xref: /freebsd/sys/kern/vfs_vnops.c (revision 3a3deb00a5e449c9478156b162dfa10ec82a2a3f)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/sx.h>
74 #include <sys/sysctl.h>
75 #include <sys/ttycom.h>
76 #include <sys/conf.h>
77 #include <sys/syslog.h>
78 #include <sys/unistd.h>
79 #include <sys/user.h>
80 
81 #include <security/audit/audit.h>
82 #include <security/mac/mac_framework.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_extern.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pager.h>
91 
92 #ifdef HWPMC_HOOKS
93 #include <sys/pmckern.h>
94 #endif
95 
96 static fo_rdwr_t	vn_read;
97 static fo_rdwr_t	vn_write;
98 static fo_rdwr_t	vn_io_fault;
99 static fo_truncate_t	vn_truncate;
100 static fo_ioctl_t	vn_ioctl;
101 static fo_poll_t	vn_poll;
102 static fo_kqfilter_t	vn_kqfilter;
103 static fo_stat_t	vn_statfile;
104 static fo_close_t	vn_closefile;
105 static fo_mmap_t	vn_mmap;
106 
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 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
223 			bwillwrite();
224 		if ((error = namei(ndp)) != 0)
225 			return (error);
226 		if (ndp->ni_vp == NULL) {
227 			VATTR_NULL(vap);
228 			vap->va_type = VREG;
229 			vap->va_mode = cmode;
230 			if (fmode & O_EXCL)
231 				vap->va_vaflags |= VA_EXCLUSIVE;
232 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
233 				NDFREE(ndp, NDF_ONLY_PNBUF);
234 				vput(ndp->ni_dvp);
235 				if ((error = vn_start_write(NULL, &mp,
236 				    V_XSLEEP | PCATCH)) != 0)
237 					return (error);
238 				goto restart;
239 			}
240 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
241 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
242 #ifdef MAC
243 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
244 			    &ndp->ni_cnd, vap);
245 			if (error == 0)
246 #endif
247 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
248 						   &ndp->ni_cnd, vap);
249 			vput(ndp->ni_dvp);
250 			vn_finished_write(mp);
251 			if (error) {
252 				NDFREE(ndp, NDF_ONLY_PNBUF);
253 				return (error);
254 			}
255 			fmode &= ~O_TRUNC;
256 			vp = ndp->ni_vp;
257 		} else {
258 			if (ndp->ni_dvp == ndp->ni_vp)
259 				vrele(ndp->ni_dvp);
260 			else
261 				vput(ndp->ni_dvp);
262 			ndp->ni_dvp = NULL;
263 			vp = ndp->ni_vp;
264 			if (fmode & O_EXCL) {
265 				error = EEXIST;
266 				goto bad;
267 			}
268 			if (vp->v_type == VDIR) {
269 				error = EISDIR;
270 				goto bad;
271 			}
272 			fmode &= ~O_CREAT;
273 		}
274 	} else {
275 		ndp->ni_cnd.cn_nameiop = LOOKUP;
276 		ndp->ni_cnd.cn_flags = ISOPEN |
277 		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
278 		if (!(fmode & FWRITE))
279 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
280 		if ((fmode & O_BENEATH) != 0)
281 			ndp->ni_cnd.cn_flags |= BENEATH;
282 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
283 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
284 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
285 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
286 		if ((error = namei(ndp)) != 0)
287 			return (error);
288 		vp = ndp->ni_vp;
289 	}
290 	error = vn_open_vnode(vp, fmode, cred, td, fp);
291 	if (error)
292 		goto bad;
293 	*flagp = fmode;
294 	return (0);
295 bad:
296 	NDFREE(ndp, NDF_ONLY_PNBUF);
297 	vput(vp);
298 	*flagp = fmode;
299 	ndp->ni_vp = NULL;
300 	return (error);
301 }
302 
303 static int
304 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
305 {
306 	struct flock lf;
307 	int error, lock_flags, type;
308 
309 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
310 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
311 		return (0);
312 	KASSERT(fp != NULL, ("open with flock requires fp"));
313 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
314 		return (EOPNOTSUPP);
315 
316 	lock_flags = VOP_ISLOCKED(vp);
317 	VOP_UNLOCK(vp, 0);
318 
319 	lf.l_whence = SEEK_SET;
320 	lf.l_start = 0;
321 	lf.l_len = 0;
322 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
323 	type = F_FLOCK;
324 	if ((fmode & FNONBLOCK) == 0)
325 		type |= F_WAIT;
326 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
327 	if (error == 0)
328 		fp->f_flag |= FHASLOCK;
329 
330 	vn_lock(vp, lock_flags | LK_RETRY);
331 	if (error == 0 && VN_IS_DOOMED(vp))
332 		error = ENOENT;
333 	return (error);
334 }
335 
336 /*
337  * Common code for vnode open operations once a vnode is located.
338  * Check permissions, and call the VOP_OPEN routine.
339  */
340 int
341 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
342     struct thread *td, struct file *fp)
343 {
344 	accmode_t accmode;
345 	int error;
346 
347 	if (vp->v_type == VLNK)
348 		return (EMLINK);
349 	if (vp->v_type == VSOCK)
350 		return (EOPNOTSUPP);
351 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
352 		return (ENOTDIR);
353 	accmode = 0;
354 	if (fmode & (FWRITE | O_TRUNC)) {
355 		if (vp->v_type == VDIR)
356 			return (EISDIR);
357 		accmode |= VWRITE;
358 	}
359 	if (fmode & FREAD)
360 		accmode |= VREAD;
361 	if (fmode & FEXEC)
362 		accmode |= VEXEC;
363 	if ((fmode & O_APPEND) && (fmode & FWRITE))
364 		accmode |= VAPPEND;
365 #ifdef MAC
366 	if (fmode & O_CREAT)
367 		accmode |= VCREAT;
368 	if (fmode & O_VERIFY)
369 		accmode |= VVERIFY;
370 	error = mac_vnode_check_open(cred, vp, accmode);
371 	if (error)
372 		return (error);
373 
374 	accmode &= ~(VCREAT | VVERIFY);
375 #endif
376 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
377 		error = VOP_ACCESS(vp, accmode, cred, td);
378 		if (error != 0)
379 			return (error);
380 	}
381 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
382 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
383 	error = VOP_OPEN(vp, fmode, cred, td, fp);
384 	if (error != 0)
385 		return (error);
386 
387 	error = vn_open_vnode_advlock(vp, fmode, fp);
388 	if (error == 0 && (fmode & FWRITE) != 0) {
389 		error = VOP_ADD_WRITECOUNT(vp, 1);
390 		if (error == 0) {
391 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
392 			     __func__, vp, vp->v_writecount);
393 		}
394 	}
395 
396 	/*
397 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
398 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
399 	 * Arrange for that by having fdrop() to use vn_closefile().
400 	 */
401 	if (error != 0) {
402 		fp->f_flag |= FOPENFAILED;
403 		fp->f_vnode = vp;
404 		if (fp->f_ops == &badfileops) {
405 			fp->f_type = DTYPE_VNODE;
406 			fp->f_ops = &vnops;
407 		}
408 		vref(vp);
409 	}
410 
411 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
412 	return (error);
413 
414 }
415 
416 /*
417  * Check for write permissions on the specified vnode.
418  * Prototype text segments cannot be written.
419  * It is racy.
420  */
421 int
422 vn_writechk(struct vnode *vp)
423 {
424 
425 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
426 	/*
427 	 * If there's shared text associated with
428 	 * the vnode, try to free it up once.  If
429 	 * we fail, we can't allow writing.
430 	 */
431 	if (VOP_IS_TEXT(vp))
432 		return (ETXTBSY);
433 
434 	return (0);
435 }
436 
437 /*
438  * Vnode close call
439  */
440 static int
441 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
442     struct thread *td, bool keep_ref)
443 {
444 	struct mount *mp;
445 	int error, lock_flags;
446 
447 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
448 	    MNT_EXTENDED_SHARED(vp->v_mount))
449 		lock_flags = LK_SHARED;
450 	else
451 		lock_flags = LK_EXCLUSIVE;
452 
453 	vn_start_write(vp, &mp, V_WAIT);
454 	vn_lock(vp, lock_flags | LK_RETRY);
455 	AUDIT_ARG_VNODE1(vp);
456 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
457 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
458 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
459 		    __func__, vp, vp->v_writecount);
460 	}
461 	error = VOP_CLOSE(vp, flags, file_cred, td);
462 	if (keep_ref)
463 		VOP_UNLOCK(vp, 0);
464 	else
465 		vput(vp);
466 	vn_finished_write(mp);
467 	return (error);
468 }
469 
470 int
471 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
472     struct thread *td)
473 {
474 
475 	return (vn_close1(vp, flags, file_cred, td, false));
476 }
477 
478 /*
479  * Heuristic to detect sequential operation.
480  */
481 static int
482 sequential_heuristic(struct uio *uio, struct file *fp)
483 {
484 
485 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
486 	if (fp->f_flag & FRDAHEAD)
487 		return (fp->f_seqcount << IO_SEQSHIFT);
488 
489 	/*
490 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
491 	 * that the first I/O is normally considered to be slightly
492 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
493 	 * unless previous seeks have reduced f_seqcount to 0, in which
494 	 * case offset 0 is not special.
495 	 */
496 	if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
497 	    uio->uio_offset == fp->f_nextoff) {
498 		/*
499 		 * f_seqcount is in units of fixed-size blocks so that it
500 		 * depends mainly on the amount of sequential I/O and not
501 		 * much on the number of sequential I/O's.  The fixed size
502 		 * of 16384 is hard-coded here since it is (not quite) just
503 		 * a magic size that works well here.  This size is more
504 		 * closely related to the best I/O size for real disks than
505 		 * to any block size used by software.
506 		 */
507 		if (uio->uio_resid >= IO_SEQMAX * 16384)
508 			fp->f_seqcount = IO_SEQMAX;
509 		else {
510 			fp->f_seqcount += howmany(uio->uio_resid, 16384);
511 			if (fp->f_seqcount > IO_SEQMAX)
512 				fp->f_seqcount = IO_SEQMAX;
513 		}
514 		return (fp->f_seqcount << IO_SEQSHIFT);
515 	}
516 
517 	/* Not sequential.  Quickly draw-down sequentiality. */
518 	if (fp->f_seqcount > 1)
519 		fp->f_seqcount = 1;
520 	else
521 		fp->f_seqcount = 0;
522 	return (0);
523 }
524 
525 /*
526  * Package up an I/O request on a vnode into a uio and do it.
