp) == NULL)
1855 		return (0);
1856 
1857 	/*
1858 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1859 	 * a vfs_ref().
1860 	 * As long as a vnode is not provided we need to acquire a
1861 	 * refcount for the provided mountpoint too, in order to
1862 	 * emulate a vfs_ref().
1863 	 */
1864 	MNT_ILOCK(mp);
1865 	if (vp == NULL && (flags & V_MNTREF) == 0)
1866 		MNT_REF(mp);
1867 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1868 		mp->mnt_secondary_writes++;
1869 		mp->mnt_secondary_accwrites++;
1870 		MNT_IUNLOCK(mp);
1871 		return (0);
1872 	}
1873 	if (flags & V_NOWAIT) {
1874 		MNT_REL(mp);
1875 		MNT_IUNLOCK(mp);
1876 		return (EWOULDBLOCK);
1877 	}
1878 	/*
1879 	 * Wait for the suspension to finish.
1880 	 */
1881 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1882 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1883 	    "suspfs", 0);
1884 	vfs_rel(mp);
1885 	if (error == 0)
1886 		goto retry;
1887 	return (error);
1888 }
1889 
1890 /*
1891  * Filesystem write operation has completed. If we are suspending and this
1892  * operation is the last one, notify the suspender that the suspension is
1893  * now in effect.
1894  */
1895 void
1896 vn_finished_write(struct mount *mp)
1897 {
1898 	int c;
1899 
1900 	if (mp == NULL)
1901 		return;
1902 
1903 	if (vfs_op_thread_enter(mp)) {
1904 		vfs_mp_count_sub_pcpu(mp, writeopcount, 1);
1905 		vfs_mp_count_sub_pcpu(mp, ref, 1);
1906 		vfs_op_thread_exit(mp);
1907 		return;
1908 	}
1909 
1910 	MNT_ILOCK(mp);
1911 	vfs_assert_mount_counters(mp);
1912 	MNT_REL(mp);
1913 	c = --mp->mnt_writeopcount;
1914 	if (mp->mnt_vfs_ops == 0) {
1915 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1916 		MNT_IUNLOCK(mp);
1917 		return;
1918 	}
1919 	if (c < 0)
1920 		vfs_dump_mount_counters(mp);
1921 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1922 		wakeup(&mp->mnt_writeopcount);
1923 	MNT_IUNLOCK(mp);
1924 }
1925 
1926 /*
1927  * Filesystem secondary write operation has completed. If we are
1928  * suspending and this operation is the last one, notify the suspender
1929  * that the suspension is now in effect.
1930  */
1931 void
1932 vn_finished_secondary_write(struct mount *mp)
1933 {
1934 	if (mp == NULL)
1935 		return;
1936 	MNT_ILOCK(mp);
1937 	MNT_REL(mp);
1938 	mp->mnt_secondary_writes--;
1939 	if (mp->mnt_secondary_writes < 0)
1940 		panic("vn_finished_secondary_write: neg cnt");
1941 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1942 	    mp->mnt_secondary_writes <= 0)
1943 		wakeup(&mp->mnt_secondary_writes);
1944 	MNT_IUNLOCK(mp);
1945 }
1946 
1947 /*
1948  * Request a filesystem to suspend write operations.
1949  */
1950 int
1951 vfs_write_suspend(struct mount *mp, int flags)
1952 {
1953 	int error;
1954 
1955 	vfs_op_enter(mp);
1956 
1957 	MNT_ILOCK(mp);
1958 	vfs_assert_mount_counters(mp);
1959 	if (mp->mnt_susp_owner == curthread) {
1960 		vfs_op_exit_locked(mp);
1961 		MNT_IUNLOCK(mp);
1962 		return (EALREADY);
1963 	}
1964 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
1965 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1966 
1967 	/*
1968 	 * Unmount holds a write reference on the mount point.  If we
1969 	 * own busy reference and drain for writers, we deadlock with
1970 	 * the reference draining in the unmount path.  Callers of
1971 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1972 	 * vfs_busy() reference is owned and caller is not in the
1973 	 * unmount context.
1974 	 */
1975 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1976 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1977 		vfs_op_exit_locked(mp);
1978 		MNT_IUNLOCK(mp);
1979 		return (EBUSY);
1980 	}
1981 
1982 	mp->mnt_kern_flag |= MNTK_SUSPEND;
1983 	mp->mnt_susp_owner = curthread;
1984 	if (mp->mnt_writeopcount > 0)
1985 		(void) msleep(&mp->mnt_writeopcount,
1986 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1987 	else
1988 		MNT_IUNLOCK(mp);
1989 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
1990 		vfs_write_resume(mp, 0);
1991 		/* vfs_write_resume does vfs_op_exit() for us */
1992 	}
1993 	return (error);
1994 }
1995 
1996 /*
1997  * Request a filesystem to resume write operations.
1998  */
1999 void
2000 vfs_write_resume(struct mount *mp, int flags)
2001 {
2002 
2003 	MNT_ILOCK(mp);
2004 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2005 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2006 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2007 				       MNTK_SUSPENDED);
2008 		mp->mnt_susp_owner = NULL;
2009 		wakeup(&mp->mnt_writeopcount);
2010 		wakeup(&mp->mnt_flag);
2011 		curthread->td_pflags &= ~TDP_IGNSUSP;
2012 		if ((flags & VR_START_WRITE) != 0) {
2013 			MNT_REF(mp);
2014 			mp->mnt_writeopcount++;
2015 		}
2016 		MNT_IUNLOCK(mp);
2017 		if ((flags & VR_NO_SUSPCLR) == 0)
2018 			VFS_SUSP_CLEAN(mp);
2019 		vfs_op_exit(mp);
2020 	} else if ((flags & VR_START_WRITE) != 0) {
2021 		MNT_REF(mp);
2022 		vn_start_write_refed(mp, 0, true);
2023 	} else {
2024 		MNT_IUNLOCK(mp);
2025 	}
2026 }
2027 
2028 /*
2029  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2030  * methods.
