xref: /freebsd/sys/kern/vfs_vnops.c (revision 59e50df3cd1493537cfa916daf3c51a01b7ff06e)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/sx.h>
74 #include <sys/sysctl.h>
75 #include <sys/ttycom.h>
76 #include <sys/conf.h>
77 #include <sys/syslog.h>
78 #include <sys/unistd.h>
79 #include <sys/user.h>
80 
81 #include <security/audit/audit.h>
82 #include <security/mac/mac_framework.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_extern.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vnode_pager.h>
91 
92 #ifdef HWPMC_HOOKS
93 #include <sys/pmckern.h>
94 #endif
95 
96 static fo_rdwr_t	vn_read;
97 static fo_rdwr_t	vn_write;
98 static fo_rdwr_t	vn_io_fault;
99 static fo_truncate_t	vn_truncate;
100 static fo_ioctl_t	vn_ioctl;
101 static fo_poll_t	vn_poll;
102 static fo_kqfilter_t	vn_kqfilter;
103 static fo_stat_t	vn_statfile;
104 static fo_close_t	vn_closefile;
105 static fo_mmap_t	vn_mmap;
106 
107 struct 	fileops vnops = {
108 	.fo_read = vn_io_fault,
109 	.fo_write = vn_io_fault,
110 	.fo_truncate = vn_truncate,
111 	.fo_ioctl = vn_ioctl,
112 	.fo_poll = vn_poll,
113 	.fo_kqfilter = vn_kqfilter,
114 	.fo_stat = vn_statfile,
115 	.fo_close = vn_closefile,
116 	.fo_chmod = vn_chmod,
117 	.fo_chown = vn_chown,
118 	.fo_sendfile = vn_sendfile,
119 	.fo_seek = vn_seek,
120 	.fo_fill_kinfo = vn_fill_kinfo,
121 	.fo_mmap = vn_mmap,
122 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
123 };
124 
125 static const int io_hold_cnt = 16;
126 static int vn_io_fault_enable = 1;
127 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
128     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
129 static int vn_io_fault_prefault = 0;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
131     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
132 static u_long vn_io_faults_cnt;
133 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
134     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
135 
136 /*
137  * Returns true if vn_io_fault mode of handling the i/o request should
138  * be used.
139  */
140 static bool
141 do_vn_io_fault(struct vnode *vp, struct uio *uio)
142 {
143 	struct mount *mp;
144 
145 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
146 	    (mp = vp->v_mount) != NULL &&
147 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
148 }
149 
150 /*
151  * Structure used to pass arguments to vn_io_fault1(), to do either
152  * file- or vnode-based I/O calls.
153  */
154 struct vn_io_fault_args {
155 	enum {
156 		VN_IO_FAULT_FOP,
157 		VN_IO_FAULT_VOP
158 	} kind;
159 	struct ucred *cred;
160 	int flags;
161 	union {
162 		struct fop_args_tag {
163 			struct file *fp;
164 			fo_rdwr_t *doio;
165 		} fop_args;
166 		struct vop_args_tag {
167 			struct vnode *vp;
168 		} vop_args;
169 	} args;
170 };
171 
172 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
173     struct vn_io_fault_args *args, struct thread *td);
174 
175 int
176 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
177 {
178 	struct thread *td = ndp->ni_cnd.cn_thread;
179 
180 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
181 }
182 
183 /*
184  * Common code for vnode open operations via a name lookup.
185  * Lookup the vnode and invoke VOP_CREATE if needed.
186  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
187  *
188  * Note that this does NOT free nameidata for the successful case,
189  * due to the NDINIT being done elsewhere.
190  */
191 int
192 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
193     struct ucred *cred, struct file *fp)
194 {
195 	struct vnode *vp;
196 	struct mount *mp;
197 	struct thread *td = ndp->ni_cnd.cn_thread;
198 	struct vattr vat;
199 	struct vattr *vap = &vat;
200 	int fmode, error;
201 
202 restart:
203 	fmode = *flagp;
204 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
205 	    O_EXCL | O_DIRECTORY))
206 		return (EINVAL);
207 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
208 		ndp->ni_cnd.cn_nameiop = CREATE;
209 		/*
210 		 * Set NOCACHE to avoid flushing the cache when
211 		 * rolling in many files at once.
212 		*/
213 		ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
214 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
215 			ndp->ni_cnd.cn_flags |= FOLLOW;
216 		if ((fmode & O_BENEATH) != 0)
217 			ndp->ni_cnd.cn_flags |= BENEATH;
218 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
219 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
220 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
221 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
222 		bwillwrite();
223 		if ((error = namei(ndp)) != 0)
224 			return (error);
225 		if (ndp->ni_vp == NULL) {
226 			VATTR_NULL(vap);
227 			vap->va_type = VREG;
228 			vap->va_mode = cmode;
229 			if (fmode & O_EXCL)
230 				vap->va_vaflags |= VA_EXCLUSIVE;
231 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
232 				NDFREE(ndp, NDF_ONLY_PNBUF);
233 				vput(ndp->ni_dvp);
234 				if ((error = vn_start_write(NULL, &mp,
235 				    V_XSLEEP | PCATCH)) != 0)
236 					return (error);
237 				goto restart;
238 			}
239 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
240 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
241 #ifdef MAC
242 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
243 			    &ndp->ni_cnd, vap);
244 			if (error == 0)
245 #endif
246 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
247 						   &ndp->ni_cnd, vap);
248 			vput(ndp->ni_dvp);
249 			vn_finished_write(mp);
250 			if (error) {
251 				NDFREE(ndp, NDF_ONLY_PNBUF);
252 				return (error);
253 			}
254 			fmode &= ~O_TRUNC;
255 			vp = ndp->ni_vp;
256 		} else {
257 			if (ndp->ni_dvp == ndp->ni_vp)
258 				vrele(ndp->ni_dvp);
259 			else
260 				vput(ndp->ni_dvp);
261 			ndp->ni_dvp = NULL;
262 			vp = ndp->ni_vp;
263 			if (fmode & O_EXCL) {
264 				error = EEXIST;
265 				goto bad;
266 			}
267 			fmode &= ~O_CREAT;
268 		}
269 	} else {
270 		ndp->ni_cnd.cn_nameiop = LOOKUP;
271 		ndp->ni_cnd.cn_flags = ISOPEN |
272 		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
273 		if (!(fmode & FWRITE))
274 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
275 		if ((fmode & O_BENEATH) != 0)
276 			ndp->ni_cnd.cn_flags |= BENEATH;
277 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
278 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
279 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
280 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
281 		if ((error = namei(ndp)) != 0)
282 			return (error);
283 		vp = ndp->ni_vp;
284 	}
285 	error = vn_open_vnode(vp, fmode, cred, td, fp);
286 	if (error)
287 		goto bad;
288 	*flagp = fmode;
289 	return (0);
290 bad:
291 	NDFREE(ndp, NDF_ONLY_PNBUF);
292 	vput(vp);
293 	*flagp = fmode;
294 	ndp->ni_vp = NULL;
295 	return (error);
296 }
297 
298 static int
299 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
300 {
301 	struct flock lf;
302 	int error, lock_flags, type;
303 
304 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
305 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
306 		return (0);
307 	KASSERT(fp != NULL, ("open with flock requires fp"));
308 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
309 		return (EOPNOTSUPP);
310 
311 	lock_flags = VOP_ISLOCKED(vp);
312 	VOP_UNLOCK(vp, 0);
313 
314 	lf.l_whence = SEEK_SET;
315 	lf.l_start = 0;
316 	lf.l_len = 0;
317 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
318 	type = F_FLOCK;
319 	if ((fmode & FNONBLOCK) == 0)
320 		type |= F_WAIT;
321 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
322 	if (error == 0)
323 		fp->f_flag |= FHASLOCK;
324 
325 	vn_lock(vp, lock_flags | LK_RETRY);
326 	if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0)
327 		error = ENOENT;
328 	return (error);
329 }
330 
331 /*
332  * Common code for vnode open operations once a vnode is located.
333  * Check permissions, and call the VOP_OPEN routine.
334  */
335 int
336 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
337     struct thread *td, struct file *fp)
338 {
339 	accmode_t accmode;
340 	int error;
341 
342 	if (vp->v_type == VLNK)
343 		return (EMLINK);
344 	if (vp->v_type == VSOCK)
345 		return (EOPNOTSUPP);
346 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
347 		return (ENOTDIR);
348 	accmode = 0;
349 	if (fmode & (FWRITE | O_TRUNC)) {
350 		if (vp->v_type == VDIR)
351 			return (EISDIR);
352 		accmode |= VWRITE;
353 	}
354 	if (fmode & FREAD)
355 		accmode |= VREAD;
356 	if (fmode & FEXEC)
357 		accmode |= VEXEC;
358 	if ((fmode & O_APPEND) && (fmode & FWRITE))
359 		accmode |= VAPPEND;
360 #ifdef MAC
361 	if (fmode & O_CREAT)
362 		accmode |= VCREAT;
363 	if (fmode & O_VERIFY)
364 		accmode |= VVERIFY;
365 	error = mac_vnode_check_open(cred, vp, accmode);
366 	if (error)
367 		return (error);
368 
369 	accmode &= ~(VCREAT | VVERIFY);
370 #endif
371 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
372 		error = VOP_ACCESS(vp, accmode, cred, td);
373 		if (error != 0)
374 			return (error);
375 	}
376 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
377 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
378 	error = VOP_OPEN(vp, fmode, cred, td, fp);
379 	if (error != 0)
380 		return (error);
381 
382 	error = vn_open_vnode_advlock(vp, fmode, fp);
383 	if (error == 0 && (fmode & FWRITE) != 0) {
384 		error = VOP_ADD_WRITECOUNT(vp, 1);
385 		if (error == 0) {
386 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
387 			     __func__, vp, vp->v_writecount);
388 		}
389 	}
390 
391 	/*
392 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
393 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
394 	 * Arrange for that by having fdrop() to use vn_closefile().
395 	 */
396 	if (error != 0) {
397 		fp->f_flag |= FOPENFAILED;
398 		fp->f_vnode = vp;
399 		if (fp->f_ops == &badfileops) {
400 			fp->f_type = DTYPE_VNODE;
401 			fp->f_ops = &vnops;
402 		}
403 		vref(vp);
404 	}
405 
406 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
407 	return (error);
408 
409 }
410 
411 /*
412  * Check for write permissions on the specified vnode.
413  * Prototype text segments cannot be written.
414  * It is racy.
415  */
416 int
417 vn_writechk(struct vnode *vp)
418 {
419 
420 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
421 	/*
422 	 * If there's shared text associated with
423 	 * the vnode, try to free it up once.  If
424 	 * we fail, we can't allow writing.
425 	 */
426 	if (VOP_IS_TEXT(vp))
427 		return (ETXTBSY);
428 
429 	return (0);
430 }
431 
432 /*
433  * Vnode close call
434  */
435 static int
436 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
437     struct thread *td, bool keep_ref)
438 {
439 	struct mount *mp;
440 	int error, lock_flags;
441 
442 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
443 	    MNT_EXTENDED_SHARED(vp->v_mount))
444 		lock_flags = LK_SHARED;
445 	else
446 		lock_flags = LK_EXCLUSIVE;
447 
448 	vn_start_write(vp, &mp, V_WAIT);
449 	vn_lock(vp, lock_flags | LK_RETRY);
450 	AUDIT_ARG_VNODE1(vp);
451 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
452 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
453 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
454 		    __func__, vp, vp->v_writecount);
455 	}
456 	error = VOP_CLOSE(vp, flags, file_cred, td);
457 	if (keep_ref)
458 		VOP_UNLOCK(vp, 0);
459 	else
460 		vput(vp);
461 	vn_finished_write(mp);
462 	return (error);
463 }
464 
465 int
466 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
467     struct thread *td)
468 {
469 
470 	return (vn_close1(vp, flags, file_cred, td, false));
471 }
472 
473 /*
474  * Heuristic to detect sequential operation.
