xref: /freebsd/sys/kern/vfs_vnops.c (revision a0409676120c1e558d0ade943019934e0f15118d)
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/prng.h>
74 #include <sys/sx.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
78 #include <sys/conf.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
81 #include <sys/user.h>
82 
83 #include <security/audit/audit.h>
84 #include <security/mac/mac_framework.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_extern.h>
88 #include <vm/pmap.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pager.h>
93 
94 #ifdef HWPMC_HOOKS
95 #include <sys/pmckern.h>
96 #endif
97 
98 static fo_rdwr_t	vn_read;
99 static fo_rdwr_t	vn_write;
100 static fo_rdwr_t	vn_io_fault;
101 static fo_truncate_t	vn_truncate;
102 static fo_ioctl_t	vn_ioctl;
103 static fo_poll_t	vn_poll;
104 static fo_kqfilter_t	vn_kqfilter;
105 static fo_stat_t	vn_statfile;
106 static fo_close_t	vn_closefile;
107 static fo_mmap_t	vn_mmap;
108 static fo_fallocate_t	vn_fallocate;
109 
110 struct 	fileops vnops = {
111 	.fo_read = vn_io_fault,
112 	.fo_write = vn_io_fault,
113 	.fo_truncate = vn_truncate,
114 	.fo_ioctl = vn_ioctl,
115 	.fo_poll = vn_poll,
116 	.fo_kqfilter = vn_kqfilter,
117 	.fo_stat = vn_statfile,
118 	.fo_close = vn_closefile,
119 	.fo_chmod = vn_chmod,
120 	.fo_chown = vn_chown,
121 	.fo_sendfile = vn_sendfile,
122 	.fo_seek = vn_seek,
123 	.fo_fill_kinfo = vn_fill_kinfo,
124 	.fo_mmap = vn_mmap,
125 	.fo_fallocate = vn_fallocate,
126 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
127 };
128 
129 const u_int io_hold_cnt = 16;
130 static int vn_io_fault_enable = 1;
131 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
132     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
133 static int vn_io_fault_prefault = 0;
134 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
135     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
136 static int vn_io_pgcache_read_enable = 1;
137 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
138     &vn_io_pgcache_read_enable, 0,
139     "Enable copying from page cache for reads, avoiding fs");
140 static u_long vn_io_faults_cnt;
141 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
142     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
143 
144 static int vfs_allow_read_dir = 0;
145 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
146     &vfs_allow_read_dir, 0,
147     "Enable read(2) of directory by root for filesystems that support it");
148 
149 /*
150  * Returns true if vn_io_fault mode of handling the i/o request should
151  * be used.
152  */
153 static bool
154 do_vn_io_fault(struct vnode *vp, struct uio *uio)
155 {
156 	struct mount *mp;
157 
158 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
159 	    (mp = vp->v_mount) != NULL &&
160 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
161 }
162 
163 /*
164  * Structure used to pass arguments to vn_io_fault1(), to do either
165  * file- or vnode-based I/O calls.
166  */
167 struct vn_io_fault_args {
168 	enum {
169 		VN_IO_FAULT_FOP,
170 		VN_IO_FAULT_VOP
171 	} kind;
172 	struct ucred *cred;
173 	int flags;
174 	union {
175 		struct fop_args_tag {
176 			struct file *fp;
177 			fo_rdwr_t *doio;
178 		} fop_args;
179 		struct vop_args_tag {
180 			struct vnode *vp;
181 		} vop_args;
182 	} args;
183 };
184 
185 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
186     struct vn_io_fault_args *args, struct thread *td);
187 
188 int
189 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
190 {
191 	struct thread *td = ndp->ni_cnd.cn_thread;
192 
193 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
194 }
195 
196 static uint64_t
197 open2nameif(int fmode, u_int vn_open_flags)
198 {
199 	uint64_t res;
200 
201 	res = ISOPEN | LOCKLEAF;
202 	if ((fmode & O_RESOLVE_BENEATH) != 0)
203 		res |= RBENEATH;
204 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
205 		res |= AUDITVNODE1;
206 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
207 		res |= NOCAPCHECK;
208 	return (res);
209 }
210 
211 /*
212  * Common code for vnode open operations via a name lookup.
213  * Lookup the vnode and invoke VOP_CREATE if needed.
214  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
215  *
216  * Note that this does NOT free nameidata for the successful case,
217  * due to the NDINIT being done elsewhere.
218  */
219 int
220 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
221     struct ucred *cred, struct file *fp)
222 {
223 	struct vnode *vp;
224 	struct mount *mp;
225 	struct thread *td = ndp->ni_cnd.cn_thread;
226 	struct vattr vat;
227 	struct vattr *vap = &vat;
228 	int fmode, error;
229 	bool first_open;
230 
231 restart:
232 	first_open = false;
233 	fmode = *flagp;
234 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
235 	    O_EXCL | O_DIRECTORY))
236 		return (EINVAL);
237 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
238 		ndp->ni_cnd.cn_nameiop = CREATE;
239 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
240 		/*
241 		 * Set NOCACHE to avoid flushing the cache when
242 		 * rolling in many files at once.
243 		 *
244 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
245 		 * exist despite NOCACHE.
246 		 */
247 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
248 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
249 			ndp->ni_cnd.cn_flags |= FOLLOW;
250 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
251 			bwillwrite();
252 		if ((error = namei(ndp)) != 0)
253 			return (error);
254 		if (ndp->ni_vp == NULL) {
255 			VATTR_NULL(vap);
256 			vap->va_type = VREG;
257 			vap->va_mode = cmode;
258 			if (fmode & O_EXCL)
259 				vap->va_vaflags |= VA_EXCLUSIVE;
260 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
261 				NDFREE(ndp, NDF_ONLY_PNBUF);
262 				vput(ndp->ni_dvp);
263 				if ((error = vn_start_write(NULL, &mp,
264 				    V_XSLEEP | PCATCH)) != 0)
265 					return (error);
266 				NDREINIT(ndp);
267 				goto restart;
268 			}
269 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
270 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
271 #ifdef MAC
272 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
273 			    &ndp->ni_cnd, vap);
274 			if (error == 0)
275 #endif
276 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
277 				    &ndp->ni_cnd, vap);
278 			vp = ndp->ni_vp;
279 			if (error == 0 && (fmode & O_EXCL) != 0 &&
280 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
281 				VI_LOCK(vp);
282 				vp->v_iflag |= VI_FOPENING;
283 				VI_UNLOCK(vp);
284 				first_open = true;
285 			}
286 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
287 			    false);
288 			vn_finished_write(mp);
289 			if (error) {
290 				NDFREE(ndp, NDF_ONLY_PNBUF);
291 				if (error == ERELOOKUP) {
292 					NDREINIT(ndp);
293 					goto restart;
294 				}
295 				return (error);
296 			}
297 			fmode &= ~O_TRUNC;
298 		} else {
299 			if (ndp->ni_dvp == ndp->ni_vp)
300 				vrele(ndp->ni_dvp);
301 			else
302 				vput(ndp->ni_dvp);
303 			ndp->ni_dvp = NULL;
304 			vp = ndp->ni_vp;
305 			if (fmode & O_EXCL) {
306 				error = EEXIST;
307 				goto bad;
308 			}
309 			if (vp->v_type == VDIR) {
310 				error = EISDIR;
311 				goto bad;
312 			}
313 			fmode &= ~O_CREAT;
314 		}
315 	} else {
316 		ndp->ni_cnd.cn_nameiop = LOOKUP;
317 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
318 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
319 		    FOLLOW;
320 		if ((fmode & FWRITE) == 0)
321 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
322 		if ((error = namei(ndp)) != 0)
323 			return (error);
324 		vp = ndp->ni_vp;
325 	}
326 	error = vn_open_vnode(vp, fmode, cred, td, fp);
327 	if (first_open) {
328 		VI_LOCK(vp);
329 		vp->v_iflag &= ~VI_FOPENING;
330 		wakeup(vp);
331 		VI_UNLOCK(vp);
332 	}
333 	if (error)
334 		goto bad;
335 	*flagp = fmode;
336 	return (0);
337 bad:
338 	NDFREE(ndp, NDF_ONLY_PNBUF);
339 	vput(vp);
340 	*flagp = fmode;
341 	ndp->ni_vp = NULL;
342 	return (error);
343 }
344 
345 static int
346 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
347 {
348 	struct flock lf;
349 	int error, lock_flags, type;
350 
351 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
352 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
353 		return (0);
354 	KASSERT(fp != NULL, ("open with flock requires fp"));
355 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
356 		return (EOPNOTSUPP);
357 
358 	lock_flags = VOP_ISLOCKED(vp);
359 	VOP_UNLOCK(vp);
360 
361 	lf.l_whence = SEEK_SET;
362 	lf.l_start = 0;
363 	lf.l_len = 0;
364 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
365 	type = F_FLOCK;
366 	if ((fmode & FNONBLOCK) == 0)
367 		type |= F_WAIT;
368 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
369 		type |= F_FIRSTOPEN;
370 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
371 	if (error == 0)
372 		fp->f_flag |= FHASLOCK;
373 
374 	vn_lock(vp, lock_flags | LK_RETRY);
375 	return (error);
376 }
377 
378 /*
379  * Common code for vnode open operations once a vnode is located.
380  * Check permissions, and call the VOP_OPEN routine.
381  */
382 int
383 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
384     struct thread *td, struct file *fp)
385 {
386 	accmode_t accmode;
387 	int error;
388 
389 	if (vp->v_type == VLNK)
390 		return (EMLINK);
391 	if (vp->v_type == VSOCK)
392 		return (EOPNOTSUPP);
393 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
394 		return (ENOTDIR);
395 	accmode = 0;
396 	if (fmode & (FWRITE | O_TRUNC)) {
397 		if (vp->v_type == VDIR)
398 			return (EISDIR);
399 		accmode |= VWRITE;
400 	}
401 	if (fmode & FREAD)
402 		accmode |= VREAD;
403 	if (fmode & FEXEC)
404 		accmode |= VEXEC;
405 	if ((fmode & O_APPEND) && (fmode & FWRITE))
406 		accmode |= VAPPEND;
407 #ifdef MAC
408 	if (fmode & O_CREAT)
409 		accmode |= VCREAT;
410 	if (fmode & O_VERIFY)
411 		accmode |= VVERIFY;
412 	error = mac_vnode_check_open(cred, vp, accmode);
413 	if (error)
414 		return (error);
415 
416 	accmode &= ~(VCREAT | VVERIFY);
417 #endif
418 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
419 		error = VOP_ACCESS(vp, accmode, cred, td);
420 		if (error != 0)
421 			return (error);
422 	}
423 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
424 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
425 	error = VOP_OPEN(vp, fmode, cred, td, fp);
426 	if (error != 0)
427 		return (error);
428 
429 	error = vn_open_vnode_advlock(vp, fmode, fp);
430 	if (error == 0 && (fmode & FWRITE) != 0) {
431 		error = VOP_ADD_WRITECOUNT(vp, 1);
432 		if (error == 0) {
433 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
434 			     __func__, vp, vp->v_writecount);
435 		}
436 	}
437 
438 	/*
439 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
440 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
441 	 * Arrange for that by having fdrop() to use vn_closefile().
442 	 */
443 	if (error != 0) {
444 		fp->f_flag |= FOPENFAILED;
445 		fp->f_vnode = vp;
446 		if (fp->f_ops == &badfileops) {
447 			fp->f_type = DTYPE_VNODE;
448 			fp->f_ops = &vnops;
449 		}
450 		vref(vp);
451 	}
452 
453 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
454 	return (error);
455 
456 }
457 
458 /*
459  * Check for write permissions on the specified vnode.