2031  */
2032 int
2033 vfs_write_suspend_umnt(struct mount *mp)
2034 {
2035 	int error;
2036 
2037 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2038 	    ("vfs_write_suspend_umnt: recursed"));
2039 
2040 	/* dounmount() already called vn_start_write(). */
2041 	for (;;) {
2042 		vn_finished_write(mp);
2043 		error = vfs_write_suspend(mp, 0);
2044 		if (error != 0) {
2045 			vn_start_write(NULL, &mp, V_WAIT);
2046 			return (error);
2047 		}
2048 		MNT_ILOCK(mp);
2049 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2050 			break;
2051 		MNT_IUNLOCK(mp);
2052 		vn_start_write(NULL, &mp, V_WAIT);
2053 	}
2054 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2055 	wakeup(&mp->mnt_flag);
2056 	MNT_IUNLOCK(mp);
2057 	curthread->td_pflags |= TDP_IGNSUSP;
2058 	return (0);
2059 }
2060 
2061 /*
2062  * Implement kqueues for files by translating it to vnode operation.
2063  */
2064 static int
2065 vn_kqfilter(struct file *fp, struct knote *kn)
2066 {
2067 
2068 	return (VOP_KQFILTER(fp->f_vnode, kn));
2069 }
2070 
2071 /*
2072  * Simplified in-kernel wrapper calls for extended attribute access.
2073  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2074  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2075  */
2076 int
2077 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2078     const char *attrname, int *buflen, char *buf, struct thread *td)
2079 {
2080 	struct uio	auio;
2081 	struct iovec	iov;
2082 	int	error;
2083 
2084 	iov.iov_len = *buflen;
2085 	iov.iov_base = buf;
2086 
2087 	auio.uio_iov = &iov;
2088 	auio.uio_iovcnt = 1;
2089 	auio.uio_rw = UIO_READ;
2090 	auio.uio_segflg = UIO_SYSSPACE;
2091 	auio.uio_td = td;
2092 	auio.uio_offset = 0;
2093 	auio.uio_resid = *buflen;
2094 
2095 	if ((ioflg & IO_NODELOCKED) == 0)
2096 		vn_lock(vp, LK_SHARED | LK_RETRY);
2097 
2098 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2099 
2100 	/* authorize attribute retrieval as kernel */
2101 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2102 	    td);
2103 
2104 	if ((ioflg & IO_NODELOCKED) == 0)
2105 		VOP_UNLOCK(vp);
2106 
2107 	if (error == 0) {
2108 		*buflen = *buflen - auio.uio_resid;
2109 	}
2110 
2111 	return (error);
2112 }
2113 
2114 /*
2115  * XXX failure mode if partially written?
2116  */
2117 int
2118 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2119     const char *attrname, int buflen, char *buf, struct thread *td)
2120 {
2121 	struct uio	auio;
2122 	struct iovec	iov;
2123 	struct mount	*mp;
2124 	int	error;
2125 
2126 	iov.iov_len = buflen;
2127 	iov.iov_base = buf;
2128 
2129 	auio.uio_iov = &iov;
2130 	auio.uio_iovcnt = 1;
2131 	auio.uio_rw = UIO_WRITE;
2132 	auio.uio_segflg = UIO_SYSSPACE;
2133 	auio.uio_td = td;
2134 	auio.uio_offset = 0;
2135 	auio.uio_resid = buflen;
2136 
2137 	if ((ioflg & IO_NODELOCKED) == 0) {
2138 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2139 			return (error);
2140 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2141 	}
2142 
2143 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2144 
2145 	/* authorize attribute setting as kernel */
2146 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2147 
2148 	if ((ioflg & IO_NODELOCKED) == 0) {
2149 		vn_finished_write(mp);
2150 		VOP_UNLOCK(vp);
2151 	}
2152 
2153 	return (error);
2154 }
2155 
2156 int
2157 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2158     const char *attrname, struct thread *td)
2159 {
2160 	struct mount	*mp;
2161 	int	error;
2162 
2163 	if ((ioflg & IO_NODELOCKED) == 0) {
2164 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2165 			return (error);
2166 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2167 	}
2168 
2169 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2170 
2171 	/* authorize attribute removal as kernel */
2172 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2173 	if (error == EOPNOTSUPP)
2174 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2175 		    NULL, td);
2176 
2177 	if ((ioflg & IO_NODELOCKED) == 0) {
2178 		vn_finished_write(mp);
2179 		VOP_UNLOCK(vp);
2180 	}
2181 
2182 	return (error);
2183 }
2184 
2185 static int
2186 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2187     struct vnode **rvp)
2188 {
2189 
2190 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2191 }
2192 
2193 int
2194 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2195 {
2196 
2197 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2198 	    lkflags, rvp));
2199 }
2200 
2201 int
2202 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2203     int lkflags, struct vnode **rvp)
2204 {
2205 	struct mount *mp;
2206 	int ltype, error;
2207 