475  */
476 static int
477 sequential_heuristic(struct uio *uio, struct file *fp)
478 {
479 
480 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
481 	if (fp->f_flag & FRDAHEAD)
482 		return (fp->f_seqcount << IO_SEQSHIFT);
483 
484 	/*
485 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
486 	 * that the first I/O is normally considered to be slightly
487 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
488 	 * unless previous seeks have reduced f_seqcount to 0, in which
489 	 * case offset 0 is not special.
490 	 */
491 	if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
492 	    uio->uio_offset == fp->f_nextoff) {
493 		/*
494 		 * f_seqcount is in units of fixed-size blocks so that it
495 		 * depends mainly on the amount of sequential I/O and not
496 		 * much on the number of sequential I/O's.  The fixed size
497 		 * of 16384 is hard-coded here since it is (not quite) just
498 		 * a magic size that works well here.  This size is more
499 		 * closely related to the best I/O size for real disks than
500 		 * to any block size used by software.
501 		 */
502 		if (uio->uio_resid >= IO_SEQMAX * 16384)
503 			fp->f_seqcount = IO_SEQMAX;
504 		else {
505 			fp->f_seqcount += howmany(uio->uio_resid, 16384);
506 			if (fp->f_seqcount > IO_SEQMAX)
507 				fp->f_seqcount = IO_SEQMAX;
508 		}
509 		return (fp->f_seqcount << IO_SEQSHIFT);
510 	}
511 
512 	/* Not sequential.  Quickly draw-down sequentiality. */
513 	if (fp->f_seqcount > 1)
514 		fp->f_seqcount = 1;
515 	else
516 		fp->f_seqcount = 0;
517 	return (0);
518 }
519 
520 /*
521  * Package up an I/O request on a vnode into a uio and do it.
522  */
523 int
524 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
525     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
526     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
527 {
528 	struct uio auio;
529 	struct iovec aiov;
530 	struct mount *mp;
531 	struct ucred *cred;
532 	void *rl_cookie;
533 	struct vn_io_fault_args args;
534 	int error, lock_flags;
535 
536 	if (offset < 0 && vp->v_type != VCHR)
537 		return (EINVAL);
538 	auio.uio_iov = &aiov;
539 	auio.uio_iovcnt = 1;
540 	aiov.iov_base = base;
541 	aiov.iov_len = len;
542 	auio.uio_resid = len;
543 	auio.uio_offset = offset;
544 	auio.uio_segflg = segflg;
545 	auio.uio_rw = rw;
546 	auio.uio_td = td;
547 	error = 0;
548 
549 	if ((ioflg & IO_NODELOCKED) == 0) {
550 		if ((ioflg & IO_RANGELOCKED) == 0) {
551 			if (rw == UIO_READ) {
552 				rl_cookie = vn_rangelock_rlock(vp, offset,
553 				    offset + len);
554 			} else {
555 				rl_cookie = vn_rangelock_wlock(vp, offset,
556 				    offset + len);
557 			}
558 		} else
559 			rl_cookie = NULL;
560 		mp = NULL;
561 		if (rw == UIO_WRITE) {
562 			if (vp->v_type != VCHR &&
563 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
564 			    != 0)
565 				goto out;
566 			if (MNT_SHARED_WRITES(mp) ||
567 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
568 				lock_flags = LK_SHARED;
569 			else
570 				lock_flags = LK_EXCLUSIVE;
571 		} else
572 			lock_flags = LK_SHARED;
573 		vn_lock(vp, lock_flags | LK_RETRY);
574 	} else
575 		rl_cookie = NULL;
576 
577 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
578 #ifdef MAC
579 	if ((ioflg & IO_NOMACCHECK) == 0) {
580 		if (rw == UIO_READ)
581 			error = mac_vnode_check_read(active_cred, file_cred,
582 			    vp);
583 		else
584 			error = mac_vnode_check_write(active_cred, file_cred,
585 			    vp);
586 	}
587 #endif
588 	if (error == 0) {
589 		if (file_cred != NULL)
590 			cred = file_cred;
591 		else
592 			cred = active_cred;
593 		if (do_vn_io_fault(vp, &auio)) {
594 			args.kind = VN_IO_FAULT_VOP;
595 			args.cred = cred;
596 			args.flags = ioflg;
597 			args.args.vop_args.vp = vp;
598 			error = vn_io_fault1(vp, &auio, &args, td);
599 		} else if (rw == UIO_READ) {
600 			error = VOP_READ(vp, &auio, ioflg, cred);
601 		} else /* if (rw == UIO_WRITE) */ {
602 			error = VOP_WRITE(vp, &auio, ioflg, cred);
603 		}
604 	}
605 	if (aresid)
606 		*aresid = auio.uio_resid;
607 	else
608 		if (auio.uio_resid && error == 0)
609 			error = EIO;
610 	if ((ioflg & IO_NODELOCKED) == 0) {
611 		VOP_UNLOCK(vp, 0);
612 		if (mp != NULL)
613 			vn_finished_write(mp);
614 	}
615  out:
616 	if (rl_cookie != NULL)
617 		vn_rangelock_unlock(vp, rl_cookie);
618 	return (error);
619 }
620 
621 /*
622  * Package up an I/O request on a vnode into a uio and do it.  The I/O
623  * request is split up into smaller chunks and we try to avoid saturating
624  * the buffer cache while potentially holding a vnode locked, so we
625  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
626  * to give other processes a chance to lock the vnode (either other processes
627  * core'ing the same binary, or unrelated processes scanning the directory).
628  */
629 int
630 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
631     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
632     struct ucred *file_cred, size_t *aresid, struct thread *td)
633 {
634 	int error = 0;
635 	ssize_t iaresid;
636 
637 	do {
638 		int chunk;
639 
640 		/*
641 		 * Force `offset' to a multiple of MAXBSIZE except possibly
642 		 * for the first chunk, so that filesystems only need to
643 		 * write full blocks except possibly for the first and last
644 		 * chunks.
645 		 */
646 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
647 
648 		if (chunk > len)
649 			chunk = len;
650 		if (rw != UIO_READ && vp->v_type == VREG)
651 			bwillwrite();
652 		iaresid = 0;
653 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
654 		    ioflg, active_cred, file_cred, &iaresid, td);
655 		len -= chunk;	/* aresid calc already includes length */
656 		if (error)
657 			break;
658 		offset += chunk;
659 		base = (char *)base + chunk;
660 		kern_yield(PRI_USER);
661 	} while (len);
662 	if (aresid)
663 		*aresid = len + iaresid;
664 	return (error);
665 }
666 
667 off_t
668 foffset_lock(struct file *fp, int flags)
669 {
670 	struct mtx *mtxp;
671 	off_t res;
672 
673 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
674 
675 #if OFF_MAX <= LONG_MAX
676 	/*
677 	 * Caller only wants the current f_offset value.  Assume that
678 	 * the long and shorter integer types reads are atomic.
679 	 */
680 	if ((flags & FOF_NOLOCK) != 0)
681 		return (fp->f_offset);
682 #endif
683 
684 	/*
685 	 * According to McKusick the vn lock was protecting f_offset here.
686 	 * It is now protected by the FOFFSET_LOCKED flag.
687 	 */
688 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
689 	mtx_lock(mtxp);
690 	if ((flags & FOF_NOLOCK) == 0) {
691 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
692 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
693 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
694 			    "vofflock", 0);
695 		}
696 		fp->f_vnread_flags |= FOFFSET_LOCKED;
697 	}
698 	res = fp->f_offset;
699 	mtx_unlock(mtxp);
700 	return (res);
701 }
702 
703 void
704 foffset_unlock(struct file *fp, off_t val, int flags)
705 {
706 	struct mtx *mtxp;
707 
708 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
709 
710 #if OFF_MAX <= LONG_MAX
711 	if ((flags & FOF_NOLOCK) != 0) {
712 		if ((flags & FOF_NOUPDATE) == 0)
713 			fp->f_offset = val;
714 		if ((flags & FOF_NEXTOFF) != 0)
715 			fp->f_nextoff = val;
716 		return;
717 	}
718 #endif
719 
720 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
721 	mtx_lock(mtxp);
722 	if ((flags & FOF_NOUPDATE) == 0)
723 		fp->f_offset = val;
724 	if ((flags & FOF_NEXTOFF) != 0)
725 		fp->f_nextoff = val;
726 	if ((flags & FOF_NOLOCK) == 0) {
727 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
728 		    ("Lost FOFFSET_LOCKED"));
729 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
730 			wakeup(&fp->f_vnread_flags);
731 		fp->f_vnread_flags = 0;
732 	}
733 	mtx_unlock(mtxp);
734 }
735 
736 void
737 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
738 {
739 
740 	if ((flags & FOF_OFFSET) == 0)
741 		uio->uio_offset = foffset_lock(fp, flags);
742 }
743 
744 void
745 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
746 {
747 
748 	if ((flags & FOF_OFFSET) == 0)
749 		foffset_unlock(fp, uio->uio_offset, flags);
750 }
751 
752 static int
753 get_advice(struct file *fp, struct uio *uio)
754 {
755 	struct mtx *mtxp;
756 	int ret;
757 
758 	ret = POSIX_FADV_NORMAL;
759 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
760 		return (ret);
761 
762 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
763 	mtx_lock(mtxp);
764 	if (fp->f_advice != NULL &&
765 	    uio->uio_offset >= fp->f_advice->fa_start &&
766 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
767 		ret = fp->f_advice->fa_advice;
768 	mtx_unlock(mtxp);
769 	return (ret);
770 }
771 
772 /*
773  * File table vnode read routine.
774  */
775 static int
776 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
777     struct thread *td)
778 {
779 	struct vnode *vp;
780 	off_t orig_offset;
781 	int error, ioflag;
782 	int advice;
783 
784 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
785 	    uio->uio_td, td));
786 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
787 	vp = fp->f_vnode;
788 	ioflag = 0;
789 	if (fp->f_flag & FNONBLOCK)
790 		ioflag |= IO_NDELAY;
791 	if (fp->f_flag & O_DIRECT)
792 		ioflag |= IO_DIRECT;
793 	advice = get_advice(fp, uio);
794 	vn_lock(vp, LK_SHARED | LK_RETRY);
795 
796 	switch (advice) {
797 	case POSIX_FADV_NORMAL:
798 	case POSIX_FADV_SEQUENTIAL:
799 	case POSIX_FADV_NOREUSE:
800 		ioflag |= sequential_heuristic(uio, fp);
801 		break;
802 	case POSIX_FADV_RANDOM:
803 		/* Disable read-ahead for random I/O. */
804 		break;
805 	}
806 	orig_offset = uio->uio_offset;
807 
808 #ifdef MAC
809 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
810 	if (error == 0)
811 #endif
812 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
813 	fp->f_nextoff = uio->uio_offset;
814 	VOP_UNLOCK(vp, 0);
815 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
816 	    orig_offset != uio->uio_offset)
817 		/*
818 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
819 		 * for the backing file after a POSIX_FADV_NOREUSE
820 		 * read(2).
821 		 */
822 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
823 		    POSIX_FADV_DONTNEED);
824 	return (error);
825 }
826 
827 /*
828  * File table vnode write routine.