527  */
528 int
529 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
530     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
531     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
532 {
533 	struct uio auio;
534 	struct iovec aiov;
535 	struct mount *mp;
536 	struct ucred *cred;
537 	void *rl_cookie;
538 	struct vn_io_fault_args args;
539 	int error, lock_flags;
540 
541 	if (offset < 0 && vp->v_type != VCHR)
542 		return (EINVAL);
543 	auio.uio_iov = &aiov;
544 	auio.uio_iovcnt = 1;
545 	aiov.iov_base = base;
546 	aiov.iov_len = len;
547 	auio.uio_resid = len;
548 	auio.uio_offset = offset;
549 	auio.uio_segflg = segflg;
550 	auio.uio_rw = rw;
551 	auio.uio_td = td;
552 	error = 0;
553 
554 	if ((ioflg & IO_NODELOCKED) == 0) {
555 		if ((ioflg & IO_RANGELOCKED) == 0) {
556 			if (rw == UIO_READ) {
557 				rl_cookie = vn_rangelock_rlock(vp, offset,
558 				    offset + len);
559 			} else {
560 				rl_cookie = vn_rangelock_wlock(vp, offset,
561 				    offset + len);
562 			}
563 		} else
564 			rl_cookie = NULL;
565 		mp = NULL;
566 		if (rw == UIO_WRITE) {
567 			if (vp->v_type != VCHR &&
568 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
569 			    != 0)
570 				goto out;
571 			if (MNT_SHARED_WRITES(mp) ||
572 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
573 				lock_flags = LK_SHARED;
574 			else
575 				lock_flags = LK_EXCLUSIVE;
576 		} else
577 			lock_flags = LK_SHARED;
578 		vn_lock(vp, lock_flags | LK_RETRY);
579 	} else
580 		rl_cookie = NULL;
581 
582 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
583 #ifdef MAC
584 	if ((ioflg & IO_NOMACCHECK) == 0) {
585 		if (rw == UIO_READ)
586 			error = mac_vnode_check_read(active_cred, file_cred,
587 			    vp);
588 		else
589 			error = mac_vnode_check_write(active_cred, file_cred,
590 			    vp);
591 	}
592 #endif
593 	if (error == 0) {
594 		if (file_cred != NULL)
595 			cred = file_cred;
596 		else
597 			cred = active_cred;
598 		if (do_vn_io_fault(vp, &auio)) {
599 			args.kind = VN_IO_FAULT_VOP;
600 			args.cred = cred;
601 			args.flags = ioflg;
602 			args.args.vop_args.vp = vp;
603 			error = vn_io_fault1(vp, &auio, &args, td);
604 		} else if (rw == UIO_READ) {
605 			error = VOP_READ(vp, &auio, ioflg, cred);
606 		} else /* if (rw == UIO_WRITE) */ {
607 			error = VOP_WRITE(vp, &auio, ioflg, cred);
608 		}
609 	}
610 	if (aresid)
611 		*aresid = auio.uio_resid;
612 	else
613 		if (auio.uio_resid && error == 0)
614 			error = EIO;
615 	if ((ioflg & IO_NODELOCKED) == 0) {
616 		VOP_UNLOCK(vp, 0);
617 		if (mp != NULL)
618 			vn_finished_write(mp);
619 	}
620  out:
621 	if (rl_cookie != NULL)
622 		vn_rangelock_unlock(vp, rl_cookie);
623 	return (error);
624 }
625 
626 /*
627  * Package up an I/O request on a vnode into a uio and do it.  The I/O
628  * request is split up into smaller chunks and we try to avoid saturating
629  * the buffer cache while potentially holding a vnode locked, so we
630  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
631  * to give other processes a chance to lock the vnode (either other processes
632  * core'ing the same binary, or unrelated processes scanning the directory).
633  */
634 int
635 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
636     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
637     struct ucred *file_cred, size_t *aresid, struct thread *td)
638 {
639 	int error = 0;
640 	ssize_t iaresid;
641 
642 	do {
643 		int chunk;
644 
645 		/*
646 		 * Force `offset' to a multiple of MAXBSIZE except possibly
647 		 * for the first chunk, so that filesystems only need to
648 		 * write full blocks except possibly for the first and last
649 		 * chunks.
650 		 */
651 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
652 
653 		if (chunk > len)
654 			chunk = len;
655 		if (rw != UIO_READ && vp->v_type == VREG)
656 			bwillwrite();
657 		iaresid = 0;
658 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
659 		    ioflg, active_cred, file_cred, &iaresid, td);
660 		len -= chunk;	/* aresid calc already includes length */
661 		if (error)
662 			break;
663 		offset += chunk;
664 		base = (char *)base + chunk;
665 		kern_yield(PRI_USER);
666 	} while (len);
667 	if (aresid)
668 		*aresid = len + iaresid;
669 	return (error);
670 }
671 
672 off_t
673 foffset_lock(struct file *fp, int flags)
674 {
675 	struct mtx *mtxp;
676 	off_t res;
677 
678 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
679 
680 #if OFF_MAX <= LONG_MAX
681 	/*
682 	 * Caller only wants the current f_offset value.  Assume that
683 	 * the long and shorter integer types reads are atomic.
684 	 */
685 	if ((flags & FOF_NOLOCK) != 0)
686 		return (fp->f_offset);
687 #endif
688 
689 	/*
690 	 * According to McKusick the vn lock was protecting f_offset here.
691 	 * It is now protected by the FOFFSET_LOCKED flag.
692 	 */
693 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
694 	mtx_lock(mtxp);
695 	if ((flags & FOF_NOLOCK) == 0) {
696 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
697 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
698 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
699 			    "vofflock", 0);
700 		}
701 		fp->f_vnread_flags |= FOFFSET_LOCKED;
702 	}
703 	res = fp->f_offset;
704 	mtx_unlock(mtxp);
705 	return (res);
706 }
707 
708 void
709 foffset_unlock(struct file *fp, off_t val, int flags)
710 {
711 	struct mtx *mtxp;
712 
713 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
714 
715 #if OFF_MAX <= LONG_MAX
716 	if ((flags & FOF_NOLOCK) != 0) {
717 		if ((flags & FOF_NOUPDATE) == 0)
718 			fp->f_offset = val;
719 		if ((flags & FOF_NEXTOFF) != 0)
720 			fp->f_nextoff = val;
721 		return;
722 	}
723 #endif
724 
725 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
726 	mtx_lock(mtxp);
727 	if ((flags & FOF_NOUPDATE) == 0)
728 		fp->f_offset = val;
729 	if ((flags & FOF_NEXTOFF) != 0)
730 		fp->f_nextoff = val;
731 	if ((flags & FOF_NOLOCK) == 0) {
732 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
733 		    ("Lost FOFFSET_LOCKED"));
734 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
735 			wakeup(&fp->f_vnread_flags);
736 		fp->f_vnread_flags = 0;
737 	}
738 	mtx_unlock(mtxp);
739 }
740 
741 void
742 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
743 {
744 
745 	if ((flags & FOF_OFFSET) == 0)
746 		uio->uio_offset = foffset_lock(fp, flags);
747 }
748 
749 void
750 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
751 {
752 
753 	if ((flags & FOF_OFFSET) == 0)
754 		foffset_unlock(fp, uio->uio_offset, flags);
755 }
756 
757 static int
758 get_advice(struct file *fp, struct uio *uio)
759 {
760 	struct mtx *mtxp;
761 	int ret;
762 
763 	ret = POSIX_FADV_NORMAL;
764 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
765 		return (ret);
766 
767 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
768 	mtx_lock(mtxp);
769 	if (fp->f_advice != NULL &&
770 	    uio->uio_offset >= fp->f_advice->fa_start &&
771 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
772 		ret = fp->f_advice->fa_advice;
773 	mtx_unlock(mtxp);
774 	return (ret);
775 }
776 
777 /*
778  * File table vnode read routine.
779  */
780 static int
781 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
782     struct thread *td)
783 {
784 	struct vnode *vp;
785 	off_t orig_offset;
786 	int error, ioflag;
787 	int advice;
788 
789 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
790 	    uio->uio_td, td));
791 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
792 	vp = fp->f_vnode;
793 	ioflag = 0;
794 	if (fp->f_flag & FNONBLOCK)
795 		ioflag |= IO_NDELAY;
796 	if (fp->f_flag & O_DIRECT)
797 		ioflag |= IO_DIRECT;
798 	advice = get_advice(fp, uio);
799 	vn_lock(vp, LK_SHARED | LK_RETRY);
800 
801 	switch (advice) {
802 	case POSIX_FADV_NORMAL:
803 	case POSIX_FADV_SEQUENTIAL:
804 	case POSIX_FADV_NOREUSE:
805 		ioflag |= sequential_heuristic(uio, fp);
806 		break;
807 	case POSIX_FADV_RANDOM:
808 		/* Disable read-ahead for random I/O. */
809 		break;
810 	}
811 	orig_offset = uio->uio_offset;
812 
813 #ifdef MAC
814 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
815 	if (error == 0)
816 #endif
817 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
818 	fp->f_nextoff = uio->uio_offset;
819 	VOP_UNLOCK(vp, 0);
820 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
821 	    orig_offset != uio->uio_offset)
822 		/*
823 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
824 		 * for the backing file after a POSIX_FADV_NOREUSE
825 		 * read(2).
826 		 */
827 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
828 		    POSIX_FADV_DONTNEED);
829 	return (error);
830 }
831 
832 /*
833  * File table vnode write routine.
834  */
835 static int
836 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
837     struct thread *td)
838 {
839 	struct vnode *vp;
840 	struct mount *mp;
841 	off_t orig_offset;
842 	int error, ioflag, lock_flags;
843 	int advice;
844 
845 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
846 	    uio->uio_td, td));
847 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
848 	vp = fp->f_vnode;
849 	if (vp->v_type == VREG)
850 		bwillwrite();
851 	ioflag = IO_UNIT;
852 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
853 		ioflag |= IO_APPEND;
854 	if (fp->f_flag & FNONBLOCK)
855 		ioflag |= IO_NDELAY;
856 	if (fp->f_flag & O_DIRECT)
857 		ioflag |= IO_DIRECT;
858 	if ((fp->f_flag & O_FSYNC) ||
859 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
860 		ioflag |= IO_SYNC;
861 	mp = NULL;
862 	if (vp->v_type != VCHR &&
863 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
864 		goto unlock;
865 
866 	advice = get_advice(fp, uio);
867 
868 	if (MNT_SHARED_WRITES(mp) ||
869 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
870 		lock_flags = LK_SHARED;
871 	} else {
872 		lock_flags = LK_EXCLUSIVE;
873 	}
874 
875 	vn_lock(vp, lock_flags | LK_RETRY);
876 	switch (advice) {
877 	case POSIX_FADV_NORMAL:
878 	case POSIX_FADV_SEQUENTIAL:
879 	case POSIX_FADV_NOREUSE:
880 		ioflag |= sequential_heuristic(uio, fp);
881 		break;
882 	case POSIX_FADV_RANDOM:
883 		/* XXX: Is this correct? */
884 		break;
885 	}
886 	orig_offset = uio->uio_offset;
887 
888 #ifdef MAC
889 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
890 	if (error == 0)
891 #endif
892 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
893 	fp->f_nextoff = uio->uio_offset;
894 	VOP_UNLOCK(vp, 0);
895 	if (vp->v_type != VCHR)
896 		vn_finished_write(mp);
897 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
898 	    orig_offset != uio->uio_offset)
899 		/*
900 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
901 		 * for the backing file after a POSIX_FADV_NOREUSE
902 		 * write(2).
903 		 */
904 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
905 		    POSIX_FADV_DONTNEED);
906 unlock:
907 	return (error);
908 }
909 
910 /*
911  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
912  * prevent the following deadlock:
913  *
914  * Assume that the thread A reads from the vnode vp1 into userspace
915  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
916  * currently not resident, then system ends up with the call chain
917  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
918  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
919  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
920  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
921  * backed by the pages of vnode vp1, and some page in buf2 is not
922  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
923  *
924  * To prevent the lock order reversal and deadlock, vn_io_fault() does
925  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
926  * Instead, it first tries to do the whole range i/o with pagefaults
927  * disabled. If all pages in the i/o buffer are resident and mapped,
928  * VOP will succeed (ignoring the genuine filesystem errors).
929  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
930  * i/o in chunks, with all pages in the chunk prefaulted and held
931  * using vm_fault_quick_hold_pages().