2208 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2209 	mp = vp->v_mount;
2210 	ltype = VOP_ISLOCKED(vp);
2211 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2212 	    ("vn_vget_ino: vp not locked"));
2213 	error = vfs_busy(mp, MBF_NOWAIT);
2214 	if (error != 0) {
2215 		vfs_ref(mp);
2216 		VOP_UNLOCK(vp);
2217 		error = vfs_busy(mp, 0);
2218 		vn_lock(vp, ltype | LK_RETRY);
2219 		vfs_rel(mp);
2220 		if (error != 0)
2221 			return (ENOENT);
2222 		if (VN_IS_DOOMED(vp)) {
2223 			vfs_unbusy(mp);
2224 			return (ENOENT);
2225 		}
2226 	}
2227 	VOP_UNLOCK(vp);
2228 	error = alloc(mp, alloc_arg, lkflags, rvp);
2229 	vfs_unbusy(mp);
2230 	if (error != 0 || *rvp != vp)
2231 		vn_lock(vp, ltype | LK_RETRY);
2232 	if (VN_IS_DOOMED(vp)) {
2233 		if (error == 0) {
2234 			if (*rvp == vp)
2235 				vunref(vp);
2236 			else
2237 				vput(*rvp);
2238 		}
2239 		error = ENOENT;
2240 	}
2241 	return (error);
2242 }
2243 
2244 int
2245 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2246     struct thread *td)
2247 {
2248 
2249 	if (vp->v_type != VREG || td == NULL)
2250 		return (0);
2251 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2252 	    lim_cur(td, RLIMIT_FSIZE)) {
2253 		PROC_LOCK(td->td_proc);
2254 		kern_psignal(td->td_proc, SIGXFSZ);
2255 		PROC_UNLOCK(td->td_proc);
2256 		return (EFBIG);
2257 	}
2258 	return (0);
2259 }
2260 
2261 int
2262 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2263     struct thread *td)
2264 {
2265 	struct vnode *vp;
2266 
2267 	vp = fp->f_vnode;
2268 #ifdef AUDIT
2269 	vn_lock(vp, LK_SHARED | LK_RETRY);
2270 	AUDIT_ARG_VNODE1(vp);
2271 	VOP_UNLOCK(vp);
2272 #endif
2273 	return (setfmode(td, active_cred, vp, mode));
2274 }
2275 
2276 int
2277 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2278     struct thread *td)
2279 {
2280 	struct vnode *vp;
2281 
2282 	vp = fp->f_vnode;
2283 #ifdef AUDIT
2284 	vn_lock(vp, LK_SHARED | LK_RETRY);
2285 	AUDIT_ARG_VNODE1(vp);
2286 	VOP_UNLOCK(vp);
2287 #endif
2288 	return (setfown(td, active_cred, vp, uid, gid));
2289 }
2290 
2291 void
2292 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2293 {
2294 	vm_object_t object;
2295 
2296 	if ((object = vp->v_object) == NULL)
2297 		return;
2298 	VM_OBJECT_WLOCK(object);
2299 	vm_object_page_remove(object, start, end, 0);
2300 	VM_OBJECT_WUNLOCK(object);
2301 }
2302 
2303 int
2304 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2305 {
2306 	struct vattr va;
2307 	daddr_t bn, bnp;
2308 	uint64_t bsize;
2309 	off_t noff;
2310 	int error;
2311 
2312 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2313 	    ("Wrong command %lu", cmd));
2314 
2315 	if (vn_lock(vp, LK_SHARED) != 0)
2316 		return (EBADF);
2317 	if (vp->v_type != VREG) {
2318 		error = ENOTTY;
2319 		goto unlock;
2320 	}
2321 	error = VOP_GETATTR(vp, &va, cred);
2322 	if (error != 0)
2323 		goto unlock;
2324 	noff = *off;
2325 	if (noff >= va.va_size) {
2326 		error = ENXIO;
2327 		goto unlock;
2328 	}
2329 	bsize = vp->v_mount->mnt_stat.f_iosize;
2330 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2331 	    noff % bsize) {
2332 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2333 		if (error == EOPNOTSUPP) {
2334 			error = ENOTTY;
2335 			goto unlock;
2336 		}
2337 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2338 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2339 			noff = bn * bsize;
2340 			if (noff < *off)
2341 				noff = *off;
2342 			goto unlock;
2343 		}
2344 	}
2345 	if (noff > va.va_size)
2346 		noff = va.va_size;
2347 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2348 	if (cmd == FIOSEEKDATA)
2349 		error = ENXIO;
2350 unlock:
2351 	VOP_UNLOCK(vp);
2352 	if (error == 0)
2353 		*off = noff;
2354 	return (error);
2355 }
2356 
2357 int
2358 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2359 {
2360 	struct ucred *cred;
2361 	struct vnode *vp;
2362 	struct vattr vattr;
2363 	off_t foffset, size;
2364 	int error, noneg;
2365 
2366 	cred = td->td_ucred;
2367 	vp = fp->f_vnode;
2368 	foffset = foffset_lock(fp, 0);
2369 	noneg = (vp->v_type != VCHR);
2370 	error = 0;
2371 	switch (whence) {
2372 	case L_INCR:
2373 		if (noneg &&
2374 		    (foffset < 0 ||
2375 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2376 			error = EOVERFLOW;
2377 			break;
2378 		}
2379 		offset += foffset;
2380 		break;
2381 	case L_XTND:
2382 		vn_lock(vp, LK_SHARED | LK_RETRY);
2383 		error = VOP_GETATTR(vp, &vattr, cred);
2384 		VOP_UNLOCK(vp);
2385 		if (error)
2386 			break;
2387 
2388 		/*
2389 		 * If the file references a disk device, then fetch
2390 		 * the media size and use that to determine the ending
2391 		 * offset.