829  */
830 static int
831 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
832     struct thread *td)
833 {
834 	struct vnode *vp;
835 	struct mount *mp;
836 	off_t orig_offset;
837 	int error, ioflag, lock_flags;
838 	int advice;
839 
840 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
841 	    uio->uio_td, td));
842 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
843 	vp = fp->f_vnode;
844 	if (vp->v_type == VREG)
845 		bwillwrite();
846 	ioflag = IO_UNIT;
847 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
848 		ioflag |= IO_APPEND;
849 	if (fp->f_flag & FNONBLOCK)
850 		ioflag |= IO_NDELAY;
851 	if (fp->f_flag & O_DIRECT)
852 		ioflag |= IO_DIRECT;
853 	if ((fp->f_flag & O_FSYNC) ||
854 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
855 		ioflag |= IO_SYNC;
856 	mp = NULL;
857 	if (vp->v_type != VCHR &&
858 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
859 		goto unlock;
860 
861 	advice = get_advice(fp, uio);
862 
863 	if (MNT_SHARED_WRITES(mp) ||
864 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
865 		lock_flags = LK_SHARED;
866 	} else {
867 		lock_flags = LK_EXCLUSIVE;
868 	}
869 
870 	vn_lock(vp, lock_flags | LK_RETRY);
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 		/* XXX: Is this correct? */
879 		break;
880 	}
881 	orig_offset = uio->uio_offset;
882 
883 #ifdef MAC
884 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
885 	if (error == 0)
886 #endif
887 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
888 	fp->f_nextoff = uio->uio_offset;
889 	VOP_UNLOCK(vp, 0);
890 	if (vp->v_type != VCHR)
891 		vn_finished_write(mp);
892 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
893 	    orig_offset != uio->uio_offset)
894 		/*
895 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
896 		 * for the backing file after a POSIX_FADV_NOREUSE
897 		 * write(2).
898 		 */
899 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
900 		    POSIX_FADV_DONTNEED);
901 unlock:
902 	return (error);
903 }
904 
905 /*
906  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
907  * prevent the following deadlock:
908  *
909  * Assume that the thread A reads from the vnode vp1 into userspace
910  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
911  * currently not resident, then system ends up with the call chain
912  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
913  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
914  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
915  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
916  * backed by the pages of vnode vp1, and some page in buf2 is not
917  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
918  *
919  * To prevent the lock order reversal and deadlock, vn_io_fault() does
920  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
921  * Instead, it first tries to do the whole range i/o with pagefaults
922  * disabled. If all pages in the i/o buffer are resident and mapped,
923  * VOP will succeed (ignoring the genuine filesystem errors).
924  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
925  * i/o in chunks, with all pages in the chunk prefaulted and held
926  * using vm_fault_quick_hold_pages().
927  *
928  * Filesystems using this deadlock avoidance scheme should use the
929  * array of the held pages from uio, saved in the curthread->td_ma,
930  * instead of doing uiomove().  A helper function
931  * vn_io_fault_uiomove() converts uiomove request into
932  * uiomove_fromphys() over td_ma array.
933  *
934  * Since vnode locks do not cover the whole i/o anymore, rangelocks
935  * make the current i/o request atomic with respect to other i/os and
936  * truncations.
937  */
938 
939 /*
940  * Decode vn_io_fault_args and perform the corresponding i/o.
941  */
942 static int
943 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
944     struct thread *td)
945 {
946 	int error, save;
947 
948 	error = 0;
949 	save = vm_fault_disable_pagefaults();
950 	switch (args->kind) {
951 	case VN_IO_FAULT_FOP:
952 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
953 		    uio, args->cred, args->flags, td);
954 		break;
955 	case VN_IO_FAULT_VOP:
956 		if (uio->uio_rw == UIO_READ) {
957 			error = VOP_READ(args->args.vop_args.vp, uio,
958 			    args->flags, args->cred);
959 		} else if (uio->uio_rw == UIO_WRITE) {
960 			error = VOP_WRITE(args->args.vop_args.vp, uio,
961 			    args->flags, args->cred);
962 		}
963 		break;
964 	default:
965 		panic("vn_io_fault_doio: unknown kind of io %d %d",
966 		    args->kind, uio->uio_rw);
967 	}
968 	vm_fault_enable_pagefaults(save);
969 	return (error);
970 }
971 
972 static int
973 vn_io_fault_touch(char *base, const struct uio *uio)
974 {
975 	int r;
976 
977 	r = fubyte(base);
978 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
979 		return (EFAULT);
980 	return (0);
981 }
982 
983 static int
984 vn_io_fault_prefault_user(const struct uio *uio)
985 {
986 	char *base;
987 	const struct iovec *iov;
988 	size_t len;
989 	ssize_t resid;
990 	int error, i;
991 
992 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
993 	    ("vn_io_fault_prefault userspace"));
994 
995 	error = i = 0;
996 	iov = uio->uio_iov;
997 	resid = uio->uio_resid;
998 	base = iov->iov_base;
999 	len = iov->iov_len;
1000 	while (resid > 0) {
1001 		error = vn_io_fault_touch(base, uio);
1002 		if (error != 0)
1003 			break;
1004 		if (len < PAGE_SIZE) {
1005 			if (len != 0) {
1006 				error = vn_io_fault_touch(base + len - 1, uio);
1007 				if (error != 0)
1008 					break;
1009 				resid -= len;
1010 			}
1011 			if (++i >= uio->uio_iovcnt)
1012 				break;
1013 			iov = uio->uio_iov + i;
1014 			base = iov->iov_base;
1015 			len = iov->iov_len;
1016 		} else {
1017 			len -= PAGE_SIZE;
1018 			base += PAGE_SIZE;
1019 			resid -= PAGE_SIZE;
1020 		}
1021 	}
1022 	return (error);
1023 }
1024 
1025 /*
1026  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1027  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1028  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1029  * into args and call vn_io_fault1() to handle faults during the user
1030  * mode buffer accesses.
1031  */
1032 static int
1033 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1034     struct thread *td)
1035 {
1036 	vm_page_t ma[io_hold_cnt + 2];
1037 	struct uio *uio_clone, short_uio;
1038 	struct iovec short_iovec[1];
1039 	vm_page_t *prev_td_ma;
1040 	vm_prot_t prot;
1041 	vm_offset_t addr, end;
1042 	size_t len, resid;
1043 	ssize_t adv;
1044 	int error, cnt, saveheld, prev_td_ma_cnt;
1045 
1046 	if (vn_io_fault_prefault) {
1047 		error = vn_io_fault_prefault_user(uio);
1048 		if (error != 0)
1049 			return (error); /* Or ignore ? */
1050 	}
1051 
1052 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1053 
1054 	/*
1055 	 * The UFS follows IO_UNIT directive and replays back both
1056 	 * uio_offset and uio_resid if an error is encountered during the
1057 	 * operation.  But, since the iovec may be already advanced,
1058 	 * uio is still in an inconsistent state.
1059 	 *
1060 	 * Cache a copy of the original uio, which is advanced to the redo
1061 	 * point using UIO_NOCOPY below.
1062 	 */
1063 	uio_clone = cloneuio(uio);
1064 	resid = uio->uio_resid;
1065 
1066 	short_uio.uio_segflg = UIO_USERSPACE;
1067 	short_uio.uio_rw = uio->uio_rw;
1068 	short_uio.uio_td = uio->uio_td;
1069 
1070 	error = vn_io_fault_doio(args, uio, td);
1071 	if (error != EFAULT)
1072 		goto out;
1073 
1074 	atomic_add_long(&vn_io_faults_cnt, 1);
1075 	uio_clone->uio_segflg = UIO_NOCOPY;
1076 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1077 	uio_clone->uio_segflg = uio->uio_segflg;
1078 
1079 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1080 	prev_td_ma = td->td_ma;
1081 	prev_td_ma_cnt = td->td_ma_cnt;
1082 
1083 	while (uio_clone->uio_resid != 0) {
1084 		len = uio_clone->uio_iov->iov_len;
1085 		if (len == 0) {
1086 			KASSERT(uio_clone->uio_iovcnt >= 1,
1087 			    ("iovcnt underflow"));
1088 			uio_clone->uio_iov++;
1089 			uio_clone->uio_iovcnt--;
1090 			continue;
1091 		}
1092 		if (len > io_hold_cnt * PAGE_SIZE)
1093 			len = io_hold_cnt * PAGE_SIZE;
1094 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1095 		end = round_page(addr + len);
1096 		if (end < addr) {
1097 			error = EFAULT;
1098 			break;
1099 		}
1100 		cnt = atop(end - trunc_page(addr));
1101 		/*
1102 		 * A perfectly misaligned address and length could cause
1103 		 * both the start and the end of the chunk to use partial
1104 		 * page.  +2 accounts for such a situation.
1105 		 */
1106 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1107 		    addr, len, prot, ma, io_hold_cnt + 2);
1108 		if (cnt == -1) {
1109 			error = EFAULT;
1110 			break;
1111 		}
1112 		short_uio.uio_iov = &short_iovec[0];
1113 		short_iovec[0].iov_base = (void *)addr;
1114 		short_uio.uio_iovcnt = 1;
1115 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1116 		short_uio.uio_offset = uio_clone->uio_offset;
1117 		td->td_ma = ma;
1118 		td->td_ma_cnt = cnt;
1119 
1120 		error = vn_io_fault_doio(args, &short_uio, td);
1121 		vm_page_unhold_pages(ma, cnt);
1122 		adv = len - short_uio.uio_resid;
1123 
1124 		uio_clone->uio_iov->iov_base =
1125 		    (char *)uio_clone->uio_iov->iov_base + adv;
1126 		uio_clone->uio_iov->iov_len -= adv;
1127 		uio_clone->uio_resid -= adv;
1128 		uio_clone->uio_offset += adv;
1129 
1130 		uio->uio_resid -= adv;
1131 		uio->uio_offset += adv;
1132 
1133 		if (error != 0 || adv == 0)
1134 			break;
1135 	}
1136 	td->td_ma = prev_td_ma;
1137 	td->td_ma_cnt = prev_td_ma_cnt;
1138 	curthread_pflags_restore(saveheld);
1139 out:
1140 	free(uio_clone, M_IOV);
1141 	return (error);
1142 }
1143 
1144 static int
1145 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1146     int flags, struct thread *td)
1147 {
1148 	fo_rdwr_t *doio;
1149 	struct vnode *vp;
1150 	void *rl_cookie;
1151 	struct vn_io_fault_args args;
1152 	int error;
1153 
1154 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1155 	vp = fp->f_vnode;
1156 	foffset_lock_uio(fp, uio, flags);
1157 	if (do_vn_io_fault(vp, uio)) {
1158 		args.kind = VN_IO_FAULT_FOP;
1159 		args.args.fop_args.fp = fp;
1160 		args.args.fop_args.doio = doio;
1161 		args.cred = active_cred;
1162 		args.flags = flags | FOF_OFFSET;
1163 		if (uio->uio_rw == UIO_READ) {
1164 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1165 			    uio->uio_offset + uio->uio_resid);
1166 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1167 		    (flags & FOF_OFFSET) == 0) {
1168 			/* For appenders, punt and lock the whole range. */
1169 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1170 		} else {
1171 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1172 			    uio->uio_offset + uio->uio_resid);
1173 		}
1174 		error = vn_io_fault1(vp, uio, &args, td);
1175 		vn_rangelock_unlock(vp, rl_cookie);
1176 	} else {
1177 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1178 	}
1179 	foffset_unlock_uio(fp, uio, flags);
1180 	return (error);
1181 }
1182 
1183 /*
1184  * Helper function to perform the requested uiomove operation using
1185  * the held pages for io->uio_iov[0].iov_base buffer instead of
1186  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1187  * instead of iov_base prevents page faults that could occur due to
1188  * pmap_collect() invalidating the mapping created by
1189  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1190  * object cleanup revoking the write access from page mappings.
1191  *
1192  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1193  * instead of plain uiomove().