460  * Prototype text segments cannot be written.
461  * It is racy.
462  */
463 int
464 vn_writechk(struct vnode *vp)
465 {
466 
467 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
468 	/*
469 	 * If there's shared text associated with
470 	 * the vnode, try to free it up once.  If
471 	 * we fail, we can't allow writing.
472 	 */
473 	if (VOP_IS_TEXT(vp))
474 		return (ETXTBSY);
475 
476 	return (0);
477 }
478 
479 /*
480  * Vnode close call
481  */
482 static int
483 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
484     struct thread *td, bool keep_ref)
485 {
486 	struct mount *mp;
487 	int error, lock_flags;
488 
489 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
490 	    MNT_EXTENDED_SHARED(vp->v_mount))
491 		lock_flags = LK_SHARED;
492 	else
493 		lock_flags = LK_EXCLUSIVE;
494 
495 	vn_start_write(vp, &mp, V_WAIT);
496 	vn_lock(vp, lock_flags | LK_RETRY);
497 	AUDIT_ARG_VNODE1(vp);
498 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
499 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
500 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
501 		    __func__, vp, vp->v_writecount);
502 	}
503 	error = VOP_CLOSE(vp, flags, file_cred, td);
504 	if (keep_ref)
505 		VOP_UNLOCK(vp);
506 	else
507 		vput(vp);
508 	vn_finished_write(mp);
509 	return (error);
510 }
511 
512 int
513 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
514     struct thread *td)
515 {
516 
517 	return (vn_close1(vp, flags, file_cred, td, false));
518 }
519 
520 /*
521  * Heuristic to detect sequential operation.
522  */
523 static int
524 sequential_heuristic(struct uio *uio, struct file *fp)
525 {
526 	enum uio_rw rw;
527 
528 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
529 
530 	rw = uio->uio_rw;
531 	if (fp->f_flag & FRDAHEAD)
532 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
533 
534 	/*
535 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
536 	 * that the first I/O is normally considered to be slightly
537 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
538 	 * unless previous seeks have reduced f_seqcount to 0, in which
539 	 * case offset 0 is not special.
540 	 */
541 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
542 	    uio->uio_offset == fp->f_nextoff[rw]) {
543 		/*
544 		 * f_seqcount is in units of fixed-size blocks so that it
545 		 * depends mainly on the amount of sequential I/O and not
546 		 * much on the number of sequential I/O's.  The fixed size
547 		 * of 16384 is hard-coded here since it is (not quite) just
548 		 * a magic size that works well here.  This size is more
549 		 * closely related to the best I/O size for real disks than
550 		 * to any block size used by software.
551 		 */
552 		if (uio->uio_resid >= IO_SEQMAX * 16384)
553 			fp->f_seqcount[rw] = IO_SEQMAX;
554 		else {
555 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
556 			if (fp->f_seqcount[rw] > IO_SEQMAX)
557 				fp->f_seqcount[rw] = IO_SEQMAX;
558 		}
559 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
560 	}
561 
562 	/* Not sequential.  Quickly draw-down sequentiality. */
563 	if (fp->f_seqcount[rw] > 1)
564 		fp->f_seqcount[rw] = 1;
565 	else
566 		fp->f_seqcount[rw] = 0;
567 	return (0);
568 }
569 
570 /*
571  * Package up an I/O request on a vnode into a uio and do it.
572  */
573 int
574 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
575     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
576     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
577 {
578 	struct uio auio;
579 	struct iovec aiov;
580 	struct mount *mp;
581 	struct ucred *cred;
582 	void *rl_cookie;
583 	struct vn_io_fault_args args;
584 	int error, lock_flags;
585 
586 	if (offset < 0 && vp->v_type != VCHR)
587 		return (EINVAL);
588 	auio.uio_iov = &aiov;
589 	auio.uio_iovcnt = 1;
590 	aiov.iov_base = base;
591 	aiov.iov_len = len;
592 	auio.uio_resid = len;
593 	auio.uio_offset = offset;
594 	auio.uio_segflg = segflg;
595 	auio.uio_rw = rw;
596 	auio.uio_td = td;
597 	error = 0;
598 
599 	if ((ioflg & IO_NODELOCKED) == 0) {
600 		if ((ioflg & IO_RANGELOCKED) == 0) {
601 			if (rw == UIO_READ) {
602 				rl_cookie = vn_rangelock_rlock(vp, offset,
603 				    offset + len);
604 			} else if ((ioflg & IO_APPEND) != 0) {
605 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
606 			} else {
607 				rl_cookie = vn_rangelock_wlock(vp, offset,
608 				    offset + len);
609 			}
610 		} else
611 			rl_cookie = NULL;
612 		mp = NULL;
613 		if (rw == UIO_WRITE) {
614 			if (vp->v_type != VCHR &&
615 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
616 			    != 0)
617 				goto out;
618 			if (MNT_SHARED_WRITES(mp) ||
619 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
620 				lock_flags = LK_SHARED;
621 			else
622 				lock_flags = LK_EXCLUSIVE;
623 		} else
624 			lock_flags = LK_SHARED;
625 		vn_lock(vp, lock_flags | LK_RETRY);
626 	} else
627 		rl_cookie = NULL;
628 
629 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
630 #ifdef MAC
631 	if ((ioflg & IO_NOMACCHECK) == 0) {
632 		if (rw == UIO_READ)
633 			error = mac_vnode_check_read(active_cred, file_cred,
634 			    vp);
635 		else
636 			error = mac_vnode_check_write(active_cred, file_cred,
637 			    vp);
638 	}
639 #endif
640 	if (error == 0) {
641 		if (file_cred != NULL)
642 			cred = file_cred;
643 		else
644 			cred = active_cred;
645 		if (do_vn_io_fault(vp, &auio)) {
646 			args.kind = VN_IO_FAULT_VOP;
647 			args.cred = cred;
648 			args.flags = ioflg;
649 			args.args.vop_args.vp = vp;
650 			error = vn_io_fault1(vp, &auio, &args, td);
651 		} else if (rw == UIO_READ) {
652 			error = VOP_READ(vp, &auio, ioflg, cred);
653 		} else /* if (rw == UIO_WRITE) */ {
654 			error = VOP_WRITE(vp, &auio, ioflg, cred);
655 		}
656 	}
657 	if (aresid)
658 		*aresid = auio.uio_resid;
659 	else
660 		if (auio.uio_resid && error == 0)
661 			error = EIO;
662 	if ((ioflg & IO_NODELOCKED) == 0) {
663 		VOP_UNLOCK(vp);
664 		if (mp != NULL)
665 			vn_finished_write(mp);
666 	}
667  out:
668 	if (rl_cookie != NULL)
669 		vn_rangelock_unlock(vp, rl_cookie);
670 	return (error);
671 }
672 
673 /*
674  * Package up an I/O request on a vnode into a uio and do it.  The I/O
675  * request is split up into smaller chunks and we try to avoid saturating
676  * the buffer cache while potentially holding a vnode locked, so we
677  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
678  * to give other processes a chance to lock the vnode (either other processes
679  * core'ing the same binary, or unrelated processes scanning the directory).
680  */
681 int
682 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
683     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
684     struct ucred *file_cred, size_t *aresid, struct thread *td)
685 {
686 	int error = 0;
687 	ssize_t iaresid;
688 
689 	do {
690 		int chunk;
691 
692 		/*
693 		 * Force `offset' to a multiple of MAXBSIZE except possibly
694 		 * for the first chunk, so that filesystems only need to
695 		 * write full blocks except possibly for the first and last
696 		 * chunks.
697 		 */
698 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
699 
700 		if (chunk > len)
701 			chunk = len;
702 		if (rw != UIO_READ && vp->v_type == VREG)
703 			bwillwrite();
704 		iaresid = 0;
705 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
706 		    ioflg, active_cred, file_cred, &iaresid, td);
707 		len -= chunk;	/* aresid calc already includes length */
708 		if (error)
709 			break;
710 		offset += chunk;
711 		base = (char *)base + chunk;
712 		kern_yield(PRI_USER);
713 	} while (len);
714 	if (aresid)
715 		*aresid = len + iaresid;
716 	return (error);
717 }
718 
719 #if OFF_MAX <= LONG_MAX
720 off_t
721 foffset_lock(struct file *fp, int flags)
722 {
723 	volatile short *flagsp;
724 	off_t res;
725 	short state;
726 
727 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
728 
729 	if ((flags & FOF_NOLOCK) != 0)
730 		return (atomic_load_long(&fp->f_offset));
731 
732 	/*
733 	 * According to McKusick the vn lock was protecting f_offset here.
734 	 * It is now protected by the FOFFSET_LOCKED flag.
735 	 */
736 	flagsp = &fp->f_vnread_flags;
737 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
738 		return (atomic_load_long(&fp->f_offset));
739 
740 	sleepq_lock(&fp->f_vnread_flags);
741 	state = atomic_load_16(flagsp);
742 	for (;;) {
743 		if ((state & FOFFSET_LOCKED) == 0) {
744 			if (!atomic_fcmpset_acq_16(flagsp, &state,
745 			    FOFFSET_LOCKED))
746 				continue;
747 			break;
748 		}
749 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
750 			if (!atomic_fcmpset_acq_16(flagsp, &state,
751 			    state | FOFFSET_LOCK_WAITING))
752 				continue;
753 		}
754 		DROP_GIANT();
755 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
756 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
757 		PICKUP_GIANT();
758 		sleepq_lock(&fp->f_vnread_flags);
759 		state = atomic_load_16(flagsp);
760 	}
761 	res = atomic_load_long(&fp->f_offset);
762 	sleepq_release(&fp->f_vnread_flags);
763 	return (res);
764 }
765 
766 void
767 foffset_unlock(struct file *fp, off_t val, int flags)
768 {
769 	volatile short *flagsp;
770 	short state;
771 
772 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
773 
774 	if ((flags & FOF_NOUPDATE) == 0)
775 		atomic_store_long(&fp->f_offset, val);
776 	if ((flags & FOF_NEXTOFF_R) != 0)
777 		fp->f_nextoff[UIO_READ] = val;
778 	if ((flags & FOF_NEXTOFF_W) != 0)
779 		fp->f_nextoff[UIO_WRITE] = val;
780 
781 	if ((flags & FOF_NOLOCK) != 0)
782 		return;
783 
784 	flagsp = &fp->f_vnread_flags;
785 	state = atomic_load_16(flagsp);
786 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
787 	    atomic_cmpset_rel_16(flagsp, state, 0))
788 		return;
789 
790 	sleepq_lock(&fp->f_vnread_flags);
791 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
792 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
793 	fp->f_vnread_flags = 0;
794 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
795 	sleepq_release(&fp->f_vnread_flags);
796 }
797 #else
798 off_t
799 foffset_lock(struct file *fp, int flags)
800 {
801 	struct mtx *mtxp;
802 	off_t res;
803 
804 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
805 
806 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
807 	mtx_lock(mtxp);
808 	if ((flags & FOF_NOLOCK) == 0) {
809 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
810 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
811 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
812 			    "vofflock", 0);
813 		}
814 		fp->f_vnread_flags |= FOFFSET_LOCKED;
815 	}
816 	res = fp->f_offset;
817 	mtx_unlock(mtxp);
818 	return (res);
819 }
820 
821 void
822 foffset_unlock(struct file *fp, off_t val, int flags)
823 {
824 	struct mtx *mtxp;
825 
826 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
827 
828 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
829 	mtx_lock(mtxp);
830 	if ((flags & FOF_NOUPDATE) == 0)
831 		fp->f_offset = val;
832 	if ((flags & FOF_NEXTOFF_R) != 0)
833 		fp->f_nextoff[UIO_READ] = val;
834 	if ((flags & FOF_NEXTOFF_W) != 0)
835 		fp->f_nextoff[UIO_WRITE] = val;
836 	if ((flags & FOF_NOLOCK) == 0) {
837 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
838 		    ("Lost FOFFSET_LOCKED"));
839 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
840 			wakeup(&fp->f_vnread_flags);
841 		fp->f_vnread_flags = 0;
842 	}
843 	mtx_unlock(mtxp);
844 }
845 #endif
846 
847 void
848 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
849 {
850 
851 	if ((flags & FOF_OFFSET) == 0)
852 		uio->uio_offset = foffset_lock(fp, flags);
853 }
854 
855 void
856 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
857 {
858 
859 	if ((flags & FOF_OFFSET) == 0)
860 		foffset_unlock(fp, uio->uio_offset, flags);
861 }
862 
863 static int
864 get_advice(struct file *fp, struct uio *uio)
865 {
866 	struct mtx *mtxp;
867 	int ret;
868 
869 	ret = POSIX_FADV_NORMAL;
870 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
871 		return (ret);
872 
873 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
874 	mtx_lock(mtxp);
875 	if (fp->f_advice != NULL &&
876 	    uio->uio_offset >= fp->f_advice->fa_start &&
877 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
878 		ret = fp->f_advice->fa_advice;
879 	mtx_unlock(mtxp);
880 	return (ret);
881 }
882 
883 int
884 vn_read_from_obj(struct vnode *vp, struct uio *uio)
885 {
886 	vm_object_t obj;
887 	vm_page_t ma[io_hold_cnt + 2];
888 	off_t off, vsz;
889 	ssize_t resid;
890 	int error, i, j;
891 
892 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
893 	obj = atomic_load_ptr(&vp->v_object);
894 	if (obj == NULL)
895 		return (EJUSTRETURN);
896 
897 	/*
898 	 * Depends on type stability of vm_objects.