2392 		 */
2393 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2394 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2395 			vattr.va_size = size;
2396 		if (noneg &&
2397 		    (vattr.va_size > OFF_MAX ||
2398 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2399 			error = EOVERFLOW;
2400 			break;
2401 		}
2402 		offset += vattr.va_size;
2403 		break;
2404 	case L_SET:
2405 		break;
2406 	case SEEK_DATA:
2407 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2408 		if (error == ENOTTY)
2409 			error = EINVAL;
2410 		break;
2411 	case SEEK_HOLE:
2412 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2413 		if (error == ENOTTY)
2414 			error = EINVAL;
2415 		break;
2416 	default:
2417 		error = EINVAL;
2418 	}
2419 	if (error == 0 && noneg && offset < 0)
2420 		error = EINVAL;
2421 	if (error != 0)
2422 		goto drop;
2423 	VFS_KNOTE_UNLOCKED(vp, 0);
2424 	td->td_uretoff.tdu_off = offset;
2425 drop:
2426 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2427 	return (error);
2428 }
2429 
2430 int
2431 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2432     struct thread *td)
2433 {
2434 	int error;
2435 
2436 	/*
2437 	 * Grant permission if the caller is the owner of the file, or
2438 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2439 	 * on the file.  If the time pointer is null, then write
2440 	 * permission on the file is also sufficient.
2441 	 *
2442 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2443 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2444 	 * will be allowed to set the times [..] to the current
2445 	 * server time.
2446 	 */
2447 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2448 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2449 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2450 	return (error);
2451 }
2452 
2453 int
2454 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2455 {
2456 	struct vnode *vp;
2457 	int error;
2458 
2459 	if (fp->f_type == DTYPE_FIFO)
2460 		kif->kf_type = KF_TYPE_FIFO;
2461 	else
2462 		kif->kf_type = KF_TYPE_VNODE;
2463 	vp = fp->f_vnode;
2464 	vref(vp);
2465 	FILEDESC_SUNLOCK(fdp);
2466 	error = vn_fill_kinfo_vnode(vp, kif);
2467 	vrele(vp);
2468 	FILEDESC_SLOCK(fdp);
2469 	return (error);
2470 }
2471 
2472 static inline void
2473 vn_fill_junk(struct kinfo_file *kif)
2474 {
2475 	size_t len, olen;
2476 
2477 	/*
2478 	 * Simulate vn_fullpath returning changing values for a given
2479 	 * vp during e.g. coredump.
2480 	 */
2481 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2482 	olen = strlen(kif->kf_path);
2483 	if (len < olen)
2484 		strcpy(&kif->kf_path[len - 1], "$");
2485 	else
2486 		for (; olen < len; olen++)
2487 			strcpy(&kif->kf_path[olen], "A");
2488 }
2489 
2490 int
2491 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2492 {
2493 	struct vattr va;
2494 	char *fullpath, *freepath;
2495 	int error;
2496 
2497 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2498 	freepath = NULL;
2499 	fullpath = "-";
2500 	error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2501 	if (error == 0) {
2502 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2503 	}
2504 	if (freepath != NULL)
2505 		free(freepath, M_TEMP);
2506 
2507 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2508 		vn_fill_junk(kif);
2509 	);
2510 
2511 	/*
2512 	 * Retrieve vnode attributes.
2513 	 */
2514 	va.va_fsid = VNOVAL;
2515 	va.va_rdev = NODEV;
2516 	vn_lock(vp, LK_SHARED | LK_RETRY);
2517 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2518 	VOP_UNLOCK(vp);
2519 	if (error != 0)
2520 		return (error);
2521 	if (va.va_fsid != VNOVAL)
2522 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2523 	else
2524 		kif->kf_un.kf_file.kf_file_fsid =
2525 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2526 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2527 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2528 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2529 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2530 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2531 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2532 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2533 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2534 	return (0);
2535 }
2536 
2537 int
2538 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2539     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2540     struct thread *td)
2541 {
2542 #ifdef HWPMC_HOOKS
2543 	struct pmckern_map_in pkm;
2544 #endif
2545 	struct mount *mp;
2546 	struct vnode *vp;
2547 	vm_object_t object;
2548 	vm_prot_t maxprot;
2549 	boolean_t writecounted;
2550 	int error;
2551 
2552 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2553     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2554 	/*
2555 	 * POSIX shared-memory objects are defined to have
2556 	 * kernel persistence, and are not defined to support
2557 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2558 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2559 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2560 	 * flag to request this behavior.
2561 	 */
2562 	if ((fp->f_flag & FPOSIXSHM) != 0)
2563 		flags |= MAP_NOSYNC;
2564 #endif
2565 	vp = fp->f_vnode;
2566 
2567 	/*
2568 	 * Ensure that file and memory protections are
2569 	 * compatible.  Note that we only worry about
2570 	 * writability if mapping is shared; in this case,
2571 	 * current and max prot are dictated by the open file.
2572 	 * XXX use the vnode instead?  Problem is: what
2573 	 * credentials do we use for determination? What if
2574 	 * proc does a setuid?
2575 	 */
2576 	mp = vp->v_mount;
2577 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2578 		maxprot = VM_PROT_NONE;
2579 		if ((prot & VM_PROT_EXECUTE) != 0)
2580 			return (EACCES);
2581 	} else
2582 		maxprot = VM_PROT_EXECUTE;
2583 	if ((fp->f_flag & FREAD) != 0)
2584 		maxprot |= VM_PROT_READ;
2585 	else if ((prot & VM_PROT_READ) != 0)
2586 		return (EACCES);
2587 
2588 	/*
2589 	 * If we are sharing potential changes via MAP_SHARED and we
2590 	 * are trying to get write permission although we opened it
2591 	 * without asking for it, bail out.