1194  */
1195 int
1196 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1197 {
1198 	struct uio transp_uio;
1199 	struct iovec transp_iov[1];
1200 	struct thread *td;
1201 	size_t adv;
1202 	int error, pgadv;
1203 
1204 	td = curthread;
1205 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1206 	    uio->uio_segflg != UIO_USERSPACE)
1207 		return (uiomove(data, xfersize, uio));
1208 
1209 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1210 	transp_iov[0].iov_base = data;
1211 	transp_uio.uio_iov = &transp_iov[0];
1212 	transp_uio.uio_iovcnt = 1;
1213 	if (xfersize > uio->uio_resid)
1214 		xfersize = uio->uio_resid;
1215 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1216 	transp_uio.uio_offset = 0;
1217 	transp_uio.uio_segflg = UIO_SYSSPACE;
1218 	/*
1219 	 * Since transp_iov points to data, and td_ma page array
1220 	 * corresponds to original uio->uio_iov, we need to invert the
1221 	 * direction of the i/o operation as passed to
1222 	 * uiomove_fromphys().
1223 	 */
1224 	switch (uio->uio_rw) {
1225 	case UIO_WRITE:
1226 		transp_uio.uio_rw = UIO_READ;
1227 		break;
1228 	case UIO_READ:
1229 		transp_uio.uio_rw = UIO_WRITE;
1230 		break;
1231 	}
1232 	transp_uio.uio_td = uio->uio_td;
1233 	error = uiomove_fromphys(td->td_ma,
1234 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1235 	    xfersize, &transp_uio);
1236 	adv = xfersize - transp_uio.uio_resid;
1237 	pgadv =
1238 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1239 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1240 	td->td_ma += pgadv;
1241 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1242 	    pgadv));
1243 	td->td_ma_cnt -= pgadv;
1244 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1245 	uio->uio_iov->iov_len -= adv;
1246 	uio->uio_resid -= adv;
1247 	uio->uio_offset += adv;
1248 	return (error);
1249 }
1250 
1251 int
1252 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1253     struct uio *uio)
1254 {
1255 	struct thread *td;
1256 	vm_offset_t iov_base;
1257 	int cnt, pgadv;
1258 
1259 	td = curthread;
1260 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1261 	    uio->uio_segflg != UIO_USERSPACE)
1262 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1263 
1264 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1265 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1266 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1267 	switch (uio->uio_rw) {
1268 	case UIO_WRITE:
1269 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1270 		    offset, cnt);
1271 		break;
1272 	case UIO_READ:
1273 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1274 		    cnt);
1275 		break;
1276 	}
1277 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1278 	td->td_ma += pgadv;
1279 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1280 	    pgadv));
1281 	td->td_ma_cnt -= pgadv;
1282 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1283 	uio->uio_iov->iov_len -= cnt;
1284 	uio->uio_resid -= cnt;
1285 	uio->uio_offset += cnt;
1286 	return (0);
1287 }
1288 
1289 
1290 /*
1291  * File table truncate routine.
1292  */
1293 static int
1294 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1295     struct thread *td)
1296 {
1297 	struct mount *mp;
1298 	struct vnode *vp;
1299 	void *rl_cookie;
1300 	int error;
1301 
1302 	vp = fp->f_vnode;
1303 
1304 	/*
1305 	 * Lock the whole range for truncation.  Otherwise split i/o
1306 	 * might happen partly before and partly after the truncation.
1307 	 */
1308 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1309 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1310 	if (error)
1311 		goto out1;
1312 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1313 	AUDIT_ARG_VNODE1(vp);
1314 	if (vp->v_type == VDIR) {
1315 		error = EISDIR;
1316 		goto out;
1317 	}
1318 #ifdef MAC
1319 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1320 	if (error)
1321 		goto out;
1322 #endif
1323 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1324 	    fp->f_cred);
1325 out:
1326 	VOP_UNLOCK(vp, 0);
1327 	vn_finished_write(mp);
1328 out1:
1329 	vn_rangelock_unlock(vp, rl_cookie);
1330 	return (error);
1331 }
1332 
1333 /*
1334  * Truncate a file that is already locked.
1335  */
1336 int
1337 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1338     struct ucred *cred)
1339 {
1340 	struct vattr vattr;
1341 	int error;
1342 
1343 	error = VOP_ADD_WRITECOUNT(vp, 1);
1344 	if (error == 0) {
1345 		VATTR_NULL(&vattr);
1346 		vattr.va_size = length;
1347 		if (sync)
1348 			vattr.va_vaflags |= VA_SYNC;
1349 		error = VOP_SETATTR(vp, &vattr, cred);
1350 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1351 	}
1352 	return (error);
1353 }
1354 
1355 /*
1356  * File table vnode stat routine.
1357  */
1358 static int
1359 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1360     struct thread *td)
1361 {
1362 	struct vnode *vp = fp->f_vnode;
1363 	int error;
1364 
1365 	vn_lock(vp, LK_SHARED | LK_RETRY);
1366 	error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1367 	VOP_UNLOCK(vp, 0);
1368 
1369 	return (error);
1370 }
1371 
1372 /*
1373  * Stat a vnode; implementation for the stat syscall
1374  */
1375 int
1376 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1377     struct ucred *file_cred, struct thread *td)
1378 {
1379 	struct vattr vattr;
1380 	struct vattr *vap;
1381 	int error;
1382 	u_short mode;
1383 
1384 	AUDIT_ARG_VNODE1(vp);
1385 #ifdef MAC
1386 	error = mac_vnode_check_stat(active_cred, file_cred, vp);
1387 	if (error)
1388 		return (error);
1389 #endif
1390 
1391 	vap = &vattr;
1392 
1393 	/*
1394 	 * Initialize defaults for new and unusual fields, so that file
1395 	 * systems which don't support these fields don't need to know
1396 	 * about them.
1397 	 */
1398 	vap->va_birthtime.tv_sec = -1;
1399 	vap->va_birthtime.tv_nsec = 0;
1400 	vap->va_fsid = VNOVAL;
1401 	vap->va_rdev = NODEV;
1402 
1403 	error = VOP_GETATTR(vp, vap, active_cred);
1404 	if (error)
1405 		return (error);
1406 
1407 	/*
1408 	 * Zero the spare stat fields
1409 	 */
1410 	bzero(sb, sizeof *sb);
1411 
1412 	/*
1413 	 * Copy from vattr table
1414 	 */
1415 	if (vap->va_fsid != VNOVAL)
1416 		sb->st_dev = vap->va_fsid;
1417 	else
1418 		sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1419 	sb->st_ino = vap->va_fileid;
1420 	mode = vap->va_mode;
1421 	switch (vap->va_type) {
1422 	case VREG:
1423 		mode |= S_IFREG;
1424 		break;
1425 	case VDIR:
1426 		mode |= S_IFDIR;
1427 		break;
1428 	case VBLK:
1429 		mode |= S_IFBLK;
1430 		break;
1431 	case VCHR:
1432 		mode |= S_IFCHR;
1433 		break;
1434 	case VLNK:
1435 		mode |= S_IFLNK;
1436 		break;
1437 	case VSOCK:
1438 		mode |= S_IFSOCK;
1439 		break;
1440 	case VFIFO:
1441 		mode |= S_IFIFO;
1442 		break;
1443 	default:
1444 		return (EBADF);
1445 	}
1446 	sb->st_mode = mode;
1447 	sb->st_nlink = vap->va_nlink;
1448 	sb->st_uid = vap->va_uid;
1449 	sb->st_gid = vap->va_gid;
1450 	sb->st_rdev = vap->va_rdev;
1451 	if (vap->va_size > OFF_MAX)
1452 		return (EOVERFLOW);
1453 	sb->st_size = vap->va_size;
1454 	sb->st_atim = vap->va_atime;
1455 	sb->st_mtim = vap->va_mtime;
1456 	sb->st_ctim = vap->va_ctime;
1457 	sb->st_birthtim = vap->va_birthtime;
1458 
1459         /*
1460 	 * According to www.opengroup.org, the meaning of st_blksize is
1461 	 *   "a filesystem-specific preferred I/O block size for this
1462 	 *    object.  In some filesystem types, this may vary from file
1463 	 *    to file"
1464 	 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1465 	 */
1466 
1467 	sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1468 
1469 	sb->st_flags = vap->va_flags;
1470 	if (priv_check(td, PRIV_VFS_GENERATION))
1471 		sb->st_gen = 0;
1472 	else
1473 		sb->st_gen = vap->va_gen;
1474 
1475 	sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1476 	return (0);
1477 }
1478 
1479 /*
1480  * File table vnode ioctl routine.
1481  */
1482 static int
1483 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1484     struct thread *td)
1485 {
1486 	struct vattr vattr;
1487 	struct vnode *vp;
1488 	struct fiobmap2_arg *bmarg;
1489 	int error;
1490 
1491 	vp = fp->f_vnode;
1492 	switch (vp->v_type) {
1493 	case VDIR:
1494 	case VREG:
1495 		switch (com) {
1496 		case FIONREAD:
1497 			vn_lock(vp, LK_SHARED | LK_RETRY);
1498 			error = VOP_GETATTR(vp, &vattr, active_cred);
1499 			VOP_UNLOCK(vp, 0);
1500 			if (error == 0)
1501 				*(int *)data = vattr.va_size - fp->f_offset;
1502 			return (error);
1503 		case FIOBMAP2:
1504 			bmarg = (struct fiobmap2_arg *)data;
1505 			vn_lock(vp, LK_SHARED | LK_RETRY);
1506 #ifdef MAC
1507 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1508 			    vp);
1509 			if (error == 0)
1510 #endif
1511 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1512 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1513 			VOP_UNLOCK(vp, 0);
1514 			return (error);
1515 		case FIONBIO:
1516 		case FIOASYNC:
1517 			return (0);
1518 		default:
1519 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1520 			    active_cred, td));
1521 		}
1522 		break;
1523 	case VCHR:
1524 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1525 		    active_cred, td));
1526 	default:
1527 		return (ENOTTY);
1528 	}
1529 }
1530 
1531 /*
1532  * File table vnode poll routine.
1533  */
1534 static int
1535 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1536     struct thread *td)
1537 {
1538 	struct vnode *vp;
1539 	int error;
1540 
1541 	vp = fp->f_vnode;
1542 #ifdef MAC
1543 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1544 	AUDIT_ARG_VNODE1(vp);
1545 	error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1546 	VOP_UNLOCK(vp, 0);
1547 	if (!error)
1548 #endif
1549 
1550 	error = VOP_POLL(vp, events, fp->f_cred, td);
1551 	return (error);
1552 }
1553 
1554 /*
1555  * Acquire the requested lock and then check for validity.  LK_RETRY
1556  * permits vn_lock to return doomed vnodes.
1557  */
1558 int
1559 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1560 {
1561 	int error;
1562 
1563 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1564 	    ("vn_lock: no locktype"));
1565 	VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count"));
1566 retry:
1567 	error = VOP_LOCK1(vp, flags, file, line);
1568 	flags &= ~LK_INTERLOCK;	/* Interlock is always dropped. */
1569 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1570 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1571 
1572 	if ((flags & LK_RETRY) == 0) {
1573 		if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0) {
1574 			VOP_UNLOCK(vp, 0);
1575 			error = ENOENT;
1576 		}
1577 	} else if (error != 0)
1578 		goto retry;
1579 	return (error);
1580 }
1581 
1582 /*
1583  * File table vnode close routine.