899 	 */
900 	vm_object_pip_add(obj, 1);
901 	if ((obj->flags & OBJ_DEAD) != 0) {
902 		/*
903 		 * Note that object might be already reused from the
904 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
905 		 * we recheck for DOOMED vnode state after all pages
906 		 * are busied, and retract then.
907 		 *
908 		 * But we check for OBJ_DEAD to ensure that we do not
909 		 * busy pages while vm_object_terminate_pages()
910 		 * processes the queue.
911 		 */
912 		error = EJUSTRETURN;
913 		goto out_pip;
914 	}
915 
916 	resid = uio->uio_resid;
917 	off = uio->uio_offset;
918 	for (i = 0; resid > 0; i++) {
919 		MPASS(i < io_hold_cnt + 2);
920 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
921 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
922 		    VM_ALLOC_NOWAIT);
923 		if (ma[i] == NULL)
924 			break;
925 
926 		/*
927 		 * Skip invalid pages.  Valid mask can be partial only
928 		 * at EOF, and we clip later.
929 		 */
930 		if (vm_page_none_valid(ma[i])) {
931 			vm_page_sunbusy(ma[i]);
932 			break;
933 		}
934 
935 		resid -= PAGE_SIZE;
936 		off += PAGE_SIZE;
937 	}
938 	if (i == 0) {
939 		error = EJUSTRETURN;
940 		goto out_pip;
941 	}
942 
943 	/*
944 	 * Check VIRF_DOOMED after we busied our pages.  Since
945 	 * vgonel() terminates the vnode' vm_object, it cannot
946 	 * process past pages busied by us.
947 	 */
948 	if (VN_IS_DOOMED(vp)) {
949 		error = EJUSTRETURN;
950 		goto out;
951 	}
952 
953 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
954 	if (resid > uio->uio_resid)
955 		resid = uio->uio_resid;
956 
957 	/*
958 	 * Unlocked read of vnp_size is safe because truncation cannot
959 	 * pass busied page.  But we load vnp_size into a local
960 	 * variable so that possible concurrent extension does not
961 	 * break calculation.
962 	 */
963 #if defined(__powerpc__) && !defined(__powerpc64__)
964 	vsz = obj->un_pager.vnp.vnp_size;
965 #else
966 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
967 #endif
968 	if (uio->uio_offset >= vsz) {
969 		error = EJUSTRETURN;
970 		goto out;
971 	}
972 	if (uio->uio_offset + resid > vsz)
973 		resid = vsz - uio->uio_offset;
974 
975 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
976 
977 out:
978 	for (j = 0; j < i; j++) {
979 		if (error == 0)
980 			vm_page_reference(ma[j]);
981 		vm_page_sunbusy(ma[j]);
982 	}
983 out_pip:
984 	vm_object_pip_wakeup(obj);
985 	if (error != 0)
986 		return (error);
987 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
988 }
989 
990 /*
991  * File table vnode read routine.
992  */
993 static int
994 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
995     struct thread *td)
996 {
997 	struct vnode *vp;
998 	off_t orig_offset;
999 	int error, ioflag;
1000 	int advice;
1001 
1002 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1003 	    uio->uio_td, td));
1004 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1005 	vp = fp->f_vnode;
1006 	ioflag = 0;
1007 	if (fp->f_flag & FNONBLOCK)
1008 		ioflag |= IO_NDELAY;
1009 	if (fp->f_flag & O_DIRECT)
1010 		ioflag |= IO_DIRECT;
1011 
1012 	/*
1013 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1014 	 * allows us to avoid unneeded work outright.
1015 	 */
1016 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1017 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1018 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1019 		if (error == 0) {
1020 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1021 			return (0);
1022 		}
1023 		if (error != EJUSTRETURN)
1024 			return (error);
1025 	}
1026 
1027 	advice = get_advice(fp, uio);
1028 	vn_lock(vp, LK_SHARED | LK_RETRY);
1029 
1030 	switch (advice) {
1031 	case POSIX_FADV_NORMAL:
1032 	case POSIX_FADV_SEQUENTIAL:
1033 	case POSIX_FADV_NOREUSE:
1034 		ioflag |= sequential_heuristic(uio, fp);
1035 		break;
1036 	case POSIX_FADV_RANDOM:
1037 		/* Disable read-ahead for random I/O. */
1038 		break;
1039 	}
1040 	orig_offset = uio->uio_offset;
1041 
1042 #ifdef MAC
1043 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1044 	if (error == 0)
1045 #endif
1046 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1047 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1048 	VOP_UNLOCK(vp);
1049 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1050 	    orig_offset != uio->uio_offset)
1051 		/*
1052 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1053 		 * for the backing file after a POSIX_FADV_NOREUSE
1054 		 * read(2).
1055 		 */
1056 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1057 		    POSIX_FADV_DONTNEED);
1058 	return (error);
1059 }
1060 
1061 /*
1062  * File table vnode write routine.
1063  */
1064 static int
1065 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1066     struct thread *td)
1067 {
1068 	struct vnode *vp;
1069 	struct mount *mp;
1070 	off_t orig_offset;
1071 	int error, ioflag, lock_flags;
1072 	int advice;
1073 
1074 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1075 	    uio->uio_td, td));
1076 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1077 	vp = fp->f_vnode;
1078 	if (vp->v_type == VREG)
1079 		bwillwrite();
1080 	ioflag = IO_UNIT;
1081 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1082 		ioflag |= IO_APPEND;
1083 	if (fp->f_flag & FNONBLOCK)
1084 		ioflag |= IO_NDELAY;
1085 	if (fp->f_flag & O_DIRECT)
1086 		ioflag |= IO_DIRECT;
1087 	if ((fp->f_flag & O_FSYNC) ||
1088 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1089 		ioflag |= IO_SYNC;
1090 	/*
1091 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1092 	 * implementations that don't understand IO_DATASYNC fall back to full
1093 	 * O_SYNC behavior.
1094 	 */
1095 	if (fp->f_flag & O_DSYNC)
1096 		ioflag |= IO_SYNC | IO_DATASYNC;
1097 	mp = NULL;
1098 	if (vp->v_type != VCHR &&
1099 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1100 		goto unlock;
1101 
1102 	advice = get_advice(fp, uio);
1103 
1104 	if (MNT_SHARED_WRITES(mp) ||
1105 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1106 		lock_flags = LK_SHARED;
1107 	} else {
1108 		lock_flags = LK_EXCLUSIVE;
1109 	}
1110 
1111 	vn_lock(vp, lock_flags | LK_RETRY);
1112 	switch (advice) {
1113 	case POSIX_FADV_NORMAL:
1114 	case POSIX_FADV_SEQUENTIAL:
1115 	case POSIX_FADV_NOREUSE:
1116 		ioflag |= sequential_heuristic(uio, fp);
1117 		break;
1118 	case POSIX_FADV_RANDOM:
1119 		/* XXX: Is this correct? */
1120 		break;
1121 	}
1122 	orig_offset = uio->uio_offset;
1123 
1124 #ifdef MAC
1125 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1126 	if (error == 0)
1127 #endif
1128 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1129 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1130 	VOP_UNLOCK(vp);
1131 	if (vp->v_type != VCHR)
1132 		vn_finished_write(mp);
1133 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1134 	    orig_offset != uio->uio_offset)
1135 		/*
1136 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1137 		 * for the backing file after a POSIX_FADV_NOREUSE
1138 		 * write(2).
1139 		 */
1140 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1141 		    POSIX_FADV_DONTNEED);
1142 unlock:
1143 	return (error);
1144 }
1145 
1146 /*
1147  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1148  * prevent the following deadlock:
1149  *
1150  * Assume that the thread A reads from the vnode vp1 into userspace
1151  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1152  * currently not resident, then system ends up with the call chain
1153  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1154  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1155  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1156  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1157  * backed by the pages of vnode vp1, and some page in buf2 is not
1158  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1159  *
1160  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1161  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1162  * Instead, it first tries to do the whole range i/o with pagefaults
1163  * disabled. If all pages in the i/o buffer are resident and mapped,
1164  * VOP will succeed (ignoring the genuine filesystem errors).
1165  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1166  * i/o in chunks, with all pages in the chunk prefaulted and held
1167  * using vm_fault_quick_hold_pages().
1168  *
1169  * Filesystems using this deadlock avoidance scheme should use the
1170  * array of the held pages from uio, saved in the curthread->td_ma,
1171  * instead of doing uiomove().  A helper function
1172  * vn_io_fault_uiomove() converts uiomove request into
1173  * uiomove_fromphys() over td_ma array.
1174  *
1175  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1176  * make the current i/o request atomic with respect to other i/os and
1177  * truncations.
1178  */
1179 
1180 /*
1181  * Decode vn_io_fault_args and perform the corresponding i/o.