2592 	 */
2593 	if ((flags & MAP_SHARED) != 0) {
2594 		if ((fp->f_flag & FWRITE) != 0)
2595 			maxprot |= VM_PROT_WRITE;
2596 		else if ((prot & VM_PROT_WRITE) != 0)
2597 			return (EACCES);
2598 	} else {
2599 		maxprot |= VM_PROT_WRITE;
2600 		cap_maxprot |= VM_PROT_WRITE;
2601 	}
2602 	maxprot &= cap_maxprot;
2603 
2604 	/*
2605 	 * For regular files and shared memory, POSIX requires that
2606 	 * the value of foff be a legitimate offset within the data
2607 	 * object.  In particular, negative offsets are invalid.
2608 	 * Blocking negative offsets and overflows here avoids
2609 	 * possible wraparound or user-level access into reserved
2610 	 * ranges of the data object later.  In contrast, POSIX does
2611 	 * not dictate how offsets are used by device drivers, so in
2612 	 * the case of a device mapping a negative offset is passed
2613 	 * on.
2614 	 */
2615 	if (
2616 #ifdef _LP64
2617 	    size > OFF_MAX ||
2618 #endif
2619 	    foff < 0 || foff > OFF_MAX - size)
2620 		return (EINVAL);
2621 
2622 	writecounted = FALSE;
2623 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2624 	    &foff, &object, &writecounted);
2625 	if (error != 0)
2626 		return (error);
2627 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2628 	    foff, writecounted, td);
2629 	if (error != 0) {
2630 		/*
2631 		 * If this mapping was accounted for in the vnode's
2632 		 * writecount, then undo that now.
2633 		 */
2634 		if (writecounted)
2635 			vm_pager_release_writecount(object, 0, size);
2636 		vm_object_deallocate(object);
2637 	}
2638 #ifdef HWPMC_HOOKS
2639 	/* Inform hwpmc(4) if an executable is being mapped. */
2640 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2641 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2642 			pkm.pm_file = vp;
2643 			pkm.pm_address = (uintptr_t) *addr;
2644 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2645 		}
2646 	}
2647 #endif
2648 	return (error);
2649 }
2650 
2651 void
2652 vn_fsid(struct vnode *vp, struct vattr *va)
2653 {
2654 	fsid_t *f;
2655 
2656 	f = &vp->v_mount->mnt_stat.f_fsid;
2657 	va->va_fsid = (uint32_t)f->val[1];
2658 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2659 	va->va_fsid += (uint32_t)f->val[0];
2660 }
2661 
2662 int
2663 vn_fsync_buf(struct vnode *vp, int waitfor)
2664 {
2665 	struct buf *bp, *nbp;
2666 	struct bufobj *bo;
2667 	struct mount *mp;
2668 	int error, maxretry;
2669 
2670 	error = 0;
2671 	maxretry = 10000;     /* large, arbitrarily chosen */
2672 	mp = NULL;
2673 	if (vp->v_type == VCHR) {
2674 		VI_LOCK(vp);
2675 		mp = vp->v_rdev->si_mountpt;
2676 		VI_UNLOCK(vp);
2677 	}
2678 	bo = &vp->v_bufobj;
2679 	BO_LOCK(bo);
2680 loop1:
2681 	/*
2682 	 * MARK/SCAN initialization to avoid infinite loops.
2683 	 */
2684         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2685 		bp->b_vflags &= ~BV_SCANNED;
2686 		bp->b_error = 0;
2687 	}
2688 
2689 	/*
2690 	 * Flush all dirty buffers associated with a vnode.
2691 	 */
2692 loop2:
2693 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2694 		if ((bp->b_vflags & BV_SCANNED) != 0)
2695 			continue;
2696 		bp->b_vflags |= BV_SCANNED;
2697 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2698 			if (waitfor != MNT_WAIT)
2699 				continue;
2700 			if (BUF_LOCK(bp,
2701 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2702 			    BO_LOCKPTR(bo)) != 0) {
2703 				BO_LOCK(bo);
2704 				goto loop1;
2705 			}
2706 			BO_LOCK(bo);
2707 		}
2708 		BO_UNLOCK(bo);
2709 		KASSERT(bp->b_bufobj == bo,
2710 		    ("bp %p wrong b_bufobj %p should be %p",
2711 		    bp, bp->b_bufobj, bo));
2712 		if ((bp->b_flags & B_DELWRI) == 0)
2713 			panic("fsync: not dirty");
2714 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2715 			vfs_bio_awrite(bp);
2716 		} else {
2717 			bremfree(bp);
2718 			bawrite(bp);
2719 		}
2720 		if (maxretry < 1000)
2721 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2722 		BO_LOCK(bo);
2723 		goto loop2;
2724 	}
2725 
2726 	/*
2727 	 * If synchronous the caller expects us to completely resolve all
2728 	 * dirty buffers in the system.  Wait for in-progress I/O to
2729 	 * complete (which could include background bitmap writes), then
2730 	 * retry if dirty blocks still exist.
2731 	 */
2732 	if (waitfor == MNT_WAIT) {
2733 		bufobj_wwait(bo, 0, 0);
2734 		if (bo->bo_dirty.bv_cnt > 0) {
2735 			/*
2736 			 * If we are unable to write any of these buffers
2737 			 * then we fail now rather than trying endlessly
2738 			 * to write them out.