1584  */
1585 static int
1586 vn_closefile(struct file *fp, struct thread *td)
1587 {
1588 	struct vnode *vp;
1589 	struct flock lf;
1590 	int error;
1591 	bool ref;
1592 
1593 	vp = fp->f_vnode;
1594 	fp->f_ops = &badfileops;
1595 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1596 
1597 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1598 
1599 	if (__predict_false(ref)) {
1600 		lf.l_whence = SEEK_SET;
1601 		lf.l_start = 0;
1602 		lf.l_len = 0;
1603 		lf.l_type = F_UNLCK;
1604 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1605 		vrele(vp);
1606 	}
1607 	return (error);
1608 }
1609 
1610 static bool
1611 vn_suspendable(struct mount *mp)
1612 {
1613 
1614 	return (mp->mnt_op->vfs_susp_clean != NULL);
1615 }
1616 
1617 /*
1618  * Preparing to start a filesystem write operation. If the operation is
1619  * permitted, then we bump the count of operations in progress and
1620  * proceed. If a suspend request is in progress, we wait until the
1621  * suspension is over, and then proceed.
1622  */
1623 static int
1624 vn_start_write_locked(struct mount *mp, int flags)
1625 {
1626 	int error, mflags;
1627 
1628 	mtx_assert(MNT_MTX(mp), MA_OWNED);
1629 	error = 0;
1630 
1631 	/*
1632 	 * Check on status of suspension.
1633 	 */
1634 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1635 	    mp->mnt_susp_owner != curthread) {
1636 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1637 		    (flags & PCATCH) : 0) | (PUSER - 1);
1638 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1639 			if (flags & V_NOWAIT) {
1640 				error = EWOULDBLOCK;
1641 				goto unlock;
1642 			}
1643 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1644 			    "suspfs", 0);
1645 			if (error)
1646 				goto unlock;
1647 		}
1648 	}
1649 	if (flags & V_XSLEEP)
1650 		goto unlock;
1651 	mp->mnt_writeopcount++;
1652 unlock:
1653 	if (error != 0 || (flags & V_XSLEEP) != 0)
1654 		MNT_REL(mp);
1655 	MNT_IUNLOCK(mp);
1656 	return (error);
1657 }
1658 
1659 int
1660 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1661 {
1662 	struct mount *mp;
1663 	int error;
1664 
1665 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1666 	    ("V_MNTREF requires mp"));
1667 
1668 	error = 0;
1669 	/*
1670 	 * If a vnode is provided, get and return the mount point that
1671 	 * to which it will write.
1672 	 */
1673 	if (vp != NULL) {
1674 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1675 			*mpp = NULL;
1676 			if (error != EOPNOTSUPP)
1677 				return (error);
1678 			return (0);
1679 		}
1680 	}
1681 	if ((mp = *mpp) == NULL)
1682 		return (0);
1683 
1684 	if (!vn_suspendable(mp)) {
1685 		if (vp != NULL || (flags & V_MNTREF) != 0)
1686 			vfs_rel(mp);
1687 		return (0);
1688 	}
1689 
1690 	/*
1691 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1692 	 * a vfs_ref().
1693 	 * As long as a vnode is not provided we need to acquire a
1694 	 * refcount for the provided mountpoint too, in order to
1695 	 * emulate a vfs_ref().
1696 	 */
1697 	MNT_ILOCK(mp);
1698 	if (vp == NULL && (flags & V_MNTREF) == 0)
1699 		MNT_REF(mp);
1700 
1701 	return (vn_start_write_locked(mp, flags));
1702 }
1703 
1704 /*
1705  * Secondary suspension. Used by operations such as vop_inactive
1706  * routines that are needed by the higher level functions. These
1707  * are allowed to proceed until all the higher level functions have
1708  * completed (indicated by mnt_writeopcount dropping to zero). At that
1709  * time, these operations are halted until the suspension is over.
1710  */
1711 int
1712 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1713 {
1714 	struct mount *mp;
1715 	int error;
1716 
1717 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1718 	    ("V_MNTREF requires mp"));
1719 
1720  retry:
1721 	if (vp != NULL) {
1722 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1723 			*mpp = NULL;
1724 			if (error != EOPNOTSUPP)
1725 				return (error);
1726 			return (0);
1727 		}
1728 	}
1729 	/*
1730 	 * If we are not suspended or have not yet reached suspended
1731 	 * mode, then let the operation proceed.
1732 	 */
1733 	if ((mp = *mpp) == NULL)
1734 		return (0);
1735 
1736 	if (!vn_suspendable(mp)) {
1737 		if (vp != NULL || (flags & V_MNTREF) != 0)
1738 			vfs_rel(mp);
1739 		return (0);
1740 	}
1741 
1742 	/*
1743 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1744 	 * a vfs_ref().
1745 	 * As long as a vnode is not provided we need to acquire a
1746 	 * refcount for the provided mountpoint too, in order to
1747 	 * emulate a vfs_ref().
1748 	 */
1749 	MNT_ILOCK(mp);
1750 	if (vp == NULL && (flags & V_MNTREF) == 0)
1751 		MNT_REF(mp);
1752 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1753 		mp->mnt_secondary_writes++;
1754 		mp->mnt_secondary_accwrites++;
1755 		MNT_IUNLOCK(mp);
1756 		return (0);
1757 	}
1758 	if (flags & V_NOWAIT) {
1759 		MNT_REL(mp);
1760 		MNT_IUNLOCK(mp);
1761 		return (EWOULDBLOCK);
1762 	}
1763 	/*
1764 	 * Wait for the suspension to finish.
1765 	 */
1766 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1767 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1768 	    "suspfs", 0);
1769 	vfs_rel(mp);
1770 	if (error == 0)
1771 		goto retry;
1772 	return (error);
1773 }
1774 
1775 /*
1776  * Filesystem write operation has completed. If we are suspending and this
1777  * operation is the last one, notify the suspender that the suspension is
1778  * now in effect.
1779  */
1780 void
1781 vn_finished_write(struct mount *mp)
1782 {
1783 	if (mp == NULL || !vn_suspendable(mp))
1784 		return;
1785 	MNT_ILOCK(mp);
1786 	MNT_REL(mp);
1787 	mp->mnt_writeopcount--;
1788 	if (mp->mnt_writeopcount < 0)
1789 		panic("vn_finished_write: neg cnt");
1790 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1791 	    mp->mnt_writeopcount <= 0)
1792 		wakeup(&mp->mnt_writeopcount);
1793 	MNT_IUNLOCK(mp);
1794 }
1795 
1796 
1797 /*
1798  * Filesystem secondary write operation has completed. If we are
1799  * suspending and this operation is the last one, notify the suspender
1800  * that the suspension is now in effect.
1801  */
1802 void
1803 vn_finished_secondary_write(struct mount *mp)
1804 {
1805 	if (mp == NULL || !vn_suspendable(mp))
1806 		return;
1807 	MNT_ILOCK(mp);
1808 	MNT_REL(mp);
1809 	mp->mnt_secondary_writes--;
1810 	if (mp->mnt_secondary_writes < 0)
1811 		panic("vn_finished_secondary_write: neg cnt");
1812 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1813 	    mp->mnt_secondary_writes <= 0)
1814 		wakeup(&mp->mnt_secondary_writes);
1815 	MNT_IUNLOCK(mp);
1816 }
1817 
1818 
1819 
1820 /*
1821  * Request a filesystem to suspend write operations.
1822  */
1823 int
1824 vfs_write_suspend(struct mount *mp, int flags)
1825 {
1826 	int error;
1827 
1828 	MPASS(vn_suspendable(mp));
1829 
1830 	MNT_ILOCK(mp);
1831 	if (mp->mnt_susp_owner == curthread) {
1832 		MNT_IUNLOCK(mp);
1833 		return (EALREADY);
1834 	}
1835 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
1836 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1837 
1838 	/*
1839 	 * Unmount holds a write reference on the mount point.  If we
1840 	 * own busy reference and drain for writers, we deadlock with
1841 	 * the reference draining in the unmount path.  Callers of
1842 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1843 	 * vfs_busy() reference is owned and caller is not in the
1844 	 * unmount context.
1845 	 */
1846 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1847 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1848 		MNT_IUNLOCK(mp);
1849 		return (EBUSY);
1850 	}
1851 
1852 	mp->mnt_kern_flag |= MNTK_SUSPEND;
1853 	mp->mnt_susp_owner = curthread;
1854 	if (mp->mnt_writeopcount > 0)
1855 		(void) msleep(&mp->mnt_writeopcount,
1856 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1857 	else
1858 		MNT_IUNLOCK(mp);
1859 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1860 		vfs_write_resume(mp, 0);
1861 	return (error);
1862 }
1863 
1864 /*
1865  * Request a filesystem to resume write operations.
1866  */
1867 void
1868 vfs_write_resume(struct mount *mp, int flags)
1869 {
1870 
1871 	MPASS(vn_suspendable(mp));
1872 
1873 	MNT_ILOCK(mp);
1874 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1875 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1876 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1877 				       MNTK_SUSPENDED);
1878 		mp->mnt_susp_owner = NULL;
1879 		wakeup(&mp->mnt_writeopcount);
1880 		wakeup(&mp->mnt_flag);
1881 		curthread->td_pflags &= ~TDP_IGNSUSP;
1882 		if ((flags & VR_START_WRITE) != 0) {
1883 			MNT_REF(mp);
1884 			mp->mnt_writeopcount++;
1885 		}
1886 		MNT_IUNLOCK(mp);
1887 		if ((flags & VR_NO_SUSPCLR) == 0)
1888 			VFS_SUSP_CLEAN(mp);
1889 	} else if ((flags & VR_START_WRITE) != 0) {
1890 		MNT_REF(mp);
1891 		vn_start_write_locked(mp, 0);
1892 	} else {
1893 		MNT_IUNLOCK(mp);
1894 	}
1895 }
1896 
1897 /*
1898  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1899  * methods.
1900  */
1901 int
1902 vfs_write_suspend_umnt(struct mount *mp)
1903 {
1904 	int error;
1905 
1906 	MPASS(vn_suspendable(mp));
1907 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1908 	    ("vfs_write_suspend_umnt: recursed"));
1909 
1910 	/* dounmount() already called vn_start_write(). */
1911 	for (;;) {
1912 		vn_finished_write(mp);
1913 		error = vfs_write_suspend(mp, 0);
1914 		if (error != 0) {
1915 			vn_start_write(NULL, &mp, V_WAIT);
1916 			return (error);
1917 		}
1918 		MNT_ILOCK(mp);
1919 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1920 			break;
1921 		MNT_IUNLOCK(mp);
1922 		vn_start_write(NULL, &mp, V_WAIT);
1923 	}
1924 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1925 	wakeup(&mp->mnt_flag);
1926 	MNT_IUNLOCK(mp);
1927 	curthread->td_pflags |= TDP_IGNSUSP;
1928 	return (0);
1929 }
1930 
1931 /*
1932  * Implement kqueues for files by translating it to vnode operation.
1933  */
1934 static int
1935 vn_kqfilter(struct file *fp, struct knote *kn)
1936 {
1937 
1938 	return (VOP_KQFILTER(fp->f_vnode, kn));
1939 }
1940 
1941 /*
1942  * Simplified in-kernel wrapper calls for extended attribute access.
1943  * Both calls pass in a NULL credential, authorizing as "kernel" access.
1944  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1945  */
1946 int
1947 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1948     const char *attrname, int *buflen, char *buf, struct thread *td)
1949 {
1950 	struct uio	auio;
1951 	struct iovec	iov;
1952 	int	error;
1953 
1954 	iov.iov_len = *buflen;
1955 	iov.iov_base = buf;
1956 
1957 	auio.uio_iov = &iov;
1958 	auio.uio_iovcnt = 1;
1959 	auio.uio_rw = UIO_READ;
1960 	auio.uio_segflg = UIO_SYSSPACE;
1961 	auio.uio_td = td;
1962 	auio.uio_offset = 0;
1963 	auio.uio_resid = *buflen;
1964 
1965 	if ((ioflg & IO_NODELOCKED) == 0)
1966 		vn_lock(vp, LK_SHARED | LK_RETRY);
1967 
1968 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1969 
1970 	/* authorize attribute retrieval as kernel */
1971 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
1972 	    td);
1973 
1974 	if ((ioflg & IO_NODELOCKED) == 0)
1975 		VOP_UNLOCK(vp, 0);
1976 
1977 	if (error == 0) {
1978 		*buflen = *buflen - auio.uio_resid;
1979 	}
1980 
1981 	return (error);
1982 }
1983 
1984 /*
1985  * XXX failure mode if partially written?