932  *
933  * Filesystems using this deadlock avoidance scheme should use the
934  * array of the held pages from uio, saved in the curthread->td_ma,
935  * instead of doing uiomove().  A helper function
936  * vn_io_fault_uiomove() converts uiomove request into
937  * uiomove_fromphys() over td_ma array.
938  *
939  * Since vnode locks do not cover the whole i/o anymore, rangelocks
940  * make the current i/o request atomic with respect to other i/os and
941  * truncations.
942  */
943 
944 /*
945  * Decode vn_io_fault_args and perform the corresponding i/o.
946  */
947 static int
948 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
949     struct thread *td)
950 {
951 	int error, save;
952 
953 	error = 0;
954 	save = vm_fault_disable_pagefaults();
955 	switch (args->kind) {
956 	case VN_IO_FAULT_FOP:
957 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
958 		    uio, args->cred, args->flags, td);
959 		break;
960 	case VN_IO_FAULT_VOP:
961 		if (uio->uio_rw == UIO_READ) {
962 			error = VOP_READ(args->args.vop_args.vp, uio,
963 			    args->flags, args->cred);
964 		} else if (uio->uio_rw == UIO_WRITE) {
965 			error = VOP_WRITE(args->args.vop_args.vp, uio,
966 			    args->flags, args->cred);
967 		}
968 		break;
969 	default:
970 		panic("vn_io_fault_doio: unknown kind of io %d %d",
971 		    args->kind, uio->uio_rw);
972 	}
973 	vm_fault_enable_pagefaults(save);
974 	return (error);
975 }
976 
977 static int
978 vn_io_fault_touch(char *base, const struct uio *uio)
979 {
980 	int r;
981 
982 	r = fubyte(base);
983 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
984 		return (EFAULT);
985 	return (0);
986 }
987 
988 static int
989 vn_io_fault_prefault_user(const struct uio *uio)
990 {
991 	char *base;
992 	const struct iovec *iov;
993 	size_t len;
994 	ssize_t resid;
995 	int error, i;
996 
997 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
998 	    ("vn_io_fault_prefault userspace"));
999 
1000 	error = i = 0;
1001 	iov = uio->uio_iov;
1002 	resid = uio->uio_resid;
1003 	base = iov->iov_base;
1004 	len = iov->iov_len;
1005 	while (resid > 0) {
1006 		error = vn_io_fault_touch(base, uio);
1007 		if (error != 0)
1008 			break;
1009 		if (len < PAGE_SIZE) {
1010 			if (len != 0) {
1011 				error = vn_io_fault_touch(base + len - 1, uio);
1012 				if (error != 0)
1013 					break;
1014 				resid -= len;
1015 			}
1016 			if (++i >= uio->uio_iovcnt)
1017 				break;
1018 			iov = uio->uio_iov + i;
1019 			base = iov->iov_base;
1020 			len = iov->iov_len;
1021 		} else {
1022 			len -= PAGE_SIZE;
1023 			base += PAGE_SIZE;
1024 			resid -= PAGE_SIZE;
1025 		}
1026 	}
1027 	return (error);
1028 }
1029 
1030 /*
1031  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1032  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1033  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1034  * into args and call vn_io_fault1() to handle faults during the user
1035  * mode buffer accesses.
1036  */
1037 static int
1038 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1039     struct thread *td)
1040 {
1041 	vm_page_t ma[io_hold_cnt + 2];
1042 	struct uio *uio_clone, short_uio;
1043 	struct iovec short_iovec[1];
1044 	vm_page_t *prev_td_ma;
1045 	vm_prot_t prot;
1046 	vm_offset_t addr, end;
1047 	size_t len, resid;
1048 	ssize_t adv;
1049 	int error, cnt, saveheld, prev_td_ma_cnt;
1050 
1051 	if (vn_io_fault_prefault) {
1052 		error = vn_io_fault_prefault_user(uio);
1053 		if (error != 0)
1054 			return (error); /* Or ignore ? */
1055 	}
1056 
1057 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1058 
1059 	/*
1060 	 * The UFS follows IO_UNIT directive and replays back both
1061 	 * uio_offset and uio_resid if an error is encountered during the
1062 	 * operation.  But, since the iovec may be already advanced,
1063 	 * uio is still in an inconsistent state.
1064 	 *
1065 	 * Cache a copy of the original uio, which is advanced to the redo
1066 	 * point using UIO_NOCOPY below.
1067 	 */
1068 	uio_clone = cloneuio(uio);
1069 	resid = uio->uio_resid;
1070 
1071 	short_uio.uio_segflg = UIO_USERSPACE;
1072 	short_uio.uio_rw = uio->uio_rw;
1073 	short_uio.uio_td = uio->uio_td;
1074 
1075 	error = vn_io_fault_doio(args, uio, td);
1076 	if (error != EFAULT)
1077 		goto out;
1078 
1079 	atomic_add_long(&vn_io_faults_cnt, 1);
1080 	uio_clone->uio_segflg = UIO_NOCOPY;
1081 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1082 	uio_clone->uio_segflg = uio->uio_segflg;
1083 
1084 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1085 	prev_td_ma = td->td_ma;
1086 	prev_td_ma_cnt = td->td_ma_cnt;
1087 
1088 	while (uio_clone->uio_resid != 0) {
1089 		len = uio_clone->uio_iov->iov_len;
1090 		if (len == 0) {
1091 			KASSERT(uio_clone->uio_iovcnt >= 1,
1092 			    ("iovcnt underflow"));
1093 			uio_clone->uio_iov++;
1094 			uio_clone->uio_iovcnt--;
1095 			continue;
1096 		}
1097 		if (len > io_hold_cnt * PAGE_SIZE)
1098 			len = io_hold_cnt * PAGE_SIZE;
1099 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1100 		end = round_page(addr + len);
1101 		if (end < addr) {
1102 			error = EFAULT;
1103 			break;
1104 		}
1105 		cnt = atop(end - trunc_page(addr));
1106 		/*
1107 		 * A perfectly misaligned address and length could cause
1108 		 * both the start and the end of the chunk to use partial
1109 		 * page.  +2 accounts for such a situation.
1110 		 */
1111 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1112 		    addr, len, prot, ma, io_hold_cnt + 2);
1113 		if (cnt == -1) {
1114 			error = EFAULT;
1115 			break;
1116 		}
1117 		short_uio.uio_iov = &short_iovec[0];
1118 		short_iovec[0].iov_base = (void *)addr;
1119 		short_uio.uio_iovcnt = 1;
1120 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1121 		short_uio.uio_offset = uio_clone->uio_offset;
1122 		td->td_ma = ma;
1123 		td->td_ma_cnt = cnt;
1124 
1125 		error = vn_io_fault_doio(args, &short_uio, td);
1126 		vm_page_unhold_pages(ma, cnt);
1127 		adv = len - short_uio.uio_resid;
1128 
1129 		uio_clone->uio_iov->iov_base =
1130 		    (char *)uio_clone->uio_iov->iov_base + adv;
1131 		uio_clone->uio_iov->iov_len -= adv;
1132 		uio_clone->uio_resid -= adv;
1133 		uio_clone->uio_offset += adv;
1134 
1135 		uio->uio_resid -= adv;
1136 		uio->uio_offset += adv;
1137 
1138 		if (error != 0 || adv == 0)
1139 			break;
1140 	}
1141 	td->td_ma = prev_td_ma;
1142 	td->td_ma_cnt = prev_td_ma_cnt;
1143 	curthread_pflags_restore(saveheld);
1144 out:
1145 	free(uio_clone, M_IOV);
1146 	return (error);
1147 }
1148 
1149 static int
1150 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1151     int flags, struct thread *td)
1152 {
1153 	fo_rdwr_t *doio;
1154 	struct vnode *vp;
1155 	void *rl_cookie;
1156 	struct vn_io_fault_args args;
1157 	int error;
1158 
1159 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1160 	vp = fp->f_vnode;
1161 	foffset_lock_uio(fp, uio, flags);
1162 	if (do_vn_io_fault(vp, uio)) {
1163 		args.kind = VN_IO_FAULT_FOP;
1164 		args.args.fop_args.fp = fp;
1165 		args.args.fop_args.doio = doio;
1166 		args.cred = active_cred;
1167 		args.flags = flags | FOF_OFFSET;
1168 		if (uio->uio_rw == UIO_READ) {
1169 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1170 			    uio->uio_offset + uio->uio_resid);
1171 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1172 		    (flags & FOF_OFFSET) == 0) {
1173 			/* For appenders, punt and lock the whole range. */
1174 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1175 		} else {
1176 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1177 			    uio->uio_offset + uio->uio_resid);
1178 		}
1179 		error = vn_io_fault1(vp, uio, &args, td);
1180 		vn_rangelock_unlock(vp, rl_cookie);
1181 	} else {
1182 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1183 	}
1184 	foffset_unlock_uio(fp, uio, flags);
1185 	return (error);
1186 }
1187 
1188 /*
1189  * Helper function to perform the requested uiomove operation using
1190  * the held pages for io->uio_iov[0].iov_base buffer instead of
1191  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1192  * instead of iov_base prevents page faults that could occur due to
1193  * pmap_collect() invalidating the mapping created by
1194  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1195  * object cleanup revoking the write access from page mappings.
1196  *
1197  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1198  * instead of plain uiomove().
1199  */
1200 int
1201 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1202 {
1203 	struct uio transp_uio;
1204 	struct iovec transp_iov[1];
1205 	struct thread *td;
1206 	size_t adv;
1207 	int error, pgadv;
1208 
1209 	td = curthread;
1210 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1211 	    uio->uio_segflg != UIO_USERSPACE)
1212 		return (uiomove(data, xfersize, uio));
1213 
1214 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1215 	transp_iov[0].iov_base = data;
1216 	transp_uio.uio_iov = &transp_iov[0];
1217 	transp_uio.uio_iovcnt = 1;
1218 	if (xfersize > uio->uio_resid)
1219 		xfersize = uio->uio_resid;
1220 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1221 	transp_uio.uio_offset = 0;
1222 	transp_uio.uio_segflg = UIO_SYSSPACE;
1223 	/*
1224 	 * Since transp_iov points to data, and td_ma page array
1225 	 * corresponds to original uio->uio_iov, we need to invert the
1226 	 * direction of the i/o operation as passed to
1227 	 * uiomove_fromphys().
1228 	 */
1229 	switch (uio->uio_rw) {
1230 	case UIO_WRITE:
1231 		transp_uio.uio_rw = UIO_READ;
1232 		break;
1233 	case UIO_READ:
1234 		transp_uio.uio_rw = UIO_WRITE;
1235 		break;
1236 	}
1237 	transp_uio.uio_td = uio->uio_td;
1238 	error = uiomove_fromphys(td->td_ma,
1239 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1240 	    xfersize, &transp_uio);
1241 	adv = xfersize - transp_uio.uio_resid;
1242 	pgadv =
1243 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1244 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1245 	td->td_ma += pgadv;
1246 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1247 	    pgadv));
1248 	td->td_ma_cnt -= pgadv;
1249 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1250 	uio->uio_iov->iov_len -= adv;
1251 	uio->uio_resid -= adv;
1252 	uio->uio_offset += adv;
1253 	return (error);
1254 }
1255 
1256 int
1257 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1258     struct uio *uio)
1259 {
1260 	struct thread *td;
1261 	vm_offset_t iov_base;
1262 	int cnt, pgadv;
1263 
1264 	td = curthread;
1265 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1266 	    uio->uio_segflg != UIO_USERSPACE)
1267 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1268 
1269 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1270 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1271 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1272 	switch (uio->uio_rw) {
1273 	case UIO_WRITE:
1274 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1275 		    offset, cnt);
1276 		break;
1277 	case UIO_READ:
1278 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1279 		    cnt);
1280 		break;
1281 	}
1282 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1283 	td->td_ma += pgadv;
1284 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1285 	    pgadv));
1286 	td->td_ma_cnt -= pgadv;
1287 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1288 	uio->uio_iov->iov_len -= cnt;
1289 	uio->uio_resid -= cnt;
1290 	uio->uio_offset += cnt;
1291 	return (0);
1292 }
1293 
1294 
1295 /*
1296  * File table truncate routine.