2739 			 */
2740 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2741 				if ((error = bp->b_error) != 0)
2742 					break;
2743 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2744 			    (error == 0 && --maxretry >= 0))
2745 				goto loop1;
2746 			if (error == 0)
2747 				error = EAGAIN;
2748 		}
2749 	}
2750 	BO_UNLOCK(bo);
2751 	if (error != 0)
2752 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2753 
2754 	return (error);
2755 }
2756 
2757 /*
2758  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2759  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2760  * to do the actual copy.
2761  * vn_generic_copy_file_range() is factored out, so it can be called
2762  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2763  * different file systems.
2764  */
2765 int
2766 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2767     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2768     struct ucred *outcred, struct thread *fsize_td)
2769 {
2770 	int error;
2771 	size_t len;
2772 	uint64_t uvalin, uvalout;
2773 
2774 	len = *lenp;
2775 	*lenp = 0;		/* For error returns. */
2776 	error = 0;
2777 
2778 	/* Do some sanity checks on the arguments. */
2779 	uvalin = *inoffp;
2780 	uvalin += len;
2781 	uvalout = *outoffp;
2782 	uvalout += len;
2783 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2784 		error = EISDIR;
2785 	else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2786 	    (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2787 	    uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2788 	    outvp->v_type != VREG)
2789 		error = EINVAL;
2790 	if (error != 0)
2791 		goto out;
2792 
2793 	/*
2794 	 * If the two vnode are for the same file system, call
2795 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2796 	 * which can handle copies across multiple file systems.
2797 	 */
2798 	*lenp = len;
2799 	if (invp->v_mount == outvp->v_mount)
2800 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2801 		    lenp, flags, incred, outcred, fsize_td);
2802 	else
2803 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2804 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2805 out:
2806 	return (error);
2807 }
2808 
2809 /*
2810  * Test len bytes of data starting at dat for all bytes == 0.
2811  * Return true if all bytes are zero, false otherwise.
2812  * Expects dat to be well aligned.
2813  */
2814 static bool
2815 mem_iszero(void *dat, int len)
2816 {
2817 	int i;
2818 	const u_int *p;
2819 	const char *cp;
2820 
2821 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2822 		if (len >= sizeof(*p)) {
2823 			if (*p != 0)
2824 				return (false);
2825 		} else {
2826 			cp = (const char *)p;
2827 			for (i = 0; i < len; i++, cp++)
2828 				if (*cp != '\0')
2829 					return (false);
2830 		}
2831 	}
2832 	return (true);
2833 }
2834 
2835 /*
2836  * Look for a hole in the output file and, if found, adjust *outoffp
2837  * and *xferp to skip past the hole.
2838  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2839  * to be written as 0's upon return.
2840  */
2841 static off_t
2842 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2843     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2844 {
2845 	int error;
2846 	off_t delta;
2847 
2848 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2849 		*dataoffp = *outoffp;
2850 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2851 		    curthread);
2852 		if (error == 0) {
2853 			*holeoffp = *dataoffp;
2854 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2855 			    curthread);
2856 		}
2857 		if (error != 0 || *holeoffp == *dataoffp) {
2858 			/*
2859 			 * Since outvp is unlocked, it may be possible for
2860 			 * another thread to do a truncate(), lseek(), write()
2861 			 * creating a hole at startoff between the above
2862 			 * VOP_IOCTL() calls, if the other thread does not do
2863 			 * rangelocking.
2864 			 * If that happens, *holeoffp == *dataoffp and finding
2865 			 * the hole has failed, so disable vn_skip_hole().
2866 			 */
2867 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2868 			return (xfer2);
2869 		}
2870 		KASSERT(*dataoffp >= *outoffp,
2871 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2872 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2873 		KASSERT(*holeoffp > *dataoffp,
2874 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2875 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2876 	}
2877 
2878 	/*
2879 	 * If there is a hole before the data starts, advance *outoffp and
2880 	 * *xferp past the hole.
2881 	 */
2882 	if (*dataoffp > *outoffp) {
2883 		delta = *dataoffp - *outoffp;
2884 		if (delta >= *xferp) {
2885 			/* Entire *xferp is a hole. */
2886 			*outoffp += *xferp;
2887 			*xferp = 0;
2888 			return (0);
2889 		}
2890 		*xferp -= delta;
2891 		*outoffp += delta;
2892 		xfer2 = MIN(xfer2, *xferp);
2893 	}
2894 
2895 	/*
2896 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2897 	 * that the write ends at the start of the hole.
2898 	 * *holeoffp should always be greater than *outoffp, but for the
2899 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2900 	 * value.
2901 	 */
2902 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2903 		xfer2 = *holeoffp - *outoffp;
2904 	return (xfer2);
2905 }
2906 
2907 /*
2908  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2909  * dat is a maximum of blksize in length and can be written repeatedly in
2910  * the chunk.
2911  * If growfile == true, just grow the file via vn_truncate_locked() instead
2912  * of doing actual writes.
2913  * If checkhole == true, a hole is being punched, so skip over any hole
2914  * already in the output file.
2915  */
2916 static int
2917 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2918     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2919 {
2920 	struct mount *mp;
2921 	off_t dataoff, holeoff, xfer2;
2922 	int error, lckf;
2923 
2924 	/*
2925 	 * Loop around doing writes of blksize until write has been completed.
2926 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2927 	 * done for each iteration, since the xfer argument can be very
2928 	 * large if there is a large hole to punch in the output file.
2929 	 */
2930 	error = 0;
2931 	holeoff = 0;
2932 	do {
2933 		xfer2 = MIN(xfer, blksize);
2934 		if (checkhole) {
2935 			/*
2936 			 * Punching a hole.  Skip writing if there is
2937 			 * already a hole in the output file.