1986  */
1987 int
1988 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
1989     const char *attrname, int buflen, char *buf, struct thread *td)
1990 {
1991 	struct uio	auio;
1992 	struct iovec	iov;
1993 	struct mount	*mp;
1994 	int	error;
1995 
1996 	iov.iov_len = buflen;
1997 	iov.iov_base = buf;
1998 
1999 	auio.uio_iov = &iov;
2000 	auio.uio_iovcnt = 1;
2001 	auio.uio_rw = UIO_WRITE;
2002 	auio.uio_segflg = UIO_SYSSPACE;
2003 	auio.uio_td = td;
2004 	auio.uio_offset = 0;
2005 	auio.uio_resid = buflen;
2006 
2007 	if ((ioflg & IO_NODELOCKED) == 0) {
2008 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2009 			return (error);
2010 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2011 	}
2012 
2013 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2014 
2015 	/* authorize attribute setting as kernel */
2016 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2017 
2018 	if ((ioflg & IO_NODELOCKED) == 0) {
2019 		vn_finished_write(mp);
2020 		VOP_UNLOCK(vp, 0);
2021 	}
2022 
2023 	return (error);
2024 }
2025 
2026 int
2027 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2028     const char *attrname, struct thread *td)
2029 {
2030 	struct mount	*mp;
2031 	int	error;
2032 
2033 	if ((ioflg & IO_NODELOCKED) == 0) {
2034 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2035 			return (error);
2036 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2037 	}
2038 
2039 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2040 
2041 	/* authorize attribute removal as kernel */
2042 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2043 	if (error == EOPNOTSUPP)
2044 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2045 		    NULL, td);
2046 
2047 	if ((ioflg & IO_NODELOCKED) == 0) {
2048 		vn_finished_write(mp);
2049 		VOP_UNLOCK(vp, 0);
2050 	}
2051 
2052 	return (error);
2053 }
2054 
2055 static int
2056 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2057     struct vnode **rvp)
2058 {
2059 
2060 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2061 }
2062 
2063 int
2064 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2065 {
2066 
2067 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2068 	    lkflags, rvp));
2069 }
2070 
2071 int
2072 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2073     int lkflags, struct vnode **rvp)
2074 {
2075 	struct mount *mp;
2076 	int ltype, error;
2077 
2078 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2079 	mp = vp->v_mount;
2080 	ltype = VOP_ISLOCKED(vp);
2081 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2082 	    ("vn_vget_ino: vp not locked"));
2083 	error = vfs_busy(mp, MBF_NOWAIT);
2084 	if (error != 0) {
2085 		vfs_ref(mp);
2086 		VOP_UNLOCK(vp, 0);
2087 		error = vfs_busy(mp, 0);
2088 		vn_lock(vp, ltype | LK_RETRY);
2089 		vfs_rel(mp);
2090 		if (error != 0)
2091 			return (ENOENT);
2092 		if (vp->v_iflag & VI_DOOMED) {
2093 			vfs_unbusy(mp);
2094 			return (ENOENT);
2095 		}
2096 	}
2097 	VOP_UNLOCK(vp, 0);
2098 	error = alloc(mp, alloc_arg, lkflags, rvp);
2099 	vfs_unbusy(mp);
2100 	if (error != 0 || *rvp != vp)
2101 		vn_lock(vp, ltype | LK_RETRY);
2102 	if (vp->v_iflag & VI_DOOMED) {
2103 		if (error == 0) {
2104 			if (*rvp == vp)
2105 				vunref(vp);
2106 			else
2107 				vput(*rvp);
2108 		}
2109 		error = ENOENT;
2110 	}
2111 	return (error);
2112 }
2113 
2114 int
2115 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2116     struct thread *td)
2117 {
2118 
2119 	if (vp->v_type != VREG || td == NULL)
2120 		return (0);
2121 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2122 	    lim_cur(td, RLIMIT_FSIZE)) {
2123 		PROC_LOCK(td->td_proc);
2124 		kern_psignal(td->td_proc, SIGXFSZ);
2125 		PROC_UNLOCK(td->td_proc);
2126 		return (EFBIG);
2127 	}
2128 	return (0);
2129 }
2130 
2131 int
2132 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2133     struct thread *td)
2134 {
2135 	struct vnode *vp;
2136 
2137 	vp = fp->f_vnode;
2138 #ifdef AUDIT
2139 	vn_lock(vp, LK_SHARED | LK_RETRY);
2140 	AUDIT_ARG_VNODE1(vp);
2141 	VOP_UNLOCK(vp, 0);
2142 #endif
2143 	return (setfmode(td, active_cred, vp, mode));
2144 }
2145 
2146 int
2147 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2148     struct thread *td)
2149 {
2150 	struct vnode *vp;
2151 
2152 	vp = fp->f_vnode;
2153 #ifdef AUDIT
2154 	vn_lock(vp, LK_SHARED | LK_RETRY);
2155 	AUDIT_ARG_VNODE1(vp);
2156 	VOP_UNLOCK(vp, 0);
2157 #endif
2158 	return (setfown(td, active_cred, vp, uid, gid));
2159 }
2160 
2161 void
2162 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2163 {
2164 	vm_object_t object;
2165 
2166 	if ((object = vp->v_object) == NULL)
2167 		return;
2168 	VM_OBJECT_WLOCK(object);
2169 	vm_object_page_remove(object, start, end, 0);
2170 	VM_OBJECT_WUNLOCK(object);
2171 }
2172 
2173 int
2174 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2175 {
2176 	struct vattr va;
2177 	daddr_t bn, bnp;
2178 	uint64_t bsize;
2179 	off_t noff;
2180 	int error;
2181 
2182 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2183 	    ("Wrong command %lu", cmd));
2184 
2185 	if (vn_lock(vp, LK_SHARED) != 0)
2186 		return (EBADF);
2187 	if (vp->v_type != VREG) {
2188 		error = ENOTTY;
2189 		goto unlock;
2190 	}
2191 	error = VOP_GETATTR(vp, &va, cred);
2192 	if (error != 0)
2193 		goto unlock;
2194 	noff = *off;
2195 	if (noff >= va.va_size) {
2196 		error = ENXIO;
2197 		goto unlock;
2198 	}
2199 	bsize = vp->v_mount->mnt_stat.f_iosize;
2200 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2201 	    noff % bsize) {
2202 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2203 		if (error == EOPNOTSUPP) {
2204 			error = ENOTTY;
2205 			goto unlock;
2206 		}
2207 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2208 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2209 			noff = bn * bsize;
2210 			if (noff < *off)
2211 				noff = *off;
2212 			goto unlock;
2213 		}
2214 	}
2215 	if (noff > va.va_size)
2216 		noff = va.va_size;
2217 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2218 	if (cmd == FIOSEEKDATA)
2219 		error = ENXIO;
2220 unlock:
2221 	VOP_UNLOCK(vp, 0);
2222 	if (error == 0)
2223 		*off = noff;
2224 	return (error);
2225 }
2226 
2227 int
2228 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2229 {
2230 	struct ucred *cred;
2231 	struct vnode *vp;
2232 	struct vattr vattr;
2233 	off_t foffset, size;
2234 	int error, noneg;
2235 
2236 	cred = td->td_ucred;
2237 	vp = fp->f_vnode;
2238 	foffset = foffset_lock(fp, 0);
2239 	noneg = (vp->v_type != VCHR);
2240 	error = 0;
2241 	switch (whence) {
2242 	case L_INCR:
2243 		if (noneg &&
2244 		    (foffset < 0 ||
2245 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2246 			error = EOVERFLOW;
2247 			break;
2248 		}
2249 		offset += foffset;
2250 		break;
2251 	case L_XTND:
2252 		vn_lock(vp, LK_SHARED | LK_RETRY);
2253 		error = VOP_GETATTR(vp, &vattr, cred);
2254 		VOP_UNLOCK(vp, 0);
2255 		if (error)
2256 			break;
2257 
2258 		/*
2259 		 * If the file references a disk device, then fetch
2260 		 * the media size and use that to determine the ending
2261 		 * offset.
2262 		 */
2263 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2264 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2265 			vattr.va_size = size;
2266 		if (noneg &&
2267 		    (vattr.va_size > OFF_MAX ||
2268 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2269 			error = EOVERFLOW;
2270 			break;
2271 		}
2272 		offset += vattr.va_size;
2273 		break;
2274 	case L_SET:
2275 		break;
2276 	case SEEK_DATA:
2277 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2278 		if (error == ENOTTY)
2279 			error = EINVAL;
2280 		break;
2281 	case SEEK_HOLE:
2282 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2283 		if (error == ENOTTY)
2284 			error = EINVAL;
2285 		break;
2286 	default:
2287 		error = EINVAL;
2288 	}
2289 	if (error == 0 && noneg && offset < 0)
2290 		error = EINVAL;
2291 	if (error != 0)
2292 		goto drop;
2293 	VFS_KNOTE_UNLOCKED(vp, 0);
2294 	td->td_uretoff.tdu_off = offset;
2295 drop:
2296 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2297 	return (error);
2298 }
2299 
2300 int
2301 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2302     struct thread *td)
2303 {
2304 	int error;
2305 
2306 	/*
2307 	 * Grant permission if the caller is the owner of the file, or
2308 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2309 	 * on the file.  If the time pointer is null, then write
2310 	 * permission on the file is also sufficient.
2311 	 *
2312 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2313 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2314 	 * will be allowed to set the times [..] to the current
2315 	 * server time.
2316 	 */
2317 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2318 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2319 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2320 	return (error);
2321 }
2322 
2323 int
2324 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2325 {
2326 	struct vnode *vp;
2327 	int error;
2328 
2329 	if (fp->f_type == DTYPE_FIFO)
2330 		kif->kf_type = KF_TYPE_FIFO;
2331 	else
2332 		kif->kf_type = KF_TYPE_VNODE;
2333 	vp = fp->f_vnode;
2334 	vref(vp);
2335 	FILEDESC_SUNLOCK(fdp);
2336 	error = vn_fill_kinfo_vnode(vp, kif);
2337 	vrele(vp);
2338 	FILEDESC_SLOCK(fdp);
2339 	return (error);
2340 }
2341 
2342 static inline void
2343 vn_fill_junk(struct kinfo_file *kif)
2344 {
2345 	size_t len, olen;
2346 
2347 	/*
2348 	 * Simulate vn_fullpath returning changing values for a given
2349 	 * vp during e.g. coredump.
2350 	 */
2351 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2352 	olen = strlen(kif->kf_path);
2353 	if (len < olen)
2354 		strcpy(&kif->kf_path[len - 1], "$");
2355 	else
2356 		for (; olen < len; olen++)
2357 			strcpy(&kif->kf_path[olen], "A");
2358 }
2359 
2360 int
2361 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2362 {
2363 	struct vattr va;
2364 	char *fullpath, *freepath;
2365 	int error;
2366 
2367 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2368 	freepath = NULL;
2369 	fullpath = "-";
2370 	error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2371 	if (error == 0) {
2372 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2373 	}
2374 	if (freepath != NULL)
2375 		free(freepath, M_TEMP);
2376 
2377 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2378 		vn_fill_junk(kif);
2379 	);
2380 
2381 	/*
2382 	 * Retrieve vnode attributes.