1182  */
1183 static int
1184 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1185     struct thread *td)
1186 {
1187 	int error, save;
1188 
1189 	error = 0;
1190 	save = vm_fault_disable_pagefaults();
1191 	switch (args->kind) {
1192 	case VN_IO_FAULT_FOP:
1193 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1194 		    uio, args->cred, args->flags, td);
1195 		break;
1196 	case VN_IO_FAULT_VOP:
1197 		if (uio->uio_rw == UIO_READ) {
1198 			error = VOP_READ(args->args.vop_args.vp, uio,
1199 			    args->flags, args->cred);
1200 		} else if (uio->uio_rw == UIO_WRITE) {
1201 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1202 			    args->flags, args->cred);
1203 		}
1204 		break;
1205 	default:
1206 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1207 		    args->kind, uio->uio_rw);
1208 	}
1209 	vm_fault_enable_pagefaults(save);
1210 	return (error);
1211 }
1212 
1213 static int
1214 vn_io_fault_touch(char *base, const struct uio *uio)
1215 {
1216 	int r;
1217 
1218 	r = fubyte(base);
1219 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1220 		return (EFAULT);
1221 	return (0);
1222 }
1223 
1224 static int
1225 vn_io_fault_prefault_user(const struct uio *uio)
1226 {
1227 	char *base;
1228 	const struct iovec *iov;
1229 	size_t len;
1230 	ssize_t resid;
1231 	int error, i;
1232 
1233 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1234 	    ("vn_io_fault_prefault userspace"));
1235 
1236 	error = i = 0;
1237 	iov = uio->uio_iov;
1238 	resid = uio->uio_resid;
1239 	base = iov->iov_base;
1240 	len = iov->iov_len;
1241 	while (resid > 0) {
1242 		error = vn_io_fault_touch(base, uio);
1243 		if (error != 0)
1244 			break;
1245 		if (len < PAGE_SIZE) {
1246 			if (len != 0) {
1247 				error = vn_io_fault_touch(base + len - 1, uio);
1248 				if (error != 0)
1249 					break;
1250 				resid -= len;
1251 			}
1252 			if (++i >= uio->uio_iovcnt)
1253 				break;
1254 			iov = uio->uio_iov + i;
1255 			base = iov->iov_base;
1256 			len = iov->iov_len;
1257 		} else {
1258 			len -= PAGE_SIZE;
1259 			base += PAGE_SIZE;
1260 			resid -= PAGE_SIZE;
1261 		}
1262 	}
1263 	return (error);
1264 }
1265 
1266 /*
1267  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1268  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1269  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1270  * into args and call vn_io_fault1() to handle faults during the user
1271  * mode buffer accesses.
1272  */
1273 static int
1274 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1275     struct thread *td)
1276 {
1277 	vm_page_t ma[io_hold_cnt + 2];
1278 	struct uio *uio_clone, short_uio;
1279 	struct iovec short_iovec[1];
1280 	vm_page_t *prev_td_ma;
1281 	vm_prot_t prot;
1282 	vm_offset_t addr, end;
1283 	size_t len, resid;
1284 	ssize_t adv;
1285 	int error, cnt, saveheld, prev_td_ma_cnt;
1286 
1287 	if (vn_io_fault_prefault) {
1288 		error = vn_io_fault_prefault_user(uio);
1289 		if (error != 0)
1290 			return (error); /* Or ignore ? */
1291 	}
1292 
1293 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1294 
1295 	/*
1296 	 * The UFS follows IO_UNIT directive and replays back both
1297 	 * uio_offset and uio_resid if an error is encountered during the
1298 	 * operation.  But, since the iovec may be already advanced,
1299 	 * uio is still in an inconsistent state.
1300 	 *
1301 	 * Cache a copy of the original uio, which is advanced to the redo
1302 	 * point using UIO_NOCOPY below.
1303 	 */
1304 	uio_clone = cloneuio(uio);
1305 	resid = uio->uio_resid;
1306 
1307 	short_uio.uio_segflg = UIO_USERSPACE;
1308 	short_uio.uio_rw = uio->uio_rw;
1309 	short_uio.uio_td = uio->uio_td;
1310 
1311 	error = vn_io_fault_doio(args, uio, td);
1312 	if (error != EFAULT)
1313 		goto out;
1314 
1315 	atomic_add_long(&vn_io_faults_cnt, 1);
1316 	uio_clone->uio_segflg = UIO_NOCOPY;
1317 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1318 	uio_clone->uio_segflg = uio->uio_segflg;
1319 
1320 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1321 	prev_td_ma = td->td_ma;
1322 	prev_td_ma_cnt = td->td_ma_cnt;
1323 
1324 	while (uio_clone->uio_resid != 0) {
1325 		len = uio_clone->uio_iov->iov_len;
1326 		if (len == 0) {
1327 			KASSERT(uio_clone->uio_iovcnt >= 1,
1328 			    ("iovcnt underflow"));
1329 			uio_clone->uio_iov++;
1330 			uio_clone->uio_iovcnt--;
1331 			continue;
1332 		}
1333 		if (len > ptoa(io_hold_cnt))
1334 			len = ptoa(io_hold_cnt);
1335 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1336 		end = round_page(addr + len);
1337 		if (end < addr) {
1338 			error = EFAULT;
1339 			break;
1340 		}
1341 		cnt = atop(end - trunc_page(addr));
1342 		/*
1343 		 * A perfectly misaligned address and length could cause
1344 		 * both the start and the end of the chunk to use partial
1345 		 * page.  +2 accounts for such a situation.
1346 		 */
1347 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1348 		    addr, len, prot, ma, io_hold_cnt + 2);
1349 		if (cnt == -1) {
1350 			error = EFAULT;
1351 			break;
1352 		}
1353 		short_uio.uio_iov = &short_iovec[0];
1354 		short_iovec[0].iov_base = (void *)addr;
1355 		short_uio.uio_iovcnt = 1;
1356 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1357 		short_uio.uio_offset = uio_clone->uio_offset;
1358 		td->td_ma = ma;
1359 		td->td_ma_cnt = cnt;
1360 
1361 		error = vn_io_fault_doio(args, &short_uio, td);
1362 		vm_page_unhold_pages(ma, cnt);
1363 		adv = len - short_uio.uio_resid;
1364 
1365 		uio_clone->uio_iov->iov_base =
1366 		    (char *)uio_clone->uio_iov->iov_base + adv;
1367 		uio_clone->uio_iov->iov_len -= adv;
1368 		uio_clone->uio_resid -= adv;
1369 		uio_clone->uio_offset += adv;
1370 
1371 		uio->uio_resid -= adv;
1372 		uio->uio_offset += adv;
1373 
1374 		if (error != 0 || adv == 0)
1375 			break;
1376 	}
1377 	td->td_ma = prev_td_ma;
1378 	td->td_ma_cnt = prev_td_ma_cnt;
1379 	curthread_pflags_restore(saveheld);
1380 out:
1381 	free(uio_clone, M_IOV);
1382 	return (error);
1383 }
1384 
1385 static int
1386 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1387     int flags, struct thread *td)
1388 {
1389 	fo_rdwr_t *doio;
1390 	struct vnode *vp;
1391 	void *rl_cookie;
1392 	struct vn_io_fault_args args;
1393 	int error;
1394 
1395 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1396 	vp = fp->f_vnode;
1397 
1398 	/*
1399 	 * The ability to read(2) on a directory has historically been
1400 	 * allowed for all users, but this can and has been the source of
1401 	 * at least one security issue in the past.  As such, it is now hidden
1402 	 * away behind a sysctl for those that actually need it to use it, and
1403 	 * restricted to root when it's turned on to make it relatively safe to
1404 	 * leave on for longer sessions of need.
1405 	 */
1406 	if (vp->v_type == VDIR) {
1407 		KASSERT(uio->uio_rw == UIO_READ,
1408 		    ("illegal write attempted on a directory"));
1409 		if (!vfs_allow_read_dir)
1410 			return (EISDIR);
1411 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1412 			return (EISDIR);
1413 	}
1414 
1415 	foffset_lock_uio(fp, uio, flags);
1416 	if (do_vn_io_fault(vp, uio)) {
1417 		args.kind = VN_IO_FAULT_FOP;
1418 		args.args.fop_args.fp = fp;
1419 		args.args.fop_args.doio = doio;
1420 		args.cred = active_cred;
1421 		args.flags = flags | FOF_OFFSET;
1422 		if (uio->uio_rw == UIO_READ) {
1423 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1424 			    uio->uio_offset + uio->uio_resid);
1425 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1426 		    (flags & FOF_OFFSET) == 0) {
1427 			/* For appenders, punt and lock the whole range. */
1428 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1429 		} else {
1430 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1431 			    uio->uio_offset + uio->uio_resid);
1432 		}
1433 		error = vn_io_fault1(vp, uio, &args, td);
1434 		vn_rangelock_unlock(vp, rl_cookie);
1435 	} else {
1436 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1437 	}
1438 	foffset_unlock_uio(fp, uio, flags);
1439 	return (error);
1440 }
1441 
1442 /*
1443  * Helper function to perform the requested uiomove operation using
1444  * the held pages for io->uio_iov[0].iov_base buffer instead of
1445  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1446  * instead of iov_base prevents page faults that could occur due to
1447  * pmap_collect() invalidating the mapping created by
1448  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1449  * object cleanup revoking the write access from page mappings.
1450  *
1451  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1452  * instead of plain uiomove().
1453  */
1454 int
1455 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1456 {
1457 	struct uio transp_uio;
1458 	struct iovec transp_iov[1];
1459 	struct thread *td;
1460 	size_t adv;
1461 	int error, pgadv;
1462 
1463 	td = curthread;
1464 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1465 	    uio->uio_segflg != UIO_USERSPACE)
1466 		return (uiomove(data, xfersize, uio));
1467 
1468 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1469 	transp_iov[0].iov_base = data;
1470 	transp_uio.uio_iov = &transp_iov[0];
1471 	transp_uio.uio_iovcnt = 1;
1472 	if (xfersize > uio->uio_resid)
1473 		xfersize = uio->uio_resid;
1474 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1475 	transp_uio.uio_offset = 0;
1476 	transp_uio.uio_segflg = UIO_SYSSPACE;
1477 	/*
1478 	 * Since transp_iov points to data, and td_ma page array
1479 	 * corresponds to original uio->uio_iov, we need to invert the
1480 	 * direction of the i/o operation as passed to
1481 	 * uiomove_fromphys().
1482 	 */
1483 	switch (uio->uio_rw) {
1484 	case UIO_WRITE:
1485 		transp_uio.uio_rw = UIO_READ;
1486 		break;
1487 	case UIO_READ:
1488 		transp_uio.uio_rw = UIO_WRITE;
1489 		break;
1490 	}
1491 	transp_uio.uio_td = uio->uio_td;
1492 	error = uiomove_fromphys(td->td_ma,
1493 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1494 	    xfersize, &transp_uio);
1495 	adv = xfersize - transp_uio.uio_resid;
1496 	pgadv =
1497 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1498 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1499 	td->td_ma += pgadv;
1500 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1501 	    pgadv));
1502 	td->td_ma_cnt -= pgadv;
1503 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1504 	uio->uio_iov->iov_len -= adv;
1505 	uio->uio_resid -= adv;
1506 	uio->uio_offset += adv;
1507 	return (error);
1508 }
1509 
1510 int
1511 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1512     struct uio *uio)
1513 {
1514 	struct thread *td;
1515 	vm_offset_t iov_base;
1516 	int cnt, pgadv;
1517 
1518 	td = curthread;
1519 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1520 	    uio->uio_segflg != UIO_USERSPACE)
1521 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1522 
1523 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1524 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1525 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1526 	switch (uio->uio_rw) {
1527 	case UIO_WRITE:
1528 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1529 		    offset, cnt);
1530 		break;
1531 	case UIO_READ:
1532 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1533 		    cnt);
1534 		break;
1535 	}
1536 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1537 	td->td_ma += pgadv;
1538 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1539 	    pgadv));
1540 	td->td_ma_cnt -= pgadv;
1541 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1542 	uio->uio_iov->iov_len -= cnt;
1543 	uio->uio_resid -= cnt;
1544 	uio->uio_offset += cnt;
1545 	return (0);
1546 }
1547 
1548 /*
1549  * File table truncate routine.
1550  */
1551 static int
1552 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1553     struct thread *td)
1554 {
1555 	struct mount *mp;
1556 	struct vnode *vp;
1557 	void *rl_cookie;
1558 	int error;
1559 
1560 	vp = fp->f_vnode;
1561 
1562 retry:
1563 	/*
1564 	 * Lock the whole range for truncation.  Otherwise split i/o
1565 	 * might happen partly before and partly after the truncation.