1297  */
1298 static int
1299 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1300     struct thread *td)
1301 {
1302 	struct mount *mp;
1303 	struct vnode *vp;
1304 	void *rl_cookie;
1305 	int error;
1306 
1307 	vp = fp->f_vnode;
1308 
1309 	/*
1310 	 * Lock the whole range for truncation.  Otherwise split i/o
1311 	 * might happen partly before and partly after the truncation.
1312 	 */
1313 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1314 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1315 	if (error)
1316 		goto out1;
1317 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1318 	AUDIT_ARG_VNODE1(vp);
1319 	if (vp->v_type == VDIR) {
1320 		error = EISDIR;
1321 		goto out;
1322 	}
1323 #ifdef MAC
1324 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1325 	if (error)
1326 		goto out;
1327 #endif
1328 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1329 	    fp->f_cred);
1330 out:
1331 	VOP_UNLOCK(vp, 0);
1332 	vn_finished_write(mp);
1333 out1:
1334 	vn_rangelock_unlock(vp, rl_cookie);
1335 	return (error);
1336 }
1337 
1338 /*
1339  * Truncate a file that is already locked.
1340  */
1341 int
1342 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1343     struct ucred *cred)
1344 {
1345 	struct vattr vattr;
1346 	int error;
1347 
1348 	error = VOP_ADD_WRITECOUNT(vp, 1);
1349 	if (error == 0) {
1350 		VATTR_NULL(&vattr);
1351 		vattr.va_size = length;
1352 		if (sync)
1353 			vattr.va_vaflags |= VA_SYNC;
1354 		error = VOP_SETATTR(vp, &vattr, cred);
1355 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1356 	}
1357 	return (error);
1358 }
1359 
1360 /*
1361  * File table vnode stat routine.
1362  */
1363 static int
1364 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1365     struct thread *td)
1366 {
1367 	struct vnode *vp = fp->f_vnode;
1368 	int error;
1369 
1370 	vn_lock(vp, LK_SHARED | LK_RETRY);
1371 	error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1372 	VOP_UNLOCK(vp, 0);
1373 
1374 	return (error);
1375 }
1376 
1377 /*
1378  * Stat a vnode; implementation for the stat syscall
1379  */
1380 int
1381 vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred,
1382     struct ucred *file_cred, struct thread *td)
1383 {
1384 	struct vattr vattr;
1385 	struct vattr *vap;
1386 	int error;
1387 	u_short mode;
1388 
1389 	AUDIT_ARG_VNODE1(vp);
1390 #ifdef MAC
1391 	error = mac_vnode_check_stat(active_cred, file_cred, vp);
1392 	if (error)
1393 		return (error);
1394 #endif
1395 
1396 	vap = &vattr;
1397 
1398 	/*
1399 	 * Initialize defaults for new and unusual fields, so that file
1400 	 * systems which don't support these fields don't need to know
1401 	 * about them.
1402 	 */
1403 	vap->va_birthtime.tv_sec = -1;
1404 	vap->va_birthtime.tv_nsec = 0;
1405 	vap->va_fsid = VNOVAL;
1406 	vap->va_rdev = NODEV;
1407 
1408 	error = VOP_GETATTR(vp, vap, active_cred);
1409 	if (error)
1410 		return (error);
1411 
1412 	/*
1413 	 * Zero the spare stat fields
1414 	 */
1415 	bzero(sb, sizeof *sb);
1416 
1417 	/*
1418 	 * Copy from vattr table
1419 	 */
1420 	if (vap->va_fsid != VNOVAL)
1421 		sb->st_dev = vap->va_fsid;
1422 	else
1423 		sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1424 	sb->st_ino = vap->va_fileid;
1425 	mode = vap->va_mode;
1426 	switch (vap->va_type) {
1427 	case VREG:
1428 		mode |= S_IFREG;
1429 		break;
1430 	case VDIR:
1431 		mode |= S_IFDIR;
1432 		break;
1433 	case VBLK:
1434 		mode |= S_IFBLK;
1435 		break;
1436 	case VCHR:
1437 		mode |= S_IFCHR;
1438 		break;
1439 	case VLNK:
1440 		mode |= S_IFLNK;
1441 		break;
1442 	case VSOCK:
1443 		mode |= S_IFSOCK;
1444 		break;
1445 	case VFIFO:
1446 		mode |= S_IFIFO;
1447 		break;
1448 	default:
1449 		return (EBADF);
1450 	}
1451 	sb->st_mode = mode;
1452 	sb->st_nlink = vap->va_nlink;
1453 	sb->st_uid = vap->va_uid;
1454 	sb->st_gid = vap->va_gid;
1455 	sb->st_rdev = vap->va_rdev;
1456 	if (vap->va_size > OFF_MAX)
1457 		return (EOVERFLOW);
1458 	sb->st_size = vap->va_size;
1459 	sb->st_atim.tv_sec = vap->va_atime.tv_sec;
1460 	sb->st_atim.tv_nsec = vap->va_atime.tv_nsec;
1461 	sb->st_mtim.tv_sec = vap->va_mtime.tv_sec;
1462 	sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec;
1463 	sb->st_ctim.tv_sec = vap->va_ctime.tv_sec;
1464 	sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec;
1465 	sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec;
1466 	sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec;
1467 
1468         /*
1469 	 * According to www.opengroup.org, the meaning of st_blksize is
1470 	 *   "a filesystem-specific preferred I/O block size for this
1471 	 *    object.  In some filesystem types, this may vary from file
1472 	 *    to file"
1473 	 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1474 	 */
1475 
1476 	sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1477 
1478 	sb->st_flags = vap->va_flags;
1479 	if (priv_check(td, PRIV_VFS_GENERATION))
1480 		sb->st_gen = 0;
1481 	else
1482 		sb->st_gen = vap->va_gen;
1483 
1484 	sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1485 	return (0);
1486 }
1487 
1488 /*
1489  * File table vnode ioctl routine.
1490  */
1491 static int
1492 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1493     struct thread *td)
1494 {
1495 	struct vattr vattr;
1496 	struct vnode *vp;
1497 	struct fiobmap2_arg *bmarg;
1498 	int error;
1499 
1500 	vp = fp->f_vnode;
1501 	switch (vp->v_type) {
1502 	case VDIR:
1503 	case VREG:
1504 		switch (com) {
1505 		case FIONREAD:
1506 			vn_lock(vp, LK_SHARED | LK_RETRY);
1507 			error = VOP_GETATTR(vp, &vattr, active_cred);
1508 			VOP_UNLOCK(vp, 0);
1509 			if (error == 0)
1510 				*(int *)data = vattr.va_size - fp->f_offset;
1511 			return (error);
1512 		case FIOBMAP2:
1513 			bmarg = (struct fiobmap2_arg *)data;
1514 			vn_lock(vp, LK_SHARED | LK_RETRY);
1515 #ifdef MAC
1516 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1517 			    vp);
1518 			if (error == 0)
1519 #endif
1520 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1521 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1522 			VOP_UNLOCK(vp, 0);
1523 			return (error);
1524 		case FIONBIO:
1525 		case FIOASYNC:
1526 			return (0);
1527 		default:
1528 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1529 			    active_cred, td));
1530 		}
1531 		break;
1532 	case VCHR:
1533 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1534 		    active_cred, td));
1535 	default:
1536 		return (ENOTTY);
1537 	}
1538 }
1539 
1540 /*
1541  * File table vnode poll routine.
1542  */
1543 static int
1544 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1545     struct thread *td)
1546 {
1547 	struct vnode *vp;
1548 	int error;
1549 
1550 	vp = fp->f_vnode;
1551 #ifdef MAC
1552 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1553 	AUDIT_ARG_VNODE1(vp);
1554 	error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1555 	VOP_UNLOCK(vp, 0);
1556 	if (!error)
1557 #endif
1558 
1559 	error = VOP_POLL(vp, events, fp->f_cred, td);
1560 	return (error);
1561 }
1562 
1563 /*
1564  * Acquire the requested lock and then check for validity.  LK_RETRY
1565  * permits vn_lock to return doomed vnodes.
1566  */
1567 int
1568 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1569 {
1570 	int error;
1571 
1572 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1573 	    ("vn_lock: no locktype"));
1574 	VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count"));
1575 retry:
1576 	error = VOP_LOCK1(vp, flags, file, line);
1577 	flags &= ~LK_INTERLOCK;	/* Interlock is always dropped. */
1578 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1579 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1580 
1581 	if ((flags & LK_RETRY) == 0) {
1582 		if (error == 0 && VN_IS_DOOMED(vp)) {
1583 			VOP_UNLOCK(vp, 0);
1584 			error = ENOENT;
1585 		}
1586 	} else if (error != 0)
1587 		goto retry;
1588 	return (error);
1589 }
1590 
1591 /*
1592  * File table vnode close routine.
1593  */
1594 static int
1595 vn_closefile(struct file *fp, struct thread *td)
1596 {
1597 	struct vnode *vp;
1598 	struct flock lf;
1599 	int error;
1600 	bool ref;
1601 
1602 	vp = fp->f_vnode;
1603 	fp->f_ops = &badfileops;
1604 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1605 
1606 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1607 
1608 	if (__predict_false(ref)) {
1609 		lf.l_whence = SEEK_SET;
1610 		lf.l_start = 0;
1611 		lf.l_len = 0;
1612 		lf.l_type = F_UNLCK;
1613 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1614 		vrele(vp);
1615 	}
1616 	return (error);
1617 }
1618 
1619 static bool
1620 vn_suspendable(struct mount *mp)
1621 {
1622 
1623 	return (mp->mnt_op->vfs_susp_clean != NULL);
1624 }
1625 
1626 /*
1627  * Preparing to start a filesystem write operation. If the operation is
1628  * permitted, then we bump the count of operations in progress and
1629  * proceed. If a suspend request is in progress, we wait until the
1630  * suspension is over, and then proceed.
1631  */
1632 static int
1633 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1634 {
1635 	int error, mflags;
1636 
1637 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1638 	    vfs_op_thread_enter(mp)) {
1639 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1640 		vfs_mp_count_add_pcpu(mp, writeopcount, 1);
1641 		vfs_op_thread_exit(mp);
1642 		return (0);
1643 	}
1644 
1645 	if (mplocked)
1646 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1647 	else
1648 		MNT_ILOCK(mp);
1649 
1650 	error = 0;
1651 
1652 	/*
1653 	 * Check on status of suspension.