2383 	 */
2384 	va.va_fsid = VNOVAL;
2385 	va.va_rdev = NODEV;
2386 	vn_lock(vp, LK_SHARED | LK_RETRY);
2387 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2388 	VOP_UNLOCK(vp, 0);
2389 	if (error != 0)
2390 		return (error);
2391 	if (va.va_fsid != VNOVAL)
2392 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2393 	else
2394 		kif->kf_un.kf_file.kf_file_fsid =
2395 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2396 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2397 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2398 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2399 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2400 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2401 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2402 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2403 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2404 	return (0);
2405 }
2406 
2407 int
2408 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2409     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2410     struct thread *td)
2411 {
2412 #ifdef HWPMC_HOOKS
2413 	struct pmckern_map_in pkm;
2414 #endif
2415 	struct mount *mp;
2416 	struct vnode *vp;
2417 	vm_object_t object;
2418 	vm_prot_t maxprot;
2419 	boolean_t writecounted;
2420 	int error;
2421 
2422 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2423     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2424 	/*
2425 	 * POSIX shared-memory objects are defined to have
2426 	 * kernel persistence, and are not defined to support
2427 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2428 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2429 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2430 	 * flag to request this behavior.
2431 	 */
2432 	if ((fp->f_flag & FPOSIXSHM) != 0)
2433 		flags |= MAP_NOSYNC;
2434 #endif
2435 	vp = fp->f_vnode;
2436 
2437 	/*
2438 	 * Ensure that file and memory protections are
2439 	 * compatible.  Note that we only worry about
2440 	 * writability if mapping is shared; in this case,
2441 	 * current and max prot are dictated by the open file.
2442 	 * XXX use the vnode instead?  Problem is: what
2443 	 * credentials do we use for determination? What if
2444 	 * proc does a setuid?
2445 	 */
2446 	mp = vp->v_mount;
2447 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2448 		maxprot = VM_PROT_NONE;
2449 		if ((prot & VM_PROT_EXECUTE) != 0)
2450 			return (EACCES);
2451 	} else
2452 		maxprot = VM_PROT_EXECUTE;
2453 	if ((fp->f_flag & FREAD) != 0)
2454 		maxprot |= VM_PROT_READ;
2455 	else if ((prot & VM_PROT_READ) != 0)
2456 		return (EACCES);
2457 
2458 	/*
2459 	 * If we are sharing potential changes via MAP_SHARED and we
2460 	 * are trying to get write permission although we opened it
2461 	 * without asking for it, bail out.
2462 	 */
2463 	if ((flags & MAP_SHARED) != 0) {
2464 		if ((fp->f_flag & FWRITE) != 0)
2465 			maxprot |= VM_PROT_WRITE;
2466 		else if ((prot & VM_PROT_WRITE) != 0)
2467 			return (EACCES);
2468 	} else {
2469 		maxprot |= VM_PROT_WRITE;
2470 		cap_maxprot |= VM_PROT_WRITE;
2471 	}
2472 	maxprot &= cap_maxprot;
2473 
2474 	/*
2475 	 * For regular files and shared memory, POSIX requires that
2476 	 * the value of foff be a legitimate offset within the data
2477 	 * object.  In particular, negative offsets are invalid.
2478 	 * Blocking negative offsets and overflows here avoids
2479 	 * possible wraparound or user-level access into reserved
2480 	 * ranges of the data object later.  In contrast, POSIX does
2481 	 * not dictate how offsets are used by device drivers, so in
2482 	 * the case of a device mapping a negative offset is passed
2483 	 * on.
2484 	 */
2485 	if (
2486 #ifdef _LP64
2487 	    size > OFF_MAX ||
2488 #endif
2489 	    foff < 0 || foff > OFF_MAX - size)
2490 		return (EINVAL);
2491 
2492 	writecounted = FALSE;
2493 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2494 	    &foff, &object, &writecounted);
2495 	if (error != 0)
2496 		return (error);
2497 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2498 	    foff, writecounted, td);
2499 	if (error != 0) {
2500 		/*
2501 		 * If this mapping was accounted for in the vnode's
2502 		 * writecount, then undo that now.
2503 		 */
2504 		if (writecounted)
2505 			vnode_pager_release_writecount(object, 0, size);
2506 		vm_object_deallocate(object);
2507 	}
2508 #ifdef HWPMC_HOOKS
2509 	/* Inform hwpmc(4) if an executable is being mapped. */
2510 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2511 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2512 			pkm.pm_file = vp;
2513 			pkm.pm_address = (uintptr_t) *addr;
2514 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2515 		}
2516 	}
2517 #endif
2518 	return (error);
2519 }
2520 
2521 void
2522 vn_fsid(struct vnode *vp, struct vattr *va)
2523 {
2524 	fsid_t *f;
2525 
2526 	f = &vp->v_mount->mnt_stat.f_fsid;
2527 	va->va_fsid = (uint32_t)f->val[1];
2528 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2529 	va->va_fsid += (uint32_t)f->val[0];
2530 }
2531 
2532 int
2533 vn_fsync_buf(struct vnode *vp, int waitfor)
2534 {
2535 	struct buf *bp, *nbp;
2536 	struct bufobj *bo;
2537 	struct mount *mp;
2538 	int error, maxretry;
2539 
2540 	error = 0;
2541 	maxretry = 10000;     /* large, arbitrarily chosen */
2542 	mp = NULL;
2543 	if (vp->v_type == VCHR) {
2544 		VI_LOCK(vp);
2545 		mp = vp->v_rdev->si_mountpt;
2546 		VI_UNLOCK(vp);
2547 	}
2548 	bo = &vp->v_bufobj;
2549 	BO_LOCK(bo);
2550 loop1:
2551 	/*
2552 	 * MARK/SCAN initialization to avoid infinite loops.
2553 	 */
2554         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2555 		bp->b_vflags &= ~BV_SCANNED;
2556 		bp->b_error = 0;
2557 	}
2558 
2559 	/*
2560 	 * Flush all dirty buffers associated with a vnode.
2561 	 */
2562 loop2:
2563 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2564 		if ((bp->b_vflags & BV_SCANNED) != 0)
2565 			continue;
2566 		bp->b_vflags |= BV_SCANNED;
2567 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2568 			if (waitfor != MNT_WAIT)
2569 				continue;
2570 			if (BUF_LOCK(bp,
2571 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2572 			    BO_LOCKPTR(bo)) != 0) {
2573 				BO_LOCK(bo);
2574 				goto loop1;
2575 			}
2576 			BO_LOCK(bo);
2577 		}
2578 		BO_UNLOCK(bo);
2579 		KASSERT(bp->b_bufobj == bo,
2580 		    ("bp %p wrong b_bufobj %p should be %p",
2581 		    bp, bp->b_bufobj, bo));
2582 		if ((bp->b_flags & B_DELWRI) == 0)
2583 			panic("fsync: not dirty");
2584 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2585 			vfs_bio_awrite(bp);
2586 		} else {
2587 			bremfree(bp);
2588 			bawrite(bp);
2589 		}
2590 		if (maxretry < 1000)
2591 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2592 		BO_LOCK(bo);
2593 		goto loop2;
2594 	}
2595 
2596 	/*
2597 	 * If synchronous the caller expects us to completely resolve all
2598 	 * dirty buffers in the system.  Wait for in-progress I/O to
2599 	 * complete (which could include background bitmap writes), then
2600 	 * retry if dirty blocks still exist.
2601 	 */
2602 	if (waitfor == MNT_WAIT) {
2603 		bufobj_wwait(bo, 0, 0);
2604 		if (bo->bo_dirty.bv_cnt > 0) {
2605 			/*
2606 			 * If we are unable to write any of these buffers
2607 			 * then we fail now rather than trying endlessly
2608 			 * to write them out.
2609 			 */
2610 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2611 				if ((error = bp->b_error) != 0)
2612 					break;
2613 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2614 			    (error == 0 && --maxretry >= 0))
2615 				goto loop1;
2616 			if (error == 0)
2617 				error = EAGAIN;
2618 		}
2619 	}
2620 	BO_UNLOCK(bo);
2621 	if (error != 0)
2622 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2623 
2624 	return (error);
2625 }
2626 
2627 /*
2628  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2629  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2630  * to do the actual copy.
2631  * vn_generic_copy_file_range() is factored out, so it can be called
2632  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2633  * different file systems.
2634  */
2635 int
2636 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2637     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2638     struct ucred *outcred, struct thread *fsize_td)
2639 {
2640 	struct vattr va;
2641 	int error;
2642 	size_t len;
2643 	uint64_t uvalin, uvalout;
2644 
2645 	len = *lenp;
2646 	*lenp = 0;		/* For error returns. */
2647 	error = 0;
2648 
2649 	/* Do some sanity checks on the arguments. */
2650 	uvalin = *inoffp;
2651 	uvalin += len;
2652 	uvalout = *outoffp;
2653 	uvalout += len;
2654 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2655 		error = EISDIR;
2656 	else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2657 	    (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2658 	    uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2659 	    outvp->v_type != VREG)
2660 		error = EINVAL;
2661 	else if (invp == outvp)
2662 		error = EBADF;
2663 	if (error != 0)
2664 		goto out;
2665 
2666 	error = vn_lock(invp, LK_SHARED);
2667 	if (error != 0)
2668 		goto out;
2669 	/* Check that the offset + len does not go past EOF of invp. */
2670 	error = VOP_GETATTR(invp, &va, incred);
2671 	if (error == 0 && va.va_size < *inoffp + len)
2672 		error = EINVAL;
2673 	VOP_UNLOCK(invp, 0);
2674 	if (error != 0)
2675 		goto out;
2676 
2677 	/*
2678 	 * If the two vnode are for the same file system, call
2679 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2680 	 * which can handle copies across multiple file systems.
2681 	 */
2682 	*lenp = len;
2683 	if (invp->v_mount == outvp->v_mount)
2684 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2685 		    lenp, flags, incred, outcred, fsize_td);
2686 	else
2687 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2688 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2689 out:
2690 	return (error);
2691 }
2692 
2693 /*
2694  * Test len bytes of data starting at dat for all bytes == 0.
2695  * Return true if all bytes are zero, false otherwise.
2696  * Expects dat to be well aligned.
2697  */
2698 static bool
2699 mem_iszero(void *dat, int len)
2700 {
2701 	int i;
2702 	const u_int *p;
2703 	const char *cp;
2704 
2705 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2706 		if (len >= sizeof(*p)) {
2707 			if (*p != 0)
2708 				return (false);
2709 		} else {
2710 			cp = (const char *)p;
2711 			for (i = 0; i < len; i++, cp++)
2712 				if (*cp != '\0')
2713 					return (false);
2714 		}
2715 	}
2716 	return (true);
2717 }
2718 
2719 /*
2720  * Look for a hole in the output file and, if found, adjust *outoffp
2721  * and *xferp to skip past the hole.
2722  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2723  * to be written as 0's upon return.
2724  */
2725 static off_t
2726 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2727     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2728 {
2729 	int error;
2730 	off_t delta;
2731 
2732 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2733 		*dataoffp = *outoffp;
2734 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2735 		    curthread);
2736 		if (error == 0) {
2737 			*holeoffp = *dataoffp;
2738 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2739 			    curthread);
2740 		}
2741 		if (error != 0 || *holeoffp == *dataoffp) {
2742 			/*
2743 			 * Since outvp is unlocked, it may be possible for
2744 			 * another thread to do a truncate(), lseek(), write()
2745 			 * creating a hole at startoff between the above
2746 			 * VOP_IOCTL() calls, if the other thread does not do
2747 			 * rangelocking.