1566 	 */
1567 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1568 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1569 	if (error)
1570 		goto out1;
1571 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1572 	AUDIT_ARG_VNODE1(vp);
1573 	if (vp->v_type == VDIR) {
1574 		error = EISDIR;
1575 		goto out;
1576 	}
1577 #ifdef MAC
1578 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1579 	if (error)
1580 		goto out;
1581 #endif
1582 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1583 	    fp->f_cred);
1584 out:
1585 	VOP_UNLOCK(vp);
1586 	vn_finished_write(mp);
1587 out1:
1588 	vn_rangelock_unlock(vp, rl_cookie);
1589 	if (error == ERELOOKUP)
1590 		goto retry;
1591 	return (error);
1592 }
1593 
1594 /*
1595  * Truncate a file that is already locked.
1596  */
1597 int
1598 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1599     struct ucred *cred)
1600 {
1601 	struct vattr vattr;
1602 	int error;
1603 
1604 	error = VOP_ADD_WRITECOUNT(vp, 1);
1605 	if (error == 0) {
1606 		VATTR_NULL(&vattr);
1607 		vattr.va_size = length;
1608 		if (sync)
1609 			vattr.va_vaflags |= VA_SYNC;
1610 		error = VOP_SETATTR(vp, &vattr, cred);
1611 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1612 	}
1613 	return (error);
1614 }
1615 
1616 /*
1617  * File table vnode stat routine.
1618  */
1619 static int
1620 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1621     struct thread *td)
1622 {
1623 	struct vnode *vp = fp->f_vnode;
1624 	int error;
1625 
1626 	vn_lock(vp, LK_SHARED | LK_RETRY);
1627 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1628 	VOP_UNLOCK(vp);
1629 
1630 	return (error);
1631 }
1632 
1633 /*
1634  * File table vnode ioctl routine.
1635  */
1636 static int
1637 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1638     struct thread *td)
1639 {
1640 	struct vattr vattr;
1641 	struct vnode *vp;
1642 	struct fiobmap2_arg *bmarg;
1643 	int error;
1644 
1645 	vp = fp->f_vnode;
1646 	switch (vp->v_type) {
1647 	case VDIR:
1648 	case VREG:
1649 		switch (com) {
1650 		case FIONREAD:
1651 			vn_lock(vp, LK_SHARED | LK_RETRY);
1652 			error = VOP_GETATTR(vp, &vattr, active_cred);
1653 			VOP_UNLOCK(vp);
1654 			if (error == 0)
1655 				*(int *)data = vattr.va_size - fp->f_offset;
1656 			return (error);
1657 		case FIOBMAP2:
1658 			bmarg = (struct fiobmap2_arg *)data;
1659 			vn_lock(vp, LK_SHARED | LK_RETRY);
1660 #ifdef MAC
1661 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1662 			    vp);
1663 			if (error == 0)
1664 #endif
1665 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1666 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1667 			VOP_UNLOCK(vp);
1668 			return (error);
1669 		case FIONBIO:
1670 		case FIOASYNC:
1671 			return (0);
1672 		default:
1673 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1674 			    active_cred, td));
1675 		}
1676 		break;
1677 	case VCHR:
1678 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1679 		    active_cred, td));
1680 	default:
1681 		return (ENOTTY);
1682 	}
1683 }
1684 
1685 /*
1686  * File table vnode poll routine.
1687  */
1688 static int
1689 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1690     struct thread *td)
1691 {
1692 	struct vnode *vp;
1693 	int error;
1694 
1695 	vp = fp->f_vnode;
1696 #if defined(MAC) || defined(AUDIT)
1697 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1698 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1699 		AUDIT_ARG_VNODE1(vp);
1700 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1701 		VOP_UNLOCK(vp);
1702 		if (error != 0)
1703 			return (error);
1704 	}
1705 #endif
1706 	error = VOP_POLL(vp, events, fp->f_cred, td);
1707 	return (error);
1708 }
1709 
1710 /*
1711  * Acquire the requested lock and then check for validity.  LK_RETRY
1712  * permits vn_lock to return doomed vnodes.
1713  */
1714 static int __noinline
1715 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1716     int error)
1717 {
1718 
1719 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1720 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1721 
1722 	if (error == 0)
1723 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1724 
1725 	if ((flags & LK_RETRY) == 0) {
1726 		if (error == 0) {
1727 			VOP_UNLOCK(vp);
1728 			error = ENOENT;
1729 		}
1730 		return (error);
1731 	}
1732 
1733 	/*
1734 	 * LK_RETRY case.
1735 	 *
1736 	 * Nothing to do if we got the lock.
1737 	 */
1738 	if (error == 0)
1739 		return (0);
1740 
1741 	/*
1742 	 * Interlock was dropped by the call in _vn_lock.
1743 	 */
1744 	flags &= ~LK_INTERLOCK;
1745 	do {
1746 		error = VOP_LOCK1(vp, flags, file, line);
1747 	} while (error != 0);
1748 	return (0);
1749 }
1750 
1751 int
1752 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1753 {
1754 	int error;
1755 
1756 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1757 	    ("vn_lock: no locktype (%d passed)", flags));
1758 	VNPASS(vp->v_holdcnt > 0, vp);
1759 	error = VOP_LOCK1(vp, flags, file, line);
1760 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1761 		return (_vn_lock_fallback(vp, flags, file, line, error));
1762 	return (0);
1763 }
1764 
1765 /*
1766  * File table vnode close routine.
1767  */
1768 static int
1769 vn_closefile(struct file *fp, struct thread *td)
1770 {
1771 	struct vnode *vp;
1772 	struct flock lf;
1773 	int error;
1774 	bool ref;
1775 
1776 	vp = fp->f_vnode;
1777 	fp->f_ops = &badfileops;
1778 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1779 
1780 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1781 
1782 	if (__predict_false(ref)) {
1783 		lf.l_whence = SEEK_SET;
1784 		lf.l_start = 0;
1785 		lf.l_len = 0;
1786 		lf.l_type = F_UNLCK;
1787 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1788 		vrele(vp);
1789 	}
1790 	return (error);
1791 }
1792 
1793 /*
1794  * Preparing to start a filesystem write operation. If the operation is
1795  * permitted, then we bump the count of operations in progress and
1796  * proceed. If a suspend request is in progress, we wait until the
1797  * suspension is over, and then proceed.
1798  */
1799 static int
1800 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1801 {
1802 	struct mount_pcpu *mpcpu;
1803 	int error, mflags;
1804 
1805 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1806 	    vfs_op_thread_enter(mp, mpcpu)) {
1807 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1808 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1809 		vfs_op_thread_exit(mp, mpcpu);
1810 		return (0);
1811 	}
1812 
1813 	if (mplocked)
1814 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1815 	else
1816 		MNT_ILOCK(mp);
1817 
1818 	error = 0;
1819 
1820 	/*
1821 	 * Check on status of suspension.
1822 	 */
1823 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1824 	    mp->mnt_susp_owner != curthread) {
1825 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1826 		    (flags & PCATCH) : 0) | (PUSER - 1);
1827 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1828 			if (flags & V_NOWAIT) {
1829 				error = EWOULDBLOCK;
1830 				goto unlock;
1831 			}
1832 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1833 			    "suspfs", 0);
1834 			if (error)
1835 				goto unlock;
1836 		}
1837 	}
1838 	if (flags & V_XSLEEP)
1839 		goto unlock;
1840 	mp->mnt_writeopcount++;
1841 unlock:
1842 	if (error != 0 || (flags & V_XSLEEP) != 0)
1843 		MNT_REL(mp);
1844 	MNT_IUNLOCK(mp);
1845 	return (error);
1846 }
1847 
1848 int
1849 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1850 {
1851 	struct mount *mp;
1852 	int error;
1853 
1854 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1855 	    ("V_MNTREF requires mp"));
1856 
1857 	error = 0;
1858 	/*
1859 	 * If a vnode is provided, get and return the mount point that
1860 	 * to which it will write.
1861 	 */
1862 	if (vp != NULL) {
1863 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1864 			*mpp = NULL;
1865 			if (error != EOPNOTSUPP)
1866 				return (error);
1867 			return (0);
1868 		}
1869 	}
1870 	if ((mp = *mpp) == NULL)
1871 		return (0);
1872 
1873 	/*
1874 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1875 	 * a vfs_ref().
1876 	 * As long as a vnode is not provided we need to acquire a
1877 	 * refcount for the provided mountpoint too, in order to
1878 	 * emulate a vfs_ref().
1879 	 */
1880 	if (vp == NULL && (flags & V_MNTREF) == 0)
1881 		vfs_ref(mp);
1882 
1883 	return (vn_start_write_refed(mp, flags, false));
1884 }
1885 
1886 /*
1887  * Secondary suspension. Used by operations such as vop_inactive
1888  * routines that are needed by the higher level functions. These
1889  * are allowed to proceed until all the higher level functions have
1890  * completed (indicated by mnt_writeopcount dropping to zero). At that
1891  * time, these operations are halted until the suspension is over.
1892  */
1893 int
1894 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1895 {
1896 	struct mount *mp;
1897 	int error;
1898 
1899 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1900 	    ("V_MNTREF requires mp"));
1901 
1902  retry:
1903 	if (vp != NULL) {
1904 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1905 			*mpp = NULL;
1906 			if (error != EOPNOTSUPP)
1907 				return (error);
1908 			return (0);
1909 		}
1910 	}
1911 	/*
1912 	 * If we are not suspended or have not yet reached suspended
1913 	 * mode, then let the operation proceed.
1914 	 */
1915 	if ((mp = *mpp) == NULL)
1916 		return (0);
1917 
1918 	/*
1919 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1920 	 * a vfs_ref().
1921 	 * As long as a vnode is not provided we need to acquire a
1922 	 * refcount for the provided mountpoint too, in order to
1923 	 * emulate a vfs_ref().
1924 	 */
1925 	MNT_ILOCK(mp);
1926 	if (vp == NULL && (flags & V_MNTREF) == 0)
1927 		MNT_REF(mp);
1928 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1929 		mp->mnt_secondary_writes++;
1930 		mp->mnt_secondary_accwrites++;
1931 		MNT_IUNLOCK(mp);
1932 		return (0);
1933 	}
1934 	if (flags & V_NOWAIT) {
1935 		MNT_REL(mp);
1936 		MNT_IUNLOCK(mp);
1937 		return (EWOULDBLOCK);
1938 	}
1939 	/*
1940 	 * Wait for the suspension to finish.
1941 	 */
1942 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1943 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1944 	    "suspfs", 0);
1945 	vfs_rel(mp);
1946 	if (error == 0)
1947 		goto retry;
1948 	return (error);
1949 }
1950 
1951 /*
1952  * Filesystem write operation has completed. If we are suspending and this
1953  * operation is the last one, notify the suspender that the suspension is
1954  * now in effect.
1955  */
1956 void
1957 vn_finished_write(struct mount *mp)
1958 {
1959 	struct mount_pcpu *mpcpu;
1960 	int c;
1961 
1962 	if (mp == NULL)
1963 		return;
1964 
1965 	if (vfs_op_thread_enter(mp, mpcpu)) {
1966 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
1967 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
1968 		vfs_op_thread_exit(mp, mpcpu);
1969 		return;
1970 	}
1971 
1972 	MNT_ILOCK(mp);
1973 	vfs_assert_mount_counters(mp);
1974 	MNT_REL(mp);
1975 	c = --mp->mnt_writeopcount;
1976 	if (mp->mnt_vfs_ops == 0) {
1977 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1978 		MNT_IUNLOCK(mp);
1979 		return;
1980 	}
1981 	if (c < 0)
1982 		vfs_dump_mount_counters(mp);
1983 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1984 		wakeup(&mp->mnt_writeopcount);
1985 	MNT_IUNLOCK(mp);
1986 }
1987 
1988 /*
1989  * Filesystem secondary write operation has completed. If we are
1990  * suspending and this operation is the last one, notify the suspender
1991  * that the suspension is now in effect.