2938 			 */
2939 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2940 			    &dataoff, &holeoff, cred);
2941 			if (xfer == 0)
2942 				break;
2943 			if (holeoff < 0)
2944 				checkhole = false;
2945 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2946 			    (intmax_t)xfer2));
2947 		}
2948 		bwillwrite();
2949 		mp = NULL;
2950 		error = vn_start_write(outvp, &mp, V_WAIT);
2951 		if (error == 0) {
2952 			if (MNT_SHARED_WRITES(mp))
2953 				lckf = LK_SHARED;
2954 			else
2955 				lckf = LK_EXCLUSIVE;
2956 			error = vn_lock(outvp, lckf);
2957 		}
2958 		if (error == 0) {
2959 			if (growfile)
2960 				error = vn_truncate_locked(outvp, outoff + xfer,
2961 				    false, cred);
2962 			else {
2963 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2964 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
2965 				    curthread->td_ucred, cred, NULL, curthread);
2966 				outoff += xfer2;
2967 				xfer -= xfer2;
2968 			}
2969 			VOP_UNLOCK(outvp);
2970 		}
2971 		if (mp != NULL)
2972 			vn_finished_write(mp);
2973 	} while (!growfile && xfer > 0 && error == 0);
2974 	return (error);
2975 }
2976 
2977 /*
2978  * Copy a byte range of one file to another.  This function can handle the
2979  * case where invp and outvp are on different file systems.
2980  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2981  * is no better file system specific way to do it.
2982  */
2983 int
2984 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2985     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
2986     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
2987 {
2988 	struct vattr va;
2989 	struct mount *mp;
2990 	struct uio io;
2991 	off_t startoff, endoff, xfer, xfer2;
2992 	u_long blksize;
2993 	int error;
2994 	bool cantseek, readzeros, eof, lastblock;
2995 	ssize_t aresid;
2996 	size_t copylen, len, savlen;
2997 	char *dat;
2998 	long holein, holeout;
2999 
3000 	holein = holeout = 0;
3001 	savlen = len = *lenp;
3002 	error = 0;
3003 	dat = NULL;
3004 
3005 	error = vn_lock(invp, LK_SHARED);
3006 	if (error != 0)
3007 		goto out;
3008 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3009 		holein = 0;
3010 	VOP_UNLOCK(invp);
3011 
3012 	mp = NULL;
3013 	error = vn_start_write(outvp, &mp, V_WAIT);
3014 	if (error == 0)
3015 		error = vn_lock(outvp, LK_EXCLUSIVE);
3016 	if (error == 0) {
3017 		/*
3018 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3019 		 * now that outvp is locked.
3020 		 */
3021 		if (fsize_td != NULL) {
3022 			io.uio_offset = *outoffp;
3023 			io.uio_resid = len;
3024 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3025 			if (error != 0)
3026 				error = EFBIG;
3027 		}
3028 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3029 			holeout = 0;
3030 		/*
3031 		 * Holes that are past EOF do not need to be written as a block
3032 		 * of zero bytes.  So, truncate the output file as far as
3033 		 * possible and then use va.va_size to decide if writing 0
3034 		 * bytes is necessary in the loop below.
3035 		 */
3036 		if (error == 0)
3037 			error = VOP_GETATTR(outvp, &va, outcred);
3038 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3039 		    *outoffp + len) {
3040 #ifdef MAC
3041 			error = mac_vnode_check_write(curthread->td_ucred,
3042 			    outcred, outvp);
3043 			if (error == 0)
3044 #endif
3045 				error = vn_truncate_locked(outvp, *outoffp,
3046 				    false, outcred);
3047 			if (error == 0)
3048 				va.va_size = *outoffp;
3049 		}
3050 		VOP_UNLOCK(outvp);
3051 	}
3052 	if (mp != NULL)
3053 		vn_finished_write(mp);
3054 	if (error != 0)
3055 		goto out;
3056 
3057 	/*
3058 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3059 	 * If hole sizes aren't available, set the blksize to the larger
3060 	 * f_iosize of invp and outvp.
3061 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3062 	 * This value is clipped at 4Kbytes and 1Mbyte.
3063 	 */
3064 	blksize = MAX(holein, holeout);
3065 	if (blksize == 0)
3066 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3067 		    outvp->v_mount->mnt_stat.f_iosize);
3068 	if (blksize < 4096)
3069 		blksize = 4096;
3070 	else if (blksize > 1024 * 1024)
3071 		blksize = 1024 * 1024;
3072 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3073 
3074 	/*
3075 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3076 	 * to find holes.  Otherwise, just scan the read block for all 0s
3077 	 * in the inner loop where the data copying is done.
3078 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3079 	 * support holes on the server, but do not support FIOSEEKHOLE.
3080 	 */
3081 	eof = false;
3082 	while (len > 0 && error == 0 && !eof) {
3083 		endoff = 0;			/* To shut up compilers. */
3084 		cantseek = true;
3085 		startoff = *inoffp;
3086 		copylen = len;
3087 
3088 		/*
3089 		 * Find the next data area.  If there is just a hole to EOF,
3090 		 * FIOSEEKDATA should fail and then we drop down into the
3091 		 * inner loop and create the hole on the outvp file.
3092 		 * (I do not know if any file system will report a hole to
3093 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3094 		 *  will fail for those file systems.)
3095 		 *
3096 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3097 		 * the code just falls through to the inner copy loop.
3098 		 */
3099 		error = EINVAL;
3100 		if (holein > 0)
3101 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3102 			    incred, curthread);
3103 		if (error == 0) {
3104 			endoff = startoff;
3105 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3106 			    incred, curthread);
3107 			/*
3108 			 * Since invp is unlocked, it may be possible for
3109 			 * another thread to do a truncate(), lseek(), write()
3110 			 * creating a hole at startoff between the above
3111 			 * VOP_IOCTL() calls, if the other thread does not do
3112 			 * rangelocking.