2748 			 * If that happens, *holeoffp == *dataoffp and finding
2749 			 * the hole has failed, so disable vn_skip_hole().
2750 			 */
2751 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2752 			return (xfer2);
2753 		}
2754 		KASSERT(*dataoffp >= *outoffp,
2755 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2756 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2757 		KASSERT(*holeoffp > *dataoffp,
2758 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2759 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2760 	}
2761 
2762 	/*
2763 	 * If there is a hole before the data starts, advance *outoffp and
2764 	 * *xferp past the hole.
2765 	 */
2766 	if (*dataoffp > *outoffp) {
2767 		delta = *dataoffp - *outoffp;
2768 		if (delta >= *xferp) {
2769 			/* Entire *xferp is a hole. */
2770 			*outoffp += *xferp;
2771 			*xferp = 0;
2772 			return (0);
2773 		}
2774 		*xferp -= delta;
2775 		*outoffp += delta;
2776 		xfer2 = MIN(xfer2, *xferp);
2777 	}
2778 
2779 	/*
2780 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2781 	 * that the write ends at the start of the hole.
2782 	 * *holeoffp should always be greater than *outoffp, but for the
2783 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2784 	 * value.
2785 	 */
2786 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2787 		xfer2 = *holeoffp - *outoffp;
2788 	return (xfer2);
2789 }
2790 
2791 /*
2792  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2793  * dat is a maximum of blksize in length and can be written repeatedly in
2794  * the chunk.
2795  * If growfile == true, just grow the file via vn_truncate_locked() instead
2796  * of doing actual writes.
2797  * If checkhole == true, a hole is being punched, so skip over any hole
2798  * already in the output file.
2799  */
2800 static int
2801 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2802     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2803 {
2804 	struct mount *mp;
2805 	off_t dataoff, holeoff, xfer2;
2806 	int error, lckf;
2807 
2808 	/*
2809 	 * Loop around doing writes of blksize until write has been completed.
2810 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2811 	 * done for each iteration, since the xfer argument can be very
2812 	 * large if there is a large hole to punch in the output file.
2813 	 */
2814 	error = 0;
2815 	holeoff = 0;
2816 	do {
2817 		xfer2 = MIN(xfer, blksize);
2818 		if (checkhole) {
2819 			/*
2820 			 * Punching a hole.  Skip writing if there is
2821 			 * already a hole in the output file.
2822 			 */
2823 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2824 			    &dataoff, &holeoff, cred);
2825 			if (xfer == 0)
2826 				break;
2827 			if (holeoff < 0)
2828 				checkhole = false;
2829 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2830 			    (intmax_t)xfer2));
2831 		}
2832 		bwillwrite();
2833 		mp = NULL;
2834 		error = vn_start_write(outvp, &mp, V_WAIT);
2835 		if (error == 0) {
2836 			if (MNT_SHARED_WRITES(mp))
2837 				lckf = LK_SHARED;
2838 			else
2839 				lckf = LK_EXCLUSIVE;
2840 			error = vn_lock(outvp, lckf);
2841 		}
2842 		if (error == 0) {
2843 			if (growfile)
2844 				error = vn_truncate_locked(outvp, outoff + xfer,
2845 				    false, cred);
2846 			else {
2847 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2848 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
2849 				    curthread->td_ucred, cred, NULL, curthread);
2850 				outoff += xfer2;
2851 				xfer -= xfer2;
2852 			}
2853 			VOP_UNLOCK(outvp, 0);
2854 		}
2855 		if (mp != NULL)
2856 			vn_finished_write(mp);
2857 	} while (!growfile && xfer > 0 && error == 0);
2858 	return (error);
2859 }
2860 
2861 /*
2862  * Copy a byte range of one file to another.  This function can handle the
2863  * case where invp and outvp are on different file systems.
2864  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2865  * is no better file system specific way to do it.
2866  */
2867 int
2868 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2869     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
2870     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
2871 {
2872 	struct vattr va;
2873 	struct mount *mp;
2874 	struct uio io;
2875 	off_t startoff, endoff, xfer, xfer2;
2876 	u_long blksize;
2877 	int error;
2878 	bool cantseek, readzeros;
2879 	ssize_t aresid;
2880 	size_t copylen, len, savlen;
2881 	char *dat;
2882 	long holein, holeout;
2883 
2884 	holein = holeout = 0;
2885 	savlen = len = *lenp;
2886 	error = 0;
2887 	dat = NULL;
2888 
2889 	error = vn_lock(invp, LK_SHARED);
2890 	if (error != 0)
2891 		goto out;
2892 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
2893 		holein = 0;
2894 	VOP_UNLOCK(invp, 0);
2895 
2896 	mp = NULL;
2897 	error = vn_start_write(outvp, &mp, V_WAIT);
2898 	if (error == 0)
2899 		error = vn_lock(outvp, LK_EXCLUSIVE);
2900 	if (error == 0) {
2901 		/*
2902 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
2903 		 * now that outvp is locked.
2904 		 */
2905 		if (fsize_td != NULL) {
2906 			io.uio_offset = *outoffp;
2907 			io.uio_resid = len;
2908 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
2909 			if (error != 0)
2910 				error = EFBIG;
2911 		}
2912 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
2913 			holeout = 0;
2914 		/*
2915 		 * Holes that are past EOF do not need to be written as a block
2916 		 * of zero bytes.  So, truncate the output file as far as
2917 		 * possible and then use va.va_size to decide if writing 0
2918 		 * bytes is necessary in the loop below.
2919 		 */
2920 		if (error == 0)
2921 			error = VOP_GETATTR(outvp, &va, outcred);
2922 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
2923 		    *outoffp + len) {
2924 #ifdef MAC
2925 			error = mac_vnode_check_write(curthread->td_ucred,
2926 			    outcred, outvp);
2927 			if (error == 0)
2928 #endif
2929 				error = vn_truncate_locked(outvp, *outoffp,
2930 				    false, outcred);
2931 			if (error == 0)
2932 				va.va_size = *outoffp;
2933 		}
2934 		VOP_UNLOCK(outvp, 0);
2935 	}
2936 	if (mp != NULL)
2937 		vn_finished_write(mp);
2938 	if (error != 0)
2939 		goto out;
2940 
2941 	/*
2942 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
2943 	 * If hole sizes aren't available, set the blksize to the larger
2944 	 * f_iosize of invp and outvp.
2945 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
2946 	 * This value is clipped at 4Kbytes and 1Mbyte.
2947 	 */
2948 	blksize = MAX(holein, holeout);
2949 	if (blksize == 0)
2950 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
2951 		    outvp->v_mount->mnt_stat.f_iosize);
2952 	if (blksize < 4096)
2953 		blksize = 4096;
2954 	else if (blksize > 1024 * 1024)
2955 		blksize = 1024 * 1024;
2956 	dat = malloc(blksize, M_TEMP, M_WAITOK);
2957 
2958 	/*
2959 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
2960 	 * to find holes.  Otherwise, just scan the read block for all 0s
2961 	 * in the inner loop where the data copying is done.
2962 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
2963 	 * support holes on the server, but do not support FIOSEEKHOLE.
2964 	 */
2965 	while (len > 0 && error == 0) {
2966 		endoff = 0;			/* To shut up compilers. */
2967 		cantseek = true;
2968 		startoff = *inoffp;
2969 		copylen = len;
2970 
2971 		/*
2972 		 * Find the next data area.  If there is just a hole to EOF,
2973 		 * FIOSEEKDATA should fail and then we drop down into the
2974 		 * inner loop and create the hole on the outvp file.
2975 		 * (I do not know if any file system will report a hole to
2976 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
2977 		 *  will fail for those file systems.)
2978 		 *
2979 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
2980 		 * the code just falls through to the inner copy loop.
2981 		 */
2982 		error = EINVAL;
2983 		if (holein > 0)
2984 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
2985 			    incred, curthread);
2986 		if (error == 0) {
2987 			endoff = startoff;
2988 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
2989 			    incred, curthread);
2990 			/*
2991 			 * Since invp is unlocked, it may be possible for
2992 			 * another thread to do a truncate(), lseek(), write()
2993 			 * creating a hole at startoff between the above
2994 			 * VOP_IOCTL() calls, if the other thread does not do
2995 			 * rangelocking.
2996 			 * If that happens, startoff == endoff and finding
2997 			 * the hole has failed, so set an error.
2998 			 */
2999 			if (error == 0 && startoff == endoff)
3000 				error = EINVAL; /* Any error. Reset to 0. */
3001 		}
3002 		if (error == 0) {
3003 			if (startoff > *inoffp) {
3004 				/* Found hole before data block. */
3005 				xfer = MIN(startoff - *inoffp, len);
3006 				if (*outoffp < va.va_size) {
3007 					/* Must write 0s to punch hole. */
3008 					xfer2 = MIN(va.va_size - *outoffp,
3009 					    xfer);
3010 					memset(dat, 0, MIN(xfer2, blksize));
3011 					error = vn_write_outvp(outvp, dat,
3012 					    *outoffp, xfer2, blksize, false,
3013 					    holeout > 0, outcred);
3014 				}
3015 
3016 				if (error == 0 && *outoffp + xfer >
3017 				    va.va_size && xfer == len)
3018 					/* Grow last block. */
3019 					error = vn_write_outvp(outvp, dat,
3020 					    *outoffp, xfer, blksize, true,
3021 					    false, outcred);
3022 				if (error == 0) {
3023 					*inoffp += xfer;
3024 					*outoffp += xfer;
3025 					len -= xfer;
3026 				}
3027 			}
3028 			copylen = MIN(len, endoff - startoff);
3029 			cantseek = false;
3030 		} else {
3031 			cantseek = true;
3032 			startoff = *inoffp;
3033 			copylen = len;
3034 			error = 0;
3035 		}
3036 
3037 		xfer = blksize;
3038 		if (cantseek) {
3039 			/*
3040 			 * Set first xfer to end at a block boundary, so that
3041 			 * holes are more likely detected in the loop below via
3042 			 * the for all bytes 0 method.
3043 			 */
3044 			xfer -= (*inoffp % blksize);
3045 		}
3046 		/* Loop copying the data block. */
3047 		while (copylen > 0 && error == 0) {
3048 			if (copylen < xfer)
3049 				xfer = copylen;
3050 			error = vn_lock(invp, LK_SHARED);
3051 			if (error != 0)
3052 				goto out;
3053 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3054 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3055 			    curthread->td_ucred, incred, &aresid,
3056 			    curthread);
3057 			VOP_UNLOCK(invp, 0);
3058 			/*
3059 			 * Linux considers a range that exceeds EOF to
3060 			 * be an error, so we will too.
3061 			 */
3062 			if (error == 0 && aresid > 0)
3063 				error = EINVAL;
3064 			if (error == 0) {
3065 				/*
3066 				 * Skip the write for holes past the initial EOF
3067 				 * of the output file, unless this is the last
3068 				 * write of the output file at EOF.
3069 				 */
3070 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3071 				    false;
3072 				if (!cantseek || *outoffp < va.va_size ||
3073 				    xfer == len || !readzeros)
3074 					error = vn_write_outvp(outvp, dat,
3075 					    *outoffp, xfer, blksize,
3076 					    readzeros && xfer == len &&
3077 					    *outoffp >= va.va_size, false,
3078 					    outcred);
3079 				if (error == 0) {
3080 					*inoffp += xfer;
3081 					startoff += xfer;
3082 					*outoffp += xfer;
3083 					copylen -= xfer;
3084 					len -= xfer;
3085 				}
3086 			}
3087 			xfer = blksize;
3088 		}
3089 	}
3090 out:
3091 	*lenp = savlen - len;
3092 	free(dat, M_TEMP);
3093 	return (error);
3094 }
3095