1992  */
1993 void
1994 vn_finished_secondary_write(struct mount *mp)
1995 {
1996 	if (mp == NULL)
1997 		return;
1998 	MNT_ILOCK(mp);
1999 	MNT_REL(mp);
2000 	mp->mnt_secondary_writes--;
2001 	if (mp->mnt_secondary_writes < 0)
2002 		panic("vn_finished_secondary_write: neg cnt");
2003 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2004 	    mp->mnt_secondary_writes <= 0)
2005 		wakeup(&mp->mnt_secondary_writes);
2006 	MNT_IUNLOCK(mp);
2007 }
2008 
2009 /*
2010  * Request a filesystem to suspend write operations.
2011  */
2012 int
2013 vfs_write_suspend(struct mount *mp, int flags)
2014 {
2015 	int error;
2016 
2017 	vfs_op_enter(mp);
2018 
2019 	MNT_ILOCK(mp);
2020 	vfs_assert_mount_counters(mp);
2021 	if (mp->mnt_susp_owner == curthread) {
2022 		vfs_op_exit_locked(mp);
2023 		MNT_IUNLOCK(mp);
2024 		return (EALREADY);
2025 	}
2026 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2027 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2028 
2029 	/*
2030 	 * Unmount holds a write reference on the mount point.  If we
2031 	 * own busy reference and drain for writers, we deadlock with
2032 	 * the reference draining in the unmount path.  Callers of
2033 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2034 	 * vfs_busy() reference is owned and caller is not in the
2035 	 * unmount context.
2036 	 */
2037 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2038 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2039 		vfs_op_exit_locked(mp);
2040 		MNT_IUNLOCK(mp);
2041 		return (EBUSY);
2042 	}
2043 
2044 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2045 	mp->mnt_susp_owner = curthread;
2046 	if (mp->mnt_writeopcount > 0)
2047 		(void) msleep(&mp->mnt_writeopcount,
2048 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2049 	else
2050 		MNT_IUNLOCK(mp);
2051 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2052 		vfs_write_resume(mp, 0);
2053 		/* vfs_write_resume does vfs_op_exit() for us */
2054 	}
2055 	return (error);
2056 }
2057 
2058 /*
2059  * Request a filesystem to resume write operations.
2060  */
2061 void
2062 vfs_write_resume(struct mount *mp, int flags)
2063 {
2064 
2065 	MNT_ILOCK(mp);
2066 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2067 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2068 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2069 				       MNTK_SUSPENDED);
2070 		mp->mnt_susp_owner = NULL;
2071 		wakeup(&mp->mnt_writeopcount);
2072 		wakeup(&mp->mnt_flag);
2073 		curthread->td_pflags &= ~TDP_IGNSUSP;
2074 		if ((flags & VR_START_WRITE) != 0) {
2075 			MNT_REF(mp);
2076 			mp->mnt_writeopcount++;
2077 		}
2078 		MNT_IUNLOCK(mp);
2079 		if ((flags & VR_NO_SUSPCLR) == 0)
2080 			VFS_SUSP_CLEAN(mp);
2081 		vfs_op_exit(mp);
2082 	} else if ((flags & VR_START_WRITE) != 0) {
2083 		MNT_REF(mp);
2084 		vn_start_write_refed(mp, 0, true);
2085 	} else {
2086 		MNT_IUNLOCK(mp);
2087 	}
2088 }
2089 
2090 /*
2091  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2092  * methods.
2093  */
2094 int
2095 vfs_write_suspend_umnt(struct mount *mp)
2096 {
2097 	int error;
2098 
2099 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2100 	    ("vfs_write_suspend_umnt: recursed"));
2101 
2102 	/* dounmount() already called vn_start_write(). */
2103 	for (;;) {
2104 		vn_finished_write(mp);
2105 		error = vfs_write_suspend(mp, 0);
2106 		if (error != 0) {
2107 			vn_start_write(NULL, &mp, V_WAIT);
2108 			return (error);
2109 		}
2110 		MNT_ILOCK(mp);
2111 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2112 			break;
2113 		MNT_IUNLOCK(mp);
2114 		vn_start_write(NULL, &mp, V_WAIT);
2115 	}
2116 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2117 	wakeup(&mp->mnt_flag);
2118 	MNT_IUNLOCK(mp);
2119 	curthread->td_pflags |= TDP_IGNSUSP;
2120 	return (0);
2121 }
2122 
2123 /*
2124  * Implement kqueues for files by translating it to vnode operation.
2125  */
2126 static int
2127 vn_kqfilter(struct file *fp, struct knote *kn)
2128 {
2129 
2130 	return (VOP_KQFILTER(fp->f_vnode, kn));
2131 }
2132 
2133 /*
2134  * Simplified in-kernel wrapper calls for extended attribute access.
2135  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2136  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2137  */
2138 int
2139 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2140     const char *attrname, int *buflen, char *buf, struct thread *td)
2141 {
2142 	struct uio	auio;
2143 	struct iovec	iov;
2144 	int	error;
2145 
2146 	iov.iov_len = *buflen;
2147 	iov.iov_base = buf;
2148 
2149 	auio.uio_iov = &iov;
2150 	auio.uio_iovcnt = 1;
2151 	auio.uio_rw = UIO_READ;
2152 	auio.uio_segflg = UIO_SYSSPACE;
2153 	auio.uio_td = td;
2154 	auio.uio_offset = 0;
2155 	auio.uio_resid = *buflen;
2156 
2157 	if ((ioflg & IO_NODELOCKED) == 0)
2158 		vn_lock(vp, LK_SHARED | LK_RETRY);
2159 
2160 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2161 
2162 	/* authorize attribute retrieval as kernel */
2163 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2164 	    td);
2165 
2166 	if ((ioflg & IO_NODELOCKED) == 0)
2167 		VOP_UNLOCK(vp);
2168 
2169 	if (error == 0) {
2170 		*buflen = *buflen - auio.uio_resid;
2171 	}
2172 
2173 	return (error);
2174 }
2175 
2176 /*
2177  * XXX failure mode if partially written?
2178  */
2179 int
2180 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2181     const char *attrname, int buflen, char *buf, struct thread *td)
2182 {
2183 	struct uio	auio;
2184 	struct iovec	iov;
2185 	struct mount	*mp;
2186 	int	error;
2187 
2188 	iov.iov_len = buflen;
2189 	iov.iov_base = buf;
2190 
2191 	auio.uio_iov = &iov;
2192 	auio.uio_iovcnt = 1;
2193 	auio.uio_rw = UIO_WRITE;
2194 	auio.uio_segflg = UIO_SYSSPACE;
2195 	auio.uio_td = td;
2196 	auio.uio_offset = 0;
2197 	auio.uio_resid = buflen;
2198 
2199 	if ((ioflg & IO_NODELOCKED) == 0) {
2200 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2201 			return (error);
2202 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2203 	}
2204 
2205 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2206 
2207 	/* authorize attribute setting as kernel */
2208 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2209 
2210 	if ((ioflg & IO_NODELOCKED) == 0) {
2211 		vn_finished_write(mp);
2212 		VOP_UNLOCK(vp);
2213 	}
2214 
2215 	return (error);
2216 }
2217 
2218 int
2219 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2220     const char *attrname, struct thread *td)
2221 {
2222 	struct mount	*mp;
2223 	int	error;
2224 
2225 	if ((ioflg & IO_NODELOCKED) == 0) {
2226 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2227 			return (error);
2228 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2229 	}
2230 
2231 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2232 
2233 	/* authorize attribute removal as kernel */
2234 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2235 	if (error == EOPNOTSUPP)
2236 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2237 		    NULL, td);
2238 
2239 	if ((ioflg & IO_NODELOCKED) == 0) {
2240 		vn_finished_write(mp);
2241 		VOP_UNLOCK(vp);
2242 	}
2243 
2244 	return (error);
2245 }
2246 
2247 static int
2248 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2249     struct vnode **rvp)
2250 {
2251 
2252 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2253 }
2254 
2255 int
2256 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2257 {
2258 
2259 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2260 	    lkflags, rvp));
2261 }
2262 
2263 int
2264 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2265     int lkflags, struct vnode **rvp)
2266 {
2267 	struct mount *mp;
2268 	int ltype, error;
2269 
2270 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2271 	mp = vp->v_mount;
2272 	ltype = VOP_ISLOCKED(vp);
2273 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2274 	    ("vn_vget_ino: vp not locked"));
2275 	error = vfs_busy(mp, MBF_NOWAIT);
2276 	if (error != 0) {
2277 		vfs_ref(mp);
2278 		VOP_UNLOCK(vp);
2279 		error = vfs_busy(mp, 0);
2280 		vn_lock(vp, ltype | LK_RETRY);
2281 		vfs_rel(mp);
2282 		if (error != 0)
2283 			return (ENOENT);
2284 		if (VN_IS_DOOMED(vp)) {
2285 			vfs_unbusy(mp);
2286 			return (ENOENT);
2287 		}
2288 	}
2289 	VOP_UNLOCK(vp);
2290 	error = alloc(mp, alloc_arg, lkflags, rvp);
2291 	vfs_unbusy(mp);
2292 	if (error != 0 || *rvp != vp)
2293 		vn_lock(vp, ltype | LK_RETRY);
2294 	if (VN_IS_DOOMED(vp)) {
2295 		if (error == 0) {
2296 			if (*rvp == vp)
2297 				vunref(vp);
2298 			else
2299 				vput(*rvp);
2300 		}
2301 		error = ENOENT;
2302 	}
2303 	return (error);
2304 }
2305 
2306 int
2307 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2308     struct thread *td)
2309 {
2310 
2311 	if (vp->v_type != VREG || td == NULL)
2312 		return (0);
2313 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2314 	    lim_cur(td, RLIMIT_FSIZE)) {
2315 		PROC_LOCK(td->td_proc);
2316 		kern_psignal(td->td_proc, SIGXFSZ);
2317 		PROC_UNLOCK(td->td_proc);
2318 		return (EFBIG);
2319 	}
2320 	return (0);
2321 }
2322 
2323 int
2324 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2325     struct thread *td)
2326 {
2327 	struct vnode *vp;
2328 
2329 	vp = fp->f_vnode;
2330 #ifdef AUDIT
2331 	vn_lock(vp, LK_SHARED | LK_RETRY);
2332 	AUDIT_ARG_VNODE1(vp);
2333 	VOP_UNLOCK(vp);
2334 #endif
2335 	return (setfmode(td, active_cred, vp, mode));
2336 }
2337 
2338 int
2339 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2340     struct thread *td)
2341 {
2342 	struct vnode *vp;
2343 
2344 	vp = fp->f_vnode;
2345 #ifdef AUDIT
2346 	vn_lock(vp, LK_SHARED | LK_RETRY);
2347 	AUDIT_ARG_VNODE1(vp);
2348 	VOP_UNLOCK(vp);
2349 #endif
2350 	return (setfown(td, active_cred, vp, uid, gid));
2351 }
2352 
2353 void
2354 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2355 {
2356 	vm_object_t object;
2357 
2358 	if ((object = vp->v_object) == NULL)
2359 		return;
2360 	VM_OBJECT_WLOCK(object);
2361 	vm_object_page_remove(object, start, end, 0);
2362 	VM_OBJECT_WUNLOCK(object);
2363 }
2364 
2365 int
2366 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2367 {
2368 	struct vattr va;
2369 	daddr_t bn, bnp;
2370 	uint64_t bsize;
2371 	off_t noff;
2372 	int error;
2373 
2374 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2375 	    ("Wrong command %lu", cmd));
2376 
2377 	if (vn_lock(vp, LK_SHARED) != 0)
2378 		return (EBADF);
2379 	if (vp->v_type != VREG) {
2380 		error = ENOTTY;
2381 		goto unlock;
2382 	}
2383 	error = VOP_GETATTR(vp, &va, cred);
2384 	if (error != 0)
2385 		goto unlock;
2386 	noff = *off;
2387 	if (noff >= va.va_size) {
2388 		error = ENXIO;
2389 		goto unlock;
2390 	}
2391 	bsize = vp->v_mount->mnt_stat.f_iosize;
2392 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2393 	    noff % bsize) {
2394 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2395 		if (error == EOPNOTSUPP) {
2396 			error = ENOTTY;
2397 			goto unlock;
2398 		}
2399 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2400 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2401 			noff = bn * bsize;
2402 			if (noff < *off)
2403 				noff = *off;
2404 			goto unlock;
2405 		}
2406 	}
2407 	if (noff > va.va_size)
2408 		noff = va.va_size;
2409 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2410 	if (cmd == FIOSEEKDATA)
2411 		error = ENXIO;
2412 unlock:
2413 	VOP_UNLOCK(vp);
2414 	if (error == 0)
2415 		*off = noff;
2416 	return (error);
2417 }
2418 
2419 int
2420 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2421 {
2422 	struct ucred *cred;
2423 	struct vnode *vp;
2424 	struct vattr vattr;
2425 	off_t foffset, size;
2426 	int error, noneg;
2427 
2428 	cred = td->td_ucred;
2429 	vp = fp->f_vnode;
2430 	foffset = foffset_lock(fp, 0);
2431 	noneg = (vp->v_type != VCHR);
2432 	error = 0;
2433 	switch (whence) {
2434 	case L_INCR:
2435 		if (noneg &&
2436 		    (foffset < 0 ||
2437 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2438 			error = EOVERFLOW;
2439 			break;
2440 		}
2441 		offset += foffset;
2442 		break;
2443 	case L_XTND:
2444 		vn_lock(vp, LK_SHARED | LK_RETRY);
2445 		error = VOP_GETATTR(vp, &vattr, cred);
2446 		VOP_UNLOCK(vp);
2447 		if (error)
2448 			break;
2449 
2450 		/*
2451 		 * If the file references a disk device, then fetch
2452 		 * the media size and use that to determine the ending
2453 		 * offset.