1654 	 */
1655 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1656 	    mp->mnt_susp_owner != curthread) {
1657 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1658 		    (flags & PCATCH) : 0) | (PUSER - 1);
1659 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1660 			if (flags & V_NOWAIT) {
1661 				error = EWOULDBLOCK;
1662 				goto unlock;
1663 			}
1664 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1665 			    "suspfs", 0);
1666 			if (error)
1667 				goto unlock;
1668 		}
1669 	}
1670 	if (flags & V_XSLEEP)
1671 		goto unlock;
1672 	mp->mnt_writeopcount++;
1673 unlock:
1674 	if (error != 0 || (flags & V_XSLEEP) != 0)
1675 		MNT_REL(mp);
1676 	MNT_IUNLOCK(mp);
1677 	return (error);
1678 }
1679 
1680 int
1681 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1682 {
1683 	struct mount *mp;
1684 	int error;
1685 
1686 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1687 	    ("V_MNTREF requires mp"));
1688 
1689 	error = 0;
1690 	/*
1691 	 * If a vnode is provided, get and return the mount point that
1692 	 * to which it will write.
1693 	 */
1694 	if (vp != NULL) {
1695 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1696 			*mpp = NULL;
1697 			if (error != EOPNOTSUPP)
1698 				return (error);
1699 			return (0);
1700 		}
1701 	}
1702 	if ((mp = *mpp) == NULL)
1703 		return (0);
1704 
1705 	if (!vn_suspendable(mp)) {
1706 		if (vp != NULL || (flags & V_MNTREF) != 0)
1707 			vfs_rel(mp);
1708 		return (0);
1709 	}
1710 
1711 	/*
1712 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1713 	 * a vfs_ref().
1714 	 * As long as a vnode is not provided we need to acquire a
1715 	 * refcount for the provided mountpoint too, in order to
1716 	 * emulate a vfs_ref().
1717 	 */
1718 	if (vp == NULL && (flags & V_MNTREF) == 0)
1719 		vfs_ref(mp);
1720 
1721 	return (vn_start_write_refed(mp, flags, false));
1722 }
1723 
1724 /*
1725  * Secondary suspension. Used by operations such as vop_inactive
1726  * routines that are needed by the higher level functions. These
1727  * are allowed to proceed until all the higher level functions have
1728  * completed (indicated by mnt_writeopcount dropping to zero). At that
1729  * time, these operations are halted until the suspension is over.
1730  */
1731 int
1732 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1733 {
1734 	struct mount *mp;
1735 	int error;
1736 
1737 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1738 	    ("V_MNTREF requires mp"));
1739 
1740  retry:
1741 	if (vp != NULL) {
1742 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1743 			*mpp = NULL;
1744 			if (error != EOPNOTSUPP)
1745 				return (error);
1746 			return (0);
1747 		}
1748 	}
1749 	/*
1750 	 * If we are not suspended or have not yet reached suspended
1751 	 * mode, then let the operation proceed.
1752 	 */
1753 	if ((mp = *mpp) == NULL)
1754 		return (0);
1755 
1756 	if (!vn_suspendable(mp)) {
1757 		if (vp != NULL || (flags & V_MNTREF) != 0)
1758 			vfs_rel(mp);
1759 		return (0);
1760 	}
1761 
1762 	/*
1763 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1764 	 * a vfs_ref().
1765 	 * As long as a vnode is not provided we need to acquire a
1766 	 * refcount for the provided mountpoint too, in order to
1767 	 * emulate a vfs_ref().
1768 	 */
1769 	MNT_ILOCK(mp);
1770 	if (vp == NULL && (flags & V_MNTREF) == 0)
1771 		MNT_REF(mp);
1772 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1773 		mp->mnt_secondary_writes++;
1774 		mp->mnt_secondary_accwrites++;
1775 		MNT_IUNLOCK(mp);
1776 		return (0);
1777 	}
1778 	if (flags & V_NOWAIT) {
1779 		MNT_REL(mp);
1780 		MNT_IUNLOCK(mp);
1781 		return (EWOULDBLOCK);
1782 	}
1783 	/*
1784 	 * Wait for the suspension to finish.
1785 	 */
1786 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1787 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1788 	    "suspfs", 0);
1789 	vfs_rel(mp);
1790 	if (error == 0)
1791 		goto retry;
1792 	return (error);
1793 }
1794 
1795 /*
1796  * Filesystem write operation has completed. If we are suspending and this
1797  * operation is the last one, notify the suspender that the suspension is
1798  * now in effect.
1799  */
1800 void
1801 vn_finished_write(struct mount *mp)
1802 {
1803 	int c;
1804 
1805 	if (mp == NULL || !vn_suspendable(mp))
1806 		return;
1807 
1808 	if (vfs_op_thread_enter(mp)) {
1809 		vfs_mp_count_sub_pcpu(mp, writeopcount, 1);
1810 		vfs_mp_count_sub_pcpu(mp, ref, 1);
1811 		vfs_op_thread_exit(mp);
1812 		return;
1813 	}
1814 
1815 	MNT_ILOCK(mp);
1816 	vfs_assert_mount_counters(mp);
1817 	MNT_REL(mp);
1818 	c = --mp->mnt_writeopcount;
1819 	if (mp->mnt_vfs_ops == 0) {
1820 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1821 		MNT_IUNLOCK(mp);
1822 		return;
1823 	}
1824 	if (c < 0)
1825 		vfs_dump_mount_counters(mp);
1826 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1827 		wakeup(&mp->mnt_writeopcount);
1828 	MNT_IUNLOCK(mp);
1829 }
1830 
1831 
1832 /*
1833  * Filesystem secondary write operation has completed. If we are
1834  * suspending and this operation is the last one, notify the suspender
1835  * that the suspension is now in effect.
1836  */
1837 void
1838 vn_finished_secondary_write(struct mount *mp)
1839 {
1840 	if (mp == NULL || !vn_suspendable(mp))
1841 		return;
1842 	MNT_ILOCK(mp);
1843 	MNT_REL(mp);
1844 	mp->mnt_secondary_writes--;
1845 	if (mp->mnt_secondary_writes < 0)
1846 		panic("vn_finished_secondary_write: neg cnt");
1847 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1848 	    mp->mnt_secondary_writes <= 0)
1849 		wakeup(&mp->mnt_secondary_writes);
1850 	MNT_IUNLOCK(mp);
1851 }
1852 
1853 
1854 
1855 /*
1856  * Request a filesystem to suspend write operations.
1857  */
1858 int
1859 vfs_write_suspend(struct mount *mp, int flags)
1860 {
1861 	int error;
1862 
1863 	MPASS(vn_suspendable(mp));
1864 
1865 	vfs_op_enter(mp);
1866 
1867 	MNT_ILOCK(mp);
1868 	vfs_assert_mount_counters(mp);
1869 	if (mp->mnt_susp_owner == curthread) {
1870 		vfs_op_exit_locked(mp);
1871 		MNT_IUNLOCK(mp);
1872 		return (EALREADY);
1873 	}
1874 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
1875 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1876 
1877 	/*
1878 	 * Unmount holds a write reference on the mount point.  If we
1879 	 * own busy reference and drain for writers, we deadlock with
1880 	 * the reference draining in the unmount path.  Callers of
1881 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1882 	 * vfs_busy() reference is owned and caller is not in the
1883 	 * unmount context.
1884 	 */
1885 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1886 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1887 		vfs_op_exit_locked(mp);
1888 		MNT_IUNLOCK(mp);
1889 		return (EBUSY);
1890 	}
1891 
1892 	mp->mnt_kern_flag |= MNTK_SUSPEND;
1893 	mp->mnt_susp_owner = curthread;
1894 	if (mp->mnt_writeopcount > 0)
1895 		(void) msleep(&mp->mnt_writeopcount,
1896 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1897 	else
1898 		MNT_IUNLOCK(mp);
1899 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
1900 		vfs_write_resume(mp, 0);
1901 		vfs_op_exit(mp);
1902 	}
1903 	return (error);
1904 }
1905 
1906 /*
1907  * Request a filesystem to resume write operations.
1908  */
1909 void
1910 vfs_write_resume(struct mount *mp, int flags)
1911 {
1912 
1913 	MPASS(vn_suspendable(mp));
1914 
1915 	MNT_ILOCK(mp);
1916 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1917 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1918 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1919 				       MNTK_SUSPENDED);
1920 		mp->mnt_susp_owner = NULL;
1921 		wakeup(&mp->mnt_writeopcount);
1922 		wakeup(&mp->mnt_flag);
1923 		curthread->td_pflags &= ~TDP_IGNSUSP;
1924 		if ((flags & VR_START_WRITE) != 0) {
1925 			MNT_REF(mp);
1926 			mp->mnt_writeopcount++;
1927 		}
1928 		MNT_IUNLOCK(mp);
1929 		if ((flags & VR_NO_SUSPCLR) == 0)
1930 			VFS_SUSP_CLEAN(mp);
1931 		vfs_op_exit(mp);
1932 	} else if ((flags & VR_START_WRITE) != 0) {
1933 		MNT_REF(mp);
1934 		vn_start_write_refed(mp, 0, true);
1935 	} else {
1936 		MNT_IUNLOCK(mp);
1937 	}
1938 }
1939 
1940 /*
1941  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1942  * methods.
1943  */
1944 int
1945 vfs_write_suspend_umnt(struct mount *mp)
1946 {
1947 	int error;
1948 
1949 	MPASS(vn_suspendable(mp));
1950 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1951 	    ("vfs_write_suspend_umnt: recursed"));
1952 
1953 	/* dounmount() already called vn_start_write(). */
1954 	for (;;) {
1955 		vn_finished_write(mp);
1956 		error = vfs_write_suspend(mp, 0);
1957 		if (error != 0) {
1958 			vn_start_write(NULL, &mp, V_WAIT);
1959 			return (error);
1960 		}
1961 		MNT_ILOCK(mp);
1962 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1963 			break;
1964 		MNT_IUNLOCK(mp);
1965 		vn_start_write(NULL, &mp, V_WAIT);
1966 	}
1967 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1968 	wakeup(&mp->mnt_flag);
1969 	MNT_IUNLOCK(mp);
1970 	curthread->td_pflags |= TDP_IGNSUSP;
1971 	return (0);
1972 }
1973 
1974 /*
1975  * Implement kqueues for files by translating it to vnode operation.
1976  */
1977 static int
1978 vn_kqfilter(struct file *fp, struct knote *kn)
1979 {
1980 
1981 	return (VOP_KQFILTER(fp->f_vnode, kn));
1982 }
1983 
1984 /*
1985  * Simplified in-kernel wrapper calls for extended attribute access.
1986  * Both calls pass in a NULL credential, authorizing as "kernel" access.
1987  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1988  */
1989 int
1990 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1991     const char *attrname, int *buflen, char *buf, struct thread *td)
1992 {
1993 	struct uio	auio;
1994 	struct iovec	iov;
1995 	int	error;
1996 
1997 	iov.iov_len = *buflen;
1998 	iov.iov_base = buf;
1999 
2000 	auio.uio_iov = &iov;
2001 	auio.uio_iovcnt = 1;
2002 	auio.uio_rw = UIO_READ;
2003 	auio.uio_segflg = UIO_SYSSPACE;
2004 	auio.uio_td = td;
2005 	auio.uio_offset = 0;
2006 	auio.uio_resid = *buflen;
2007 
2008 	if ((ioflg & IO_NODELOCKED) == 0)
2009 		vn_lock(vp, LK_SHARED | LK_RETRY);
2010 
2011 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2012 
2013 	/* authorize attribute retrieval as kernel */
2014 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2015 	    td);
2016 
2017 	if ((ioflg & IO_NODELOCKED) == 0)
2018 		VOP_UNLOCK(vp, 0);
2019 
2020 	if (error == 0) {
2021 		*buflen = *buflen - auio.uio_resid;
2022 	}
2023 
2024 	return (error);
2025 }
2026 
2027 /*
2028  * XXX failure mode if partially written?