3113 			 * If that happens, startoff == endoff and finding
3114 			 * the hole has failed, so set an error.
3115 			 */
3116 			if (error == 0 && startoff == endoff)
3117 				error = EINVAL; /* Any error. Reset to 0. */
3118 		}
3119 		if (error == 0) {
3120 			if (startoff > *inoffp) {
3121 				/* Found hole before data block. */
3122 				xfer = MIN(startoff - *inoffp, len);
3123 				if (*outoffp < va.va_size) {
3124 					/* Must write 0s to punch hole. */
3125 					xfer2 = MIN(va.va_size - *outoffp,
3126 					    xfer);
3127 					memset(dat, 0, MIN(xfer2, blksize));
3128 					error = vn_write_outvp(outvp, dat,
3129 					    *outoffp, xfer2, blksize, false,
3130 					    holeout > 0, outcred);
3131 				}
3132 
3133 				if (error == 0 && *outoffp + xfer >
3134 				    va.va_size && xfer == len)
3135 					/* Grow last block. */
3136 					error = vn_write_outvp(outvp, dat,
3137 					    *outoffp, xfer, blksize, true,
3138 					    false, outcred);
3139 				if (error == 0) {
3140 					*inoffp += xfer;
3141 					*outoffp += xfer;
3142 					len -= xfer;
3143 				}
3144 			}
3145 			copylen = MIN(len, endoff - startoff);
3146 			cantseek = false;
3147 		} else {
3148 			cantseek = true;
3149 			startoff = *inoffp;
3150 			copylen = len;
3151 			error = 0;
3152 		}
3153 
3154 		xfer = blksize;
3155 		if (cantseek) {
3156 			/*
3157 			 * Set first xfer to end at a block boundary, so that
3158 			 * holes are more likely detected in the loop below via
3159 			 * the for all bytes 0 method.
3160 			 */
3161 			xfer -= (*inoffp % blksize);
3162 		}
3163 		/* Loop copying the data block. */
3164 		while (copylen > 0 && error == 0 && !eof) {
3165 			if (copylen < xfer)
3166 				xfer = copylen;
3167 			error = vn_lock(invp, LK_SHARED);
3168 			if (error != 0)
3169 				goto out;
3170 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3171 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3172 			    curthread->td_ucred, incred, &aresid,
3173 			    curthread);
3174 			VOP_UNLOCK(invp);
3175 			lastblock = false;
3176 			if (error == 0 && aresid > 0) {
3177 				/* Stop the copy at EOF on the input file. */
3178 				xfer -= aresid;
3179 				eof = true;
3180 				lastblock = true;
3181 			}
3182 			if (error == 0) {
3183 				/*
3184 				 * Skip the write for holes past the initial EOF
3185 				 * of the output file, unless this is the last
3186 				 * write of the output file at EOF.
3187 				 */
3188 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3189 				    false;
3190 				if (xfer == len)
3191 					lastblock = true;
3192 				if (!cantseek || *outoffp < va.va_size ||
3193 				    lastblock || !readzeros)
3194 					error = vn_write_outvp(outvp, dat,
3195 					    *outoffp, xfer, blksize,
3196 					    readzeros && lastblock &&
3197 					    *outoffp >= va.va_size, false,
3198 					    outcred);
3199 				if (error == 0) {
3200 					*inoffp += xfer;
3201 					startoff += xfer;
3202 					*outoffp += xfer;
3203 					copylen -= xfer;
3204 					len -= xfer;
3205 				}
3206 			}
3207 			xfer = blksize;
3208 		}
3209 	}
3210 out:
3211 	*lenp = savlen - len;
3212 	free(dat, M_TEMP);
3213 	return (error);
3214 }
3215 
3216 static int
3217 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3218 {
3219 	struct mount *mp;
3220 	struct vnode *vp;
3221 	off_t olen, ooffset;
3222 	int error;
3223 #ifdef AUDIT
3224 	int audited_vnode1 = 0;
3225 #endif
3226 
3227 	vp = fp->f_vnode;
3228 	if (vp->v_type != VREG)
3229 		return (ENODEV);
3230 
3231 	/* Allocating blocks may take a long time, so iterate. */
3232 	for (;;) {
3233 		olen = len;
3234 		ooffset = offset;
3235 
3236 		bwillwrite();
3237 		mp = NULL;
3238 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3239 		if (error != 0)
3240 			break;
3241 		error = vn_lock(vp, LK_EXCLUSIVE);
3242 		if (error != 0) {
3243 			vn_finished_write(mp);
3244 			break;
3245 		}
3246 #ifdef AUDIT
3247 		if (!audited_vnode1) {
3248 			AUDIT_ARG_VNODE1(vp);
3249 			audited_vnode1 = 1;
3250 		}
3251 #endif
3252 #ifdef MAC
3253 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3254 		if (error == 0)
3255 #endif
3256 			error = VOP_ALLOCATE(vp, &offset, &len);
3257 		VOP_UNLOCK(vp);
3258 		vn_finished_write(mp);
3259 
3260 		if (olen + ooffset != offset + len) {
3261 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3262 			    ooffset, olen, offset, len);
3263 		}
3264 		if (error != 0 || len == 0)
3265 			break;
3266 		KASSERT(olen > len, ("Iteration did not make progress?"));
3267 		maybe_yield();
3268 	}
3269 
3270 	return (error);
3271 }
3272