2454 		 */
2455 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2456 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2457 			vattr.va_size = size;
2458 		if (noneg &&
2459 		    (vattr.va_size > OFF_MAX ||
2460 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2461 			error = EOVERFLOW;
2462 			break;
2463 		}
2464 		offset += vattr.va_size;
2465 		break;
2466 	case L_SET:
2467 		break;
2468 	case SEEK_DATA:
2469 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2470 		if (error == ENOTTY)
2471 			error = EINVAL;
2472 		break;
2473 	case SEEK_HOLE:
2474 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2475 		if (error == ENOTTY)
2476 			error = EINVAL;
2477 		break;
2478 	default:
2479 		error = EINVAL;
2480 	}
2481 	if (error == 0 && noneg && offset < 0)
2482 		error = EINVAL;
2483 	if (error != 0)
2484 		goto drop;
2485 	VFS_KNOTE_UNLOCKED(vp, 0);
2486 	td->td_uretoff.tdu_off = offset;
2487 drop:
2488 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2489 	return (error);
2490 }
2491 
2492 int
2493 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2494     struct thread *td)
2495 {
2496 	int error;
2497 
2498 	/*
2499 	 * Grant permission if the caller is the owner of the file, or
2500 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2501 	 * on the file.  If the time pointer is null, then write
2502 	 * permission on the file is also sufficient.
2503 	 *
2504 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2505 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2506 	 * will be allowed to set the times [..] to the current
2507 	 * server time.
2508 	 */
2509 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2510 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2511 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2512 	return (error);
2513 }
2514 
2515 int
2516 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2517 {
2518 	struct vnode *vp;
2519 	int error;
2520 
2521 	if (fp->f_type == DTYPE_FIFO)
2522 		kif->kf_type = KF_TYPE_FIFO;
2523 	else
2524 		kif->kf_type = KF_TYPE_VNODE;
2525 	vp = fp->f_vnode;
2526 	vref(vp);
2527 	FILEDESC_SUNLOCK(fdp);
2528 	error = vn_fill_kinfo_vnode(vp, kif);
2529 	vrele(vp);
2530 	FILEDESC_SLOCK(fdp);
2531 	return (error);
2532 }
2533 
2534 static inline void
2535 vn_fill_junk(struct kinfo_file *kif)
2536 {
2537 	size_t len, olen;
2538 
2539 	/*
2540 	 * Simulate vn_fullpath returning changing values for a given
2541 	 * vp during e.g. coredump.
2542 	 */
2543 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2544 	olen = strlen(kif->kf_path);
2545 	if (len < olen)
2546 		strcpy(&kif->kf_path[len - 1], "$");
2547 	else
2548 		for (; olen < len; olen++)
2549 			strcpy(&kif->kf_path[olen], "A");
2550 }
2551 
2552 int
2553 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2554 {
2555 	struct vattr va;
2556 	char *fullpath, *freepath;
2557 	int error;
2558 
2559 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2560 	freepath = NULL;
2561 	fullpath = "-";
2562 	error = vn_fullpath(vp, &fullpath, &freepath);
2563 	if (error == 0) {
2564 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2565 	}
2566 	if (freepath != NULL)
2567 		free(freepath, M_TEMP);
2568 
2569 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2570 		vn_fill_junk(kif);
2571 	);
2572 
2573 	/*
2574 	 * Retrieve vnode attributes.
2575 	 */
2576 	va.va_fsid = VNOVAL;
2577 	va.va_rdev = NODEV;
2578 	vn_lock(vp, LK_SHARED | LK_RETRY);
2579 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2580 	VOP_UNLOCK(vp);
2581 	if (error != 0)
2582 		return (error);
2583 	if (va.va_fsid != VNOVAL)
2584 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2585 	else
2586 		kif->kf_un.kf_file.kf_file_fsid =
2587 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2588 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2589 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2590 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2591 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2592 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2593 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2594 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2595 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2596 	return (0);
2597 }
2598 
2599 int
2600 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2601     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2602     struct thread *td)
2603 {
2604 #ifdef HWPMC_HOOKS
2605 	struct pmckern_map_in pkm;
2606 #endif
2607 	struct mount *mp;
2608 	struct vnode *vp;
2609 	vm_object_t object;
2610 	vm_prot_t maxprot;
2611 	boolean_t writecounted;
2612 	int error;
2613 
2614 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2615     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2616 	/*
2617 	 * POSIX shared-memory objects are defined to have
2618 	 * kernel persistence, and are not defined to support
2619 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2620 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2621 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2622 	 * flag to request this behavior.
2623 	 */
2624 	if ((fp->f_flag & FPOSIXSHM) != 0)
2625 		flags |= MAP_NOSYNC;
2626 #endif
2627 	vp = fp->f_vnode;
2628 
2629 	/*
2630 	 * Ensure that file and memory protections are
2631 	 * compatible.  Note that we only worry about
2632 	 * writability if mapping is shared; in this case,
2633 	 * current and max prot are dictated by the open file.
2634 	 * XXX use the vnode instead?  Problem is: what
2635 	 * credentials do we use for determination? What if
2636 	 * proc does a setuid?
2637 	 */
2638 	mp = vp->v_mount;
2639 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2640 		maxprot = VM_PROT_NONE;
2641 		if ((prot & VM_PROT_EXECUTE) != 0)
2642 			return (EACCES);
2643 	} else
2644 		maxprot = VM_PROT_EXECUTE;
2645 	if ((fp->f_flag & FREAD) != 0)
2646 		maxprot |= VM_PROT_READ;
2647 	else if ((prot & VM_PROT_READ) != 0)
2648 		return (EACCES);
2649 
2650 	/*
2651 	 * If we are sharing potential changes via MAP_SHARED and we
2652 	 * are trying to get write permission although we opened it
2653 	 * without asking for it, bail out.
2654 	 */
2655 	if ((flags & MAP_SHARED) != 0) {
2656 		if ((fp->f_flag & FWRITE) != 0)
2657 			maxprot |= VM_PROT_WRITE;
2658 		else if ((prot & VM_PROT_WRITE) != 0)
2659 			return (EACCES);
2660 	} else {
2661 		maxprot |= VM_PROT_WRITE;
2662 		cap_maxprot |= VM_PROT_WRITE;
2663 	}
2664 	maxprot &= cap_maxprot;
2665 
2666 	/*
2667 	 * For regular files and shared memory, POSIX requires that
2668 	 * the value of foff be a legitimate offset within the data
2669 	 * object.  In particular, negative offsets are invalid.
2670 	 * Blocking negative offsets and overflows here avoids
2671 	 * possible wraparound or user-level access into reserved
2672 	 * ranges of the data object later.  In contrast, POSIX does
2673 	 * not dictate how offsets are used by device drivers, so in
2674 	 * the case of a device mapping a negative offset is passed
2675 	 * on.
2676 	 */
2677 	if (
2678 #ifdef _LP64
2679 	    size > OFF_MAX ||
2680 #endif
2681 	    foff > OFF_MAX - size)
2682 		return (EINVAL);
2683 
2684 	writecounted = FALSE;
2685 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2686 	    &foff, &object, &writecounted);
2687 	if (error != 0)
2688 		return (error);
2689 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2690 	    foff, writecounted, td);
2691 	if (error != 0) {
2692 		/*
2693 		 * If this mapping was accounted for in the vnode's
2694 		 * writecount, then undo that now.
2695 		 */
2696 		if (writecounted)
2697 			vm_pager_release_writecount(object, 0, size);
2698 		vm_object_deallocate(object);
2699 	}
2700 #ifdef HWPMC_HOOKS
2701 	/* Inform hwpmc(4) if an executable is being mapped. */
2702 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2703 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2704 			pkm.pm_file = vp;
2705 			pkm.pm_address = (uintptr_t) *addr;
2706 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2707 		}
2708 	}
2709 #endif
2710 	return (error);
2711 }
2712 
2713 void
2714 vn_fsid(struct vnode *vp, struct vattr *va)
2715 {
2716 	fsid_t *f;
2717 
2718 	f = &vp->v_mount->mnt_stat.f_fsid;
2719 	va->va_fsid = (uint32_t)f->val[1];
2720 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2721 	va->va_fsid += (uint32_t)f->val[0];
2722 }
2723 
2724 int
2725 vn_fsync_buf(struct vnode *vp, int waitfor)
2726 {
2727 	struct buf *bp, *nbp;
2728 	struct bufobj *bo;
2729 	struct mount *mp;
2730 	int error, maxretry;
2731 
2732 	error = 0;
2733 	maxretry = 10000;     /* large, arbitrarily chosen */
2734 	mp = NULL;
2735 	if (vp->v_type == VCHR) {
2736 		VI_LOCK(vp);
2737 		mp = vp->v_rdev->si_mountpt;
2738 		VI_UNLOCK(vp);
2739 	}
2740 	bo = &vp->v_bufobj;
2741 	BO_LOCK(bo);
2742 loop1:
2743 	/*
2744 	 * MARK/SCAN initialization to avoid infinite loops.