2029  */
2030 int
2031 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2032     const char *attrname, int buflen, char *buf, struct thread *td)
2033 {
2034 	struct uio	auio;
2035 	struct iovec	iov;
2036 	struct mount	*mp;
2037 	int	error;
2038 
2039 	iov.iov_len = buflen;
2040 	iov.iov_base = buf;
2041 
2042 	auio.uio_iov = &iov;
2043 	auio.uio_iovcnt = 1;
2044 	auio.uio_rw = UIO_WRITE;
2045 	auio.uio_segflg = UIO_SYSSPACE;
2046 	auio.uio_td = td;
2047 	auio.uio_offset = 0;
2048 	auio.uio_resid = buflen;
2049 
2050 	if ((ioflg & IO_NODELOCKED) == 0) {
2051 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2052 			return (error);
2053 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2054 	}
2055 
2056 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2057 
2058 	/* authorize attribute setting as kernel */
2059 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2060 
2061 	if ((ioflg & IO_NODELOCKED) == 0) {
2062 		vn_finished_write(mp);
2063 		VOP_UNLOCK(vp, 0);
2064 	}
2065 
2066 	return (error);
2067 }
2068 
2069 int
2070 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2071     const char *attrname, struct thread *td)
2072 {
2073 	struct mount	*mp;
2074 	int	error;
2075 
2076 	if ((ioflg & IO_NODELOCKED) == 0) {
2077 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2078 			return (error);
2079 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2080 	}
2081 
2082 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2083 
2084 	/* authorize attribute removal as kernel */
2085 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2086 	if (error == EOPNOTSUPP)
2087 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2088 		    NULL, td);
2089 
2090 	if ((ioflg & IO_NODELOCKED) == 0) {
2091 		vn_finished_write(mp);
2092 		VOP_UNLOCK(vp, 0);
2093 	}
2094 
2095 	return (error);
2096 }
2097 
2098 static int
2099 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2100     struct vnode **rvp)
2101 {
2102 
2103 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2104 }
2105 
2106 int
2107 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2108 {
2109 
2110 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2111 	    lkflags, rvp));
2112 }
2113 
2114 int
2115 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2116     int lkflags, struct vnode **rvp)
2117 {
2118 	struct mount *mp;
2119 	int ltype, error;
2120 
2121 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2122 	mp = vp->v_mount;
2123 	ltype = VOP_ISLOCKED(vp);
2124 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2125 	    ("vn_vget_ino: vp not locked"));
2126 	error = vfs_busy(mp, MBF_NOWAIT);
2127 	if (error != 0) {
2128 		vfs_ref(mp);
2129 		VOP_UNLOCK(vp, 0);
2130 		error = vfs_busy(mp, 0);
2131 		vn_lock(vp, ltype | LK_RETRY);
2132 		vfs_rel(mp);
2133 		if (error != 0)
2134 			return (ENOENT);
2135 		if (VN_IS_DOOMED(vp)) {
2136 			vfs_unbusy(mp);
2137 			return (ENOENT);
2138 		}
2139 	}
2140 	VOP_UNLOCK(vp, 0);
2141 	error = alloc(mp, alloc_arg, lkflags, rvp);
2142 	vfs_unbusy(mp);
2143 	if (error != 0 || *rvp != vp)
2144 		vn_lock(vp, ltype | LK_RETRY);
2145 	if (VN_IS_DOOMED(vp)) {
2146 		if (error == 0) {
2147 			if (*rvp == vp)
2148 				vunref(vp);
2149 			else
2150 				vput(*rvp);
2151 		}
2152 		error = ENOENT;
2153 	}
2154 	return (error);
2155 }
2156 
2157 int
2158 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2159     struct thread *td)
2160 {
2161 
2162 	if (vp->v_type != VREG || td == NULL)
2163 		return (0);
2164 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2165 	    lim_cur(td, RLIMIT_FSIZE)) {
2166 		PROC_LOCK(td->td_proc);
2167 		kern_psignal(td->td_proc, SIGXFSZ);
2168 		PROC_UNLOCK(td->td_proc);
2169 		return (EFBIG);
2170 	}
2171 	return (0);
2172 }
2173 
2174 int
2175 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2176     struct thread *td)
2177 {
2178 	struct vnode *vp;
2179 
2180 	vp = fp->f_vnode;
2181 #ifdef AUDIT
2182 	vn_lock(vp, LK_SHARED | LK_RETRY);
2183 	AUDIT_ARG_VNODE1(vp);
2184 	VOP_UNLOCK(vp, 0);
2185 #endif
2186 	return (setfmode(td, active_cred, vp, mode));
2187 }
2188 
2189 int
2190 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2191     struct thread *td)
2192 {
2193 	struct vnode *vp;
2194 
2195 	vp = fp->f_vnode;
2196 #ifdef AUDIT
2197 	vn_lock(vp, LK_SHARED | LK_RETRY);
2198 	AUDIT_ARG_VNODE1(vp);
2199 	VOP_UNLOCK(vp, 0);
2200 #endif
2201 	return (setfown(td, active_cred, vp, uid, gid));
2202 }
2203 
2204 void
2205 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2206 {
2207 	vm_object_t object;
2208 
2209 	if ((object = vp->v_object) == NULL)
2210 		return;
2211 	VM_OBJECT_WLOCK(object);
2212 	vm_object_page_remove(object, start, end, 0);
2213 	VM_OBJECT_WUNLOCK(object);
2214 }
2215 
2216 int
2217 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2218 {
2219 	struct vattr va;
2220 	daddr_t bn, bnp;
2221 	uint64_t bsize;
2222 	off_t noff;
2223 	int error;
2224 
2225 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2226 	    ("Wrong command %lu", cmd));
2227 
2228 	if (vn_lock(vp, LK_SHARED) != 0)
2229 		return (EBADF);
2230 	if (vp->v_type != VREG) {
2231 		error = ENOTTY;
2232 		goto unlock;
2233 	}
2234 	error = VOP_GETATTR(vp, &va, cred);
2235 	if (error != 0)
2236 		goto unlock;
2237 	noff = *off;
2238 	if (noff >= va.va_size) {
2239 		error = ENXIO;
2240 		goto unlock;
2241 	}
2242 	bsize = vp->v_mount->mnt_stat.f_iosize;
2243 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2244 	    noff % bsize) {
2245 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2246 		if (error == EOPNOTSUPP) {
2247 			error = ENOTTY;
2248 			goto unlock;
2249 		}
2250 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2251 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2252 			noff = bn * bsize;
2253 			if (noff < *off)
2254 				noff = *off;
2255 			goto unlock;
2256 		}
2257 	}
2258 	if (noff > va.va_size)
2259 		noff = va.va_size;
2260 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2261 	if (cmd == FIOSEEKDATA)
2262 		error = ENXIO;
2263 unlock:
2264 	VOP_UNLOCK(vp, 0);
2265 	if (error == 0)
2266 		*off = noff;
2267 	return (error);
2268 }
2269 
2270 int
2271 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2272 {
2273 	struct ucred *cred;
2274 	struct vnode *vp;
2275 	struct vattr vattr;
2276 	off_t foffset, size;
2277 	int error, noneg;
2278 
2279 	cred = td->td_ucred;
2280 	vp = fp->f_vnode;
2281 	foffset = foffset_lock(fp, 0);
2282 	noneg = (vp->v_type != VCHR);
2283 	error = 0;
2284 	switch (whence) {
2285 	case L_INCR:
2286 		if (noneg &&
2287 		    (foffset < 0 ||
2288 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2289 			error = EOVERFLOW;
2290 			break;
2291 		}
2292 		offset += foffset;
2293 		break;
2294 	case L_XTND:
2295 		vn_lock(vp, LK_SHARED | LK_RETRY);
2296 		error = VOP_GETATTR(vp, &vattr, cred);
2297 		VOP_UNLOCK(vp, 0);
2298 		if (error)
2299 			break;
2300 
2301 		/*
2302 		 * If the file references a disk device, then fetch
2303 		 * the media size and use that to determine the ending
2304 		 * offset.
2305 		 */
2306 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2307 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2308 			vattr.va_size = size;
2309 		if (noneg &&
2310 		    (vattr.va_size > OFF_MAX ||
2311 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2312 			error = EOVERFLOW;
2313 			break;
2314 		}
2315 		offset += vattr.va_size;
2316 		break;
2317 	case L_SET:
2318 		break;
2319 	case SEEK_DATA:
2320 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2321 		if (error == ENOTTY)
2322 			error = EINVAL;
2323 		break;
2324 	case SEEK_HOLE:
2325 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2326 		if (error == ENOTTY)
2327 			error = EINVAL;
2328 		break;
2329 	default:
2330 		error = EINVAL;
2331 	}
2332 	if (error == 0 && noneg && offset < 0)
2333 		error = EINVAL;
2334 	if (error != 0)
2335 		goto drop;
2336 	VFS_KNOTE_UNLOCKED(vp, 0);
2337 	td->td_uretoff.tdu_off = offset;
2338 drop:
2339 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2340 	return (error);
2341 }
2342 
2343 int
2344 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2345     struct thread *td)
2346 {
2347 	int error;
2348 
2349 	/*
2350 	 * Grant permission if the caller is the owner of the file, or
2351 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2352 	 * on the file.  If the time pointer is null, then write
2353 	 * permission on the file is also sufficient.
2354 	 *
2355 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2356 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2357 	 * will be allowed to set the times [..] to the current
2358 	 * server time.
2359 	 */
2360 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2361 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2362 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2363 	return (error);
2364 }
2365 
2366 int
2367 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2368 {
2369 	struct vnode *vp;
2370 	int error;
2371 
2372 	if (fp->f_type == DTYPE_FIFO)
2373 		kif->kf_type = KF_TYPE_FIFO;
2374 	else
2375 		kif->kf_type = KF_TYPE_VNODE;
2376 	vp = fp->f_vnode;
2377 	vref(vp);
2378 	FILEDESC_SUNLOCK(fdp);
2379 	error = vn_fill_kinfo_vnode(vp, kif);
2380 	vrele(vp);
2381 	FILEDESC_SLOCK(fdp);
2382 	return (error);
2383 }
2384 
2385 static inline void
2386 vn_fill_junk(struct kinfo_file *kif)
2387 {
2388 	size_t len, olen;
2389 
2390 	/*
2391 	 * Simulate vn_fullpath returning changing values for a given
2392 	 * vp during e.g. coredump.
2393 	 */
2394 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2395 	olen = strlen(kif->kf_path);
2396 	if (len < olen)
2397 		strcpy(&kif->kf_path[len - 1], "$");
2398 	else
2399 		for (; olen < len; olen++)
2400 			strcpy(&kif->kf_path[olen], "A");
2401 }
2402 
2403 int
2404 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2405 {
2406 	struct vattr va;
2407 	char *fullpath, *freepath;
2408 	int error;
2409 
2410 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2411 	freepath = NULL;
2412 	fullpath = "-";
2413 	error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2414 	if (error == 0) {
2415 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2416 	}
2417 	if (freepath != NULL)
2418 		free(freepath, M_TEMP);
2419 
2420 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2421 		vn_fill_junk(kif);
2422 	);
2423 
2424 	/*
2425 	 * Retrieve vnode attributes.
2426 	 */
2427 	va.va_fsid = VNOVAL;
2428 	va.va_rdev = NODEV;
2429 	vn_lock(vp, LK_SHARED | LK_RETRY);
2430 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2431 	VOP_UNLOCK(vp, 0);
2432 	if (error != 0)
2433 		return (error);
2434 	if (va.va_fsid != VNOVAL)
2435 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2436 	else
2437 		kif->kf_un.kf_file.kf_file_fsid =
2438 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2439 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2440 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2441 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2442 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2443 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2444 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2445 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2446 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2447 	return (0);
2448 }
2449 
2450 int
2451 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2452     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2453     struct thread *td)
2454 {
2455 #ifdef HWPMC_HOOKS
2456 	struct pmckern_map_in pkm;
2457 #endif
2458 	struct mount *mp;
2459 	struct vnode *vp;
2460 	vm_object_t object;
2461 	vm_prot_t maxprot;
2462 	boolean_t writecounted;
2463 	int error;
2464 
2465 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2466     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2467 	/*
2468 	 * POSIX shared-memory objects are defined to have
2469 	 * kernel persistence, and are not defined to support
2470 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2471 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2472 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2473 	 * flag to request this behavior.
2474 	 */
2475 	if ((fp->f_flag & FPOSIXSHM) != 0)
2476 		flags |= MAP_NOSYNC;
2477 #endif
2478 	vp = fp->f_vnode;
2479 
2480 	/*
2481 	 * Ensure that file and memory protections are
2482 	 * compatible.  Note that we only worry about
2483 	 * writability if mapping is shared; in this case,
2484 	 * current and max prot are dictated by the open file.
2485 	 * XXX use the vnode instead?  Problem is: what
2486 	 * credentials do we use for determination? What if
2487 	 * proc does a setuid?
2488 	 */
2489 	mp = vp->v_mount;
2490 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2491 		maxprot = VM_PROT_NONE;
2492 		if ((prot & VM_PROT_EXECUTE) != 0)
2493 			return (EACCES);
2494 	} else
2495 		maxprot = VM_PROT_EXECUTE;
2496 	if ((fp->f_flag & FREAD) != 0)
2497 		maxprot |= VM_PROT_READ;
2498 	else if ((prot & VM_PROT_READ) != 0)
2499 		return (EACCES);
2500 
2501 	/*
2502 	 * If we are sharing potential changes via MAP_SHARED and we
2503 	 * are trying to get write permission although we opened it
2504 	 * without asking for it, bail out.
2505 	 */
2506 	if ((flags & MAP_SHARED) != 0) {
2507 		if ((fp->f_flag & FWRITE) != 0)
2508 			maxprot |= VM_PROT_WRITE;
2509 		else if ((prot & VM_PROT_WRITE) != 0)
2510 			return (EACCES);
2511 	} else {
2512 		maxprot |= VM_PROT_WRITE;
2513 		cap_maxprot |= VM_PROT_WRITE;
2514 	}
2515 	maxprot &= cap_maxprot;
2516 
2517 	/*
2518 	 * For regular files and shared memory, POSIX requires that
2519 	 * the value of foff be a legitimate offset within the data
2520 	 * object.  In particular, negative offsets are invalid.
2521 	 * Blocking negative offsets and overflows here avoids
2522 	 * possible wraparound or user-level access into reserved
2523 	 * ranges of the data object later.  In contrast, POSIX does
2524 	 * not dictate how offsets are used by device drivers, so in
2525 	 * the case of a device mapping a negative offset is passed
2526 	 * on.
2527 	 */
2528 	if (
2529 #ifdef _LP64
2530 	    size > OFF_MAX ||
2531 #endif
2532 	    foff < 0 || foff > OFF_MAX - size)
2533 		return (EINVAL);
2534 
2535 	writecounted = FALSE;
2536 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2537 	    &foff, &object, &writecounted);
2538 	if (error != 0)
2539 		return (error);
2540 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2541 	    foff, writecounted, td);
2542 	if (error != 0) {
2543 		/*
2544 		 * If this mapping was accounted for in the vnode's
2545 		 * writecount, then undo that now.
2546 		 */
2547 		if (writecounted)
2548 			vm_pager_release_writecount(object, 0, size);
2549 		vm_object_deallocate(object);
2550 	}
2551 #ifdef HWPMC_HOOKS
2552 	/* Inform hwpmc(4) if an executable is being mapped. */
2553 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2554 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2555 			pkm.pm_file = vp;
2556 			pkm.pm_address = (uintptr_t) *addr;
2557 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2558 		}
2559 	}
2560 #endif
2561 	return (error);
2562 }
2563 
2564 void
2565 vn_fsid(struct vnode *vp, struct vattr *va)
2566 {
2567 	fsid_t *f;
2568 
2569 	f = &vp->v_mount->mnt_stat.f_fsid;
2570 	va->va_fsid = (uint32_t)f->val[1];
2571 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2572 	va->va_fsid += (uint32_t)f->val[0];
2573 }
2574 
2575 int
2576 vn_fsync_buf(struct vnode *vp, int waitfor)
2577 {
2578 	struct buf *bp, *nbp;
2579 	struct bufobj *bo;
2580 	struct mount *mp;
2581 	int error, maxretry;
2582 
2583 	error = 0;
2584 	maxretry = 10000;     /* large, arbitrarily chosen */
2585 	mp = NULL;
2586 	if (vp->v_type == VCHR) {
2587 		VI_LOCK(vp);
2588 		mp = vp->v_rdev->si_mountpt;
2589 		VI_UNLOCK(vp);
2590 	}
2591 	bo = &vp->v_bufobj;
2592 	BO_LOCK(bo);
2593 loop1:
2594 	/*
2595 	 * MARK/SCAN initialization to avoid infinite loops.
2596 	 */
2597         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2598 		bp->b_vflags &= ~BV_SCANNED;
2599 		bp->b_error = 0;
2600 	}
2601 
2602 	/*
2603 	 * Flush all dirty buffers associated with a vnode.
2604 	 */
2605 loop2:
2606 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2607 		if ((bp->b_vflags & BV_SCANNED) != 0)
2608 			continue;
2609 		bp->b_vflags |= BV_SCANNED;
2610 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2611 			if (waitfor != MNT_WAIT)
2612 				continue;
2613 			if (BUF_LOCK(bp,
2614 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2615 			    BO_LOCKPTR(bo)) != 0) {
2616 				BO_LOCK(bo);
2617 				goto loop1;
2618 			}
2619 			BO_LOCK(bo);
2620 		}
2621 		BO_UNLOCK(bo);
2622 		KASSERT(bp->b_bufobj == bo,
2623 		    ("bp %p wrong b_bufobj %p should be %p",
2624 		    bp, bp->b_bufobj, bo));
2625 		if ((bp->b_flags & B_DELWRI) == 0)
2626 			panic("fsync: not dirty");
2627 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2628 			vfs_bio_awrite(bp);
2629 		} else {
2630 			bremfree(bp);
2631 			bawrite(bp);
2632 		}
2633 		if (maxretry < 1000)
2634 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2635 		BO_LOCK(bo);
2636 		goto loop2;
2637 	}
2638 
2639 	/*
2640 	 * If synchronous the caller expects us to completely resolve all
2641 	 * dirty buffers in the system.  Wait for in-progress I/O to
2642 	 * complete (which could include background bitmap writes), then
2643 	 * retry if dirty blocks still exist.
2644 	 */
2645 	if (waitfor == MNT_WAIT) {
2646 		bufobj_wwait(bo, 0, 0);
2647 		if (bo->bo_dirty.bv_cnt > 0) {
2648 			/*
2649 			 * If we are unable to write any of these buffers
2650 			 * then we fail now rather than trying endlessly
2651 			 * to write them out.
2652 			 */
2653 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2654 				if ((error = bp->b_error) != 0)
2655 					break;
2656 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2657 			    (error == 0 && --maxretry >= 0))
2658 				goto loop1;
2659 			if (error == 0)
2660 				error = EAGAIN;
2661 		}
2662 	}
2663 	BO_UNLOCK(bo);
2664 	if (error != 0)
2665 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2666 
2667 	return (error);
2668 }
2669 
2670 /*
2671  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2672  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2673  * to do the actual copy.
2674  * vn_generic_copy_file_range() is factored out, so it can be called
2675  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2676  * different file systems.
2677  */
2678 int
2679 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2680     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2681     struct ucred *outcred, struct thread *fsize_td)
2682 {
2683 	int error;
2684 	size_t len;
2685 	uint64_t uvalin, uvalout;
2686 
2687 	len = *lenp;
2688 	*lenp = 0;		/* For error returns. */
2689 	error = 0;
2690 
2691 	/* Do some sanity checks on the arguments. */
2692 	uvalin = *inoffp;
2693 	uvalin += len;
2694 	uvalout = *outoffp;
2695 	uvalout += len;
2696 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2697 		error = EISDIR;
2698 	else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin <
2699 	    (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX ||
2700 	    uvalout < (uint64_t)*outoffp || invp->v_type != VREG ||
2701 	    outvp->v_type != VREG)
2702 		error = EINVAL;
2703 	if (error != 0)
2704 		goto out;
2705 
2706 	/*
2707 	 * If the two vnode are for the same file system, call
2708 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2709 	 * which can handle copies across multiple file systems.
2710 	 */
2711 	*lenp = len;
2712 	if (invp->v_mount == outvp->v_mount)
2713 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2714 		    lenp, flags, incred, outcred, fsize_td);
2715 	else
2716 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2717 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2718 out:
2719 	return (error);
2720 }
2721 
2722 /*
2723  * Test len bytes of data starting at dat for all bytes == 0.
2724  * Return true if all bytes are zero, false otherwise.
2725  * Expects dat to be well aligned.
2726  */
2727 static bool
2728 mem_iszero(void *dat, int len)
2729 {
2730 	int i;
2731 	const u_int *p;
2732 	const char *cp;
2733 
2734 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2735 		if (len >= sizeof(*p)) {
2736 			if (*p != 0)
2737 				return (false);
2738 		} else {
2739 			cp = (const char *)p;
2740 			for (i = 0; i < len; i++, cp++)
2741 				if (*cp != '\0')
2742 					return (false);
2743 		}
2744 	}
2745 	return (true);
2746 }
2747 
2748 /*
2749  * Look for a hole in the output file and, if found, adjust *outoffp
2750  * and *xferp to skip past the hole.
2751  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2752  * to be written as 0's upon return.
2753  */
2754 static off_t
2755 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2756     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2757 {
2758 	int error;
2759 	off_t delta;
2760 
2761 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2762 		*dataoffp = *outoffp;
2763 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2764 		    curthread);
2765 		if (error == 0) {
2766 			*holeoffp = *dataoffp;
2767 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2768 			    curthread);
2769 		}
2770 		if (error != 0 || *holeoffp == *dataoffp) {
2771 			/*
2772 			 * Since outvp is unlocked, it may be possible for
2773 			 * another thread to do a truncate(), lseek(), write()
2774 			 * creating a hole at startoff between the above
2775 			 * VOP_IOCTL() calls, if the other thread does not do
2776 			 * rangelocking.
2777 			 * If that happens, *holeoffp == *dataoffp and finding
2778 			 * the hole has failed, so disable vn_skip_hole().
2779 			 */
2780 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2781 			return (xfer2);
2782 		}
2783 		KASSERT(*dataoffp >= *outoffp,
2784 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2785 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2786 		KASSERT(*holeoffp > *dataoffp,
2787 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2788 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2789 	}
2790 
2791 	/*
2792 	 * If there is a hole before the data starts, advance *outoffp and
2793 	 * *xferp past the hole.
2794 	 */
2795 	if (*dataoffp > *outoffp) {
2796 		delta = *dataoffp - *outoffp;
2797 		if (delta >= *xferp) {
2798 			/* Entire *xferp is a hole. */
2799 			*outoffp += *xferp;
2800 			*xferp = 0;
2801 			return (0);
2802 		}
2803 		*xferp -= delta;
2804 		*outoffp += delta;
2805 		xfer2 = MIN(xfer2, *xferp);
2806 	}
2807 
2808 	/*
2809 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2810 	 * that the write ends at the start of the hole.
2811 	 * *holeoffp should always be greater than *outoffp, but for the
2812 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2813 	 * value.
2814 	 */
2815 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2816 		xfer2 = *holeoffp - *outoffp;
2817 	return (xfer2);
2818 }
2819 
2820 /*
2821  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2822  * dat is a maximum of blksize in length and can be written repeatedly in
2823  * the chunk.
2824  * If growfile == true, just grow the file via vn_truncate_locked() instead
2825  * of doing actual writes.
2826  * If checkhole == true, a hole is being punched, so skip over any hole
2827  * already in the output file.
2828  */
2829 static int
2830 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2831     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2832 {
2833 	struct mount *mp;
2834 	off_t dataoff, holeoff, xfer2;
2835 	int error, lckf;
2836 
2837 	/*
2838 	 * Loop around doing writes of blksize until write has been completed.
2839 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2840 	 * done for each iteration, since the xfer argument can be very
2841 	 * large if there is a large hole to punch in the output file.
2842 	 */
2843 	error = 0;
2844 	holeoff = 0;
2845 	do {
2846 		xfer2 = MIN(xfer, blksize);
2847 		if (checkhole) {
2848 			/*
2849 			 * Punching a hole.  Skip writing if there is
2850 			 * already a hole in the output file.
2851 			 */
2852 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2853 			    &dataoff, &holeoff, cred);
2854 			if (xfer == 0)
2855 				break;
2856 			if (holeoff < 0)
2857 				checkhole = false;
2858 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2859 			    (intmax_t)xfer2));
2860 		}
2861 		bwillwrite();
2862 		mp = NULL;
2863 		error = vn_start_write(outvp, &mp, V_WAIT);
2864 		if (error == 0) {
2865 			if (MNT_SHARED_WRITES(mp))
2866 				lckf = LK_SHARED;
2867 			else
2868 				lckf = LK_EXCLUSIVE;
2869 			error = vn_lock(outvp, lckf);
2870 		}
2871 		if (error == 0) {
2872 			if (growfile)
2873 				error = vn_truncate_locked(outvp, outoff + xfer,
2874 				    false, cred);
2875 			else {
2876 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
2877 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
2878 				    curthread->td_ucred, cred, NULL, curthread);
2879 				outoff += xfer2;
2880 				xfer -= xfer2;
2881 			}
2882 			VOP_UNLOCK(outvp, 0);
2883 		}
2884 		if (mp != NULL)
2885 			vn_finished_write(mp);
2886 	} while (!growfile && xfer > 0 && error == 0);
2887 	return (error);
2888 }
2889 
2890 /*
2891  * Copy a byte range of one file to another.  This function can handle the
2892  * case where invp and outvp are on different file systems.
2893  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
2894  * is no better file system specific way to do it.
2895  */
2896 int
2897 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
2898     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
2899     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
2900 {
2901 	struct vattr va;
2902 	struct mount *mp;
2903 	struct uio io;
2904 	off_t startoff, endoff, xfer, xfer2;
2905 	u_long blksize;
2906 	int error;
2907 	bool cantseek, readzeros, eof, lastblock;
2908 	ssize_t aresid;
2909 	size_t copylen, len, savlen;
2910 	char *dat;
2911 	long holein, holeout;
2912 
2913 	holein = holeout = 0;
2914 	savlen = len = *lenp;
2915 	error = 0;
2916 	dat = NULL;
2917 
2918 	error = vn_lock(invp, LK_SHARED);
2919 	if (error != 0)
2920 		goto out;
2921 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
2922 		holein = 0;
2923 	VOP_UNLOCK(invp, 0);
2924 
2925 	mp = NULL;
2926 	error = vn_start_write(outvp, &mp, V_WAIT);
2927 	if (error == 0)
2928 		error = vn_lock(outvp, LK_EXCLUSIVE);
2929 	if (error == 0) {
2930 		/*
2931 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
2932 		 * now that outvp is locked.
2933 		 */
2934 		if (fsize_td != NULL) {
2935 			io.uio_offset = *outoffp;
2936 			io.uio_resid = len;
2937 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
2938 			if (error != 0)
2939 				error = EFBIG;
2940 		}
2941 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
2942 			holeout = 0;
2943 		/*
2944 		 * Holes that are past EOF do not need to be written as a block
2945 		 * of zero bytes.  So, truncate the output file as far as
2946 		 * possible and then use va.va_size to decide if writing 0
2947 		 * bytes is necessary in the loop below.
2948 		 */
2949 		if (error == 0)
2950 			error = VOP_GETATTR(outvp, &va, outcred);
2951 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
2952 		    *outoffp + len) {
2953 #ifdef MAC
2954 			error = mac_vnode_check_write(curthread->td_ucred,
2955 			    outcred, outvp);
2956 			if (error == 0)
2957 #endif
2958 				error = vn_truncate_locked(outvp, *outoffp,
2959 				    false, outcred);
2960 			if (error == 0)
2961 				va.va_size = *outoffp;
2962 		}
2963 		VOP_UNLOCK(outvp, 0);
2964 	}
2965 	if (mp != NULL)
2966 		vn_finished_write(mp);
2967 	if (error != 0)
2968 		goto out;
2969 
2970 	/*
2971 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
2972 	 * If hole sizes aren't available, set the blksize to the larger
2973 	 * f_iosize of invp and outvp.
2974 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
2975 	 * This value is clipped at 4Kbytes and 1Mbyte.
2976 	 */
2977 	blksize = MAX(holein, holeout);
2978 	if (blksize == 0)
2979 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
2980 		    outvp->v_mount->mnt_stat.f_iosize);
2981 	if (blksize < 4096)
2982 		blksize = 4096;
2983 	else if (blksize > 1024 * 1024)
2984 		blksize = 1024 * 1024;
2985 	dat = malloc(blksize, M_TEMP, M_WAITOK);
2986 
2987 	/*
2988 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
2989 	 * to find holes.  Otherwise, just scan the read block for all 0s
2990 	 * in the inner loop where the data copying is done.
2991 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
2992 	 * support holes on the server, but do not support FIOSEEKHOLE.
2993 	 */
2994 	eof = false;
2995 	while (len > 0 && error == 0 && !eof) {
2996 		endoff = 0;			/* To shut up compilers. */
2997 		cantseek = true;
2998 		startoff = *inoffp;
2999 		copylen = len;
3000 
3001 		/*
3002 		 * Find the next data area.  If there is just a hole to EOF,
3003 		 * FIOSEEKDATA should fail and then we drop down into the
3004 		 * inner loop and create the hole on the outvp file.
3005 		 * (I do not know if any file system will report a hole to
3006 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3007 		 *  will fail for those file systems.)
3008 		 *
3009 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3010 		 * the code just falls through to the inner copy loop.
3011 		 */
3012 		error = EINVAL;
3013 		if (holein > 0)
3014 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3015 			    incred, curthread);
3016 		if (error == 0) {
3017 			endoff = startoff;
3018 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3019 			    incred, curthread);
3020 			/*
3021 			 * Since invp is unlocked, it may be possible for
3022 			 * another thread to do a truncate(), lseek(), write()
3023 			 * creating a hole at startoff between the above
3024 			 * VOP_IOCTL() calls, if the other thread does not do
3025 			 * rangelocking.
3026 			 * If that happens, startoff == endoff and finding
3027 			 * the hole has failed, so set an error.
3028 			 */
3029 			if (error == 0 && startoff == endoff)
3030 				error = EINVAL; /* Any error. Reset to 0. */
3031 		}
3032 		if (error == 0) {
3033 			if (startoff > *inoffp) {
3034 				/* Found hole before data block. */
3035 				xfer = MIN(startoff - *inoffp, len);
3036 				if (*outoffp < va.va_size) {
3037 					/* Must write 0s to punch hole. */
3038 					xfer2 = MIN(va.va_size - *outoffp,
3039 					    xfer);
3040 					memset(dat, 0, MIN(xfer2, blksize));
3041 					error = vn_write_outvp(outvp, dat,
3042 					    *outoffp, xfer2, blksize, false,
3043 					    holeout > 0, outcred);
3044 				}
3045 
3046 				if (error == 0 && *outoffp + xfer >
3047 				    va.va_size && xfer == len)
3048 					/* Grow last block. */
3049 					error = vn_write_outvp(outvp, dat,
3050 					    *outoffp, xfer, blksize, true,
3051 					    false, outcred);
3052 				if (error == 0) {
3053 					*inoffp += xfer;
3054 					*outoffp += xfer;
3055 					len -= xfer;
3056 				}
3057 			}
3058 			copylen = MIN(len, endoff - startoff);
3059 			cantseek = false;
3060 		} else {
3061 			cantseek = true;
3062 			startoff = *inoffp;
3063 			copylen = len;
3064 			error = 0;
3065 		}
3066 
3067 		xfer = blksize;
3068 		if (cantseek) {
3069 			/*
3070 			 * Set first xfer to end at a block boundary, so that
3071 			 * holes are more likely detected in the loop below via
3072 			 * the for all bytes 0 method.
3073 			 */
3074 			xfer -= (*inoffp % blksize);
3075 		}
3076 		/* Loop copying the data block. */
3077 		while (copylen > 0 && error == 0 && !eof) {
3078 			if (copylen < xfer)
3079 				xfer = copylen;
3080 			error = vn_lock(invp, LK_SHARED);
3081 			if (error != 0)
3082 				goto out;
3083 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3084 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3085 			    curthread->td_ucred, incred, &aresid,
3086 			    curthread);
3087 			VOP_UNLOCK(invp, 0);
3088 			lastblock = false;
3089 			if (error == 0 && aresid > 0) {
3090 				/* Stop the copy at EOF on the input file. */
3091 				xfer -= aresid;
3092 				eof = true;
3093 				lastblock = true;
3094 			}
3095 			if (error == 0) {
3096 				/*
3097 				 * Skip the write for holes past the initial EOF
3098 				 * of the output file, unless this is the last
3099 				 * write of the output file at EOF.
3100 				 */
3101 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3102 				    false;
3103 				if (xfer == len)
3104 					lastblock = true;
3105 				if (!cantseek || *outoffp < va.va_size ||
3106 				    lastblock || !readzeros)
3107 					error = vn_write_outvp(outvp, dat,
3108 					    *outoffp, xfer, blksize,
3109 					    readzeros && lastblock &&
3110 					    *outoffp >= va.va_size, false,
3111 					    outcred);
3112 				if (error == 0) {
3113 					*inoffp += xfer;
3114 					startoff += xfer;
3115 					*outoffp += xfer;
3116 					copylen -= xfer;
3117 					len -= xfer;
3118 				}
3119 			}
3120 			xfer = blksize;
3121 		}
3122 	}
3123 out:
3124 	*lenp = savlen - len;
3125 	free(dat, M_TEMP);
3126 	return (error);
3127 }
3128