2745 	 */
2746         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2747 		bp->b_vflags &= ~BV_SCANNED;
2748 		bp->b_error = 0;
2749 	}
2750 
2751 	/*
2752 	 * Flush all dirty buffers associated with a vnode.
2753 	 */
2754 loop2:
2755 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2756 		if ((bp->b_vflags & BV_SCANNED) != 0)
2757 			continue;
2758 		bp->b_vflags |= BV_SCANNED;
2759 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2760 			if (waitfor != MNT_WAIT)
2761 				continue;
2762 			if (BUF_LOCK(bp,
2763 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2764 			    BO_LOCKPTR(bo)) != 0) {
2765 				BO_LOCK(bo);
2766 				goto loop1;
2767 			}
2768 			BO_LOCK(bo);
2769 		}
2770 		BO_UNLOCK(bo);
2771 		KASSERT(bp->b_bufobj == bo,
2772 		    ("bp %p wrong b_bufobj %p should be %p",
2773 		    bp, bp->b_bufobj, bo));
2774 		if ((bp->b_flags & B_DELWRI) == 0)
2775 			panic("fsync: not dirty");
2776 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2777 			vfs_bio_awrite(bp);
2778 		} else {
2779 			bremfree(bp);
2780 			bawrite(bp);
2781 		}
2782 		if (maxretry < 1000)
2783 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2784 		BO_LOCK(bo);
2785 		goto loop2;
2786 	}
2787 
2788 	/*
2789 	 * If synchronous the caller expects us to completely resolve all
2790 	 * dirty buffers in the system.  Wait for in-progress I/O to
2791 	 * complete (which could include background bitmap writes), then
2792 	 * retry if dirty blocks still exist.
2793 	 */
2794 	if (waitfor == MNT_WAIT) {
2795 		bufobj_wwait(bo, 0, 0);
2796 		if (bo->bo_dirty.bv_cnt > 0) {
2797 			/*
2798 			 * If we are unable to write any of these buffers
2799 			 * then we fail now rather than trying endlessly
2800 			 * to write them out.
2801 			 */
2802 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2803 				if ((error = bp->b_error) != 0)
2804 					break;
2805 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2806 			    (error == 0 && --maxretry >= 0))
2807 				goto loop1;
2808 			if (error == 0)
2809 				error = EAGAIN;
2810 		}
2811 	}
2812 	BO_UNLOCK(bo);
2813 	if (error != 0)
2814 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2815 
2816 	return (error);
2817 }
2818 
2819 /*
2820  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2821  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2822  * to do the actual copy.
2823  * vn_generic_copy_file_range() is factored out, so it can be called
2824  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2825  * different file systems.
2826  */
2827 int
2828 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2829     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2830     struct ucred *outcred, struct thread *fsize_td)
2831 {
2832 	int error;
2833 	size_t len;
2834 	uint64_t uval;
2835 
2836 	len = *lenp;
2837 	*lenp = 0;		/* For error returns. */
2838 	error = 0;
2839 
2840 	/* Do some sanity checks on the arguments. */
2841 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2842 		error = EISDIR;
2843 	else if (*inoffp < 0 || *outoffp < 0 ||
2844 	    invp->v_type != VREG || outvp->v_type != VREG)
2845 		error = EINVAL;
2846 	if (error != 0)
2847 		goto out;
2848 
2849 	/* Ensure offset + len does not wrap around. */
2850 	uval = *inoffp;
2851 	uval += len;
2852 	if (uval > INT64_MAX)
2853 		len = INT64_MAX - *inoffp;
2854 	uval = *outoffp;
2855 	uval += len;
2856 	if (uval > INT64_MAX)
2857 		len = INT64_MAX - *outoffp;
2858 	if (len == 0)
2859 		goto out;
2860 
2861 	/*
2862 	 * If the two vnode are for the same file system, call
2863 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2864 	 * which can handle copies across multiple file systems.
2865 	 */
2866 	*lenp = len;
2867 	if (invp->v_mount == outvp->v_mount)
2868 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2869 		    lenp, flags, incred, outcred, fsize_td);
2870 	else
2871 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2872 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2873 out:
2874 	return (error);
2875 }
2876 
2877 /*
2878  * Test len bytes of data starting at dat for all bytes == 0.
2879  * Return true if all bytes are zero, false otherwise.
2880  * Expects dat to be well aligned.
2881  */
2882 static bool
2883 mem_iszero(void *dat, int len)
2884 {
2885 	int i;
2886 	const u_int *p;
2887 	const char *cp;
2888 
2889 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2890 		if (len >= sizeof(*p)) {
2891 			if (*p != 0)
2892 				return (false);
2893 		} else {
2894 			cp = (const char *)p;
2895 			for (i = 0; i < len; i++, cp++)
2896 				if (*cp != '\0')
2897 					return (false);
2898 		}
2899 	}
2900 	return (true);
2901 }
2902 
2903 /*
2904  * Look for a hole in the output file and, if found, adjust *outoffp
2905  * and *xferp to skip past the hole.
2906  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2907  * to be written as 0's upon return.
2908  */
2909 static off_t
2910 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2911     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2912 {
2913 	int error;
2914 	off_t delta;
2915 
2916 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2917 		*dataoffp = *outoffp;
2918 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2919 		    curthread);
2920 		if (error == 0) {
2921 			*holeoffp = *dataoffp;
2922 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2923 			    curthread);
2924 		}
2925 		if (error != 0 || *holeoffp == *dataoffp) {
2926 			/*
2927 			 * Since outvp is unlocked, it may be possible for
2928 			 * another thread to do a truncate(), lseek(), write()
2929 			 * creating a hole at startoff between the above
2930 			 * VOP_IOCTL() calls, if the other thread does not do
2931 			 * rangelocking.
2932 			 * If that happens, *holeoffp == *dataoffp and finding
2933 			 * the hole has failed, so disable vn_skip_hole().
2934 			 */
2935 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2936 			return (xfer2);
2937 		}
2938 		KASSERT(*dataoffp >= *outoffp,
2939 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2940 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2941 		KASSERT(*holeoffp > *dataoffp,
2942 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2943 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2944 	}
2945 
2946 	/*
2947 	 * If there is a hole before the data starts, advance *outoffp and
2948 	 * *xferp past the hole.
2949 	 */
2950 	if (*dataoffp > *outoffp) {
2951 		delta = *dataoffp - *outoffp;
2952 		if (delta >= *xferp) {
2953 			/* Entire *xferp is a hole. */
2954 			*outoffp += *xferp;
2955 			*xferp = 0;
2956 			return (0);
2957 		}
2958 		*xferp -= delta;
2959 		*outoffp += delta;
2960 		xfer2 = MIN(xfer2, *xferp);
2961 	}
2962 
2963 	/*
2964 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2965 	 * that the write ends at the start of the hole.
2966 	 * *holeoffp should always be greater than *outoffp, but for the
2967 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2968 	 * value.
2969 	 */
2970 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2971 		xfer2 = *holeoffp - *outoffp;
2972 	return (xfer2);
2973 }
2974 
2975 /*
2976  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2977  * dat is a maximum of blksize in length and can be written repeatedly in
2978  * the chunk.
2979  * If growfile == true, just grow the file via vn_truncate_locked() instead
2980  * of doing actual writes.
2981  * If checkhole == true, a hole is being punched, so skip over any hole
2982  * already in the output file.
2983  */
2984 static int
2985 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2986     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2987 {
2988 	struct mount *mp;
2989 	off_t dataoff, holeoff, xfer2;
2990 	int error, lckf;
2991 
2992 	/*
2993 	 * Loop around doing writes of blksize until write has been completed.
2994 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2995 	 * done for each iteration, since the xfer argument can be very
2996 	 * large if there is a large hole to punch in the output file.
2997 	 */
2998 	error = 0;
2999 	holeoff = 0;
3000 	do {
3001 		xfer2 = MIN(xfer, blksize);
3002 		if (checkhole) {
3003 			/*
3004 			 * Punching a hole.  Skip writing if there is
3005 			 * already a hole in the output file.
3006 			 */
3007 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3008 			    &dataoff, &holeoff, cred);
3009 			if (xfer == 0)
3010 				break;
3011 			if (holeoff < 0)
3012 				checkhole = false;
3013 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3014 			    (intmax_t)xfer2));
3015 		}
3016 		bwillwrite();
3017 		mp = NULL;
3018 		error = vn_start_write(outvp, &mp, V_WAIT);
3019 		if (error != 0)
3020 			break;
3021 		if (growfile) {
3022 			error = vn_lock(outvp, LK_EXCLUSIVE);
3023 			if (error == 0) {
3024 				error = vn_truncate_locked(outvp, outoff + xfer,
3025 				    false, cred);
3026 				VOP_UNLOCK(outvp);
3027 			}
3028 		} else {
3029 			if (MNT_SHARED_WRITES(mp))
3030 				lckf = LK_SHARED;
3031 			else
3032 				lckf = LK_EXCLUSIVE;
3033 			error = vn_lock(outvp, lckf);
3034 			if (error == 0) {
3035 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3036 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3037 				    curthread->td_ucred, cred, NULL, curthread);
3038 				outoff += xfer2;
3039 				xfer -= xfer2;
3040 				VOP_UNLOCK(outvp);
3041 			}
3042 		}
3043 		if (mp != NULL)
3044 			vn_finished_write(mp);
3045 	} while (!growfile && xfer > 0 && error == 0);
3046 	return (error);
3047 }
3048 
3049 /*
3050  * Copy a byte range of one file to another.  This function can handle the
3051  * case where invp and outvp are on different file systems.
3052  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3053  * is no better file system specific way to do it.
3054  */
3055 int
3056 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3057     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3058     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3059 {
3060 	struct vattr va;
3061 	struct mount *mp;
3062 	struct uio io;
3063 	off_t startoff, endoff, xfer, xfer2;
3064 	u_long blksize;
3065 	int error, interrupted;
3066 	bool cantseek, readzeros, eof, lastblock;
3067 	ssize_t aresid;
3068 	size_t copylen, len, rem, savlen;
3069 	char *dat;
3070 	long holein, holeout;
3071 
3072 	holein = holeout = 0;
3073 	savlen = len = *lenp;
3074 	error = 0;
3075 	interrupted = 0;
3076 	dat = NULL;
3077 
3078 	error = vn_lock(invp, LK_SHARED);
3079 	if (error != 0)
3080 		goto out;
3081 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3082 		holein = 0;
3083 	VOP_UNLOCK(invp);
3084 
3085 	mp = NULL;
3086 	error = vn_start_write(outvp, &mp, V_WAIT);
3087 	if (error == 0)
3088 		error = vn_lock(outvp, LK_EXCLUSIVE);
3089 	if (error == 0) {
3090 		/*
3091 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3092 		 * now that outvp is locked.
3093 		 */
3094 		if (fsize_td != NULL) {
3095 			io.uio_offset = *outoffp;
3096 			io.uio_resid = len;
3097 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3098 			if (error != 0)
3099 				error = EFBIG;
3100 		}
3101 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3102 			holeout = 0;
3103 		/*
3104 		 * Holes that are past EOF do not need to be written as a block
3105 		 * of zero bytes.  So, truncate the output file as far as
3106 		 * possible and then use va.va_size to decide if writing 0
3107 		 * bytes is necessary in the loop below.
3108 		 */
3109 		if (error == 0)
3110 			error = VOP_GETATTR(outvp, &va, outcred);
3111 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3112 		    *outoffp + len) {
3113 #ifdef MAC
3114 			error = mac_vnode_check_write(curthread->td_ucred,
3115 			    outcred, outvp);
3116 			if (error == 0)
3117 #endif
3118 				error = vn_truncate_locked(outvp, *outoffp,
3119 				    false, outcred);
3120 			if (error == 0)
3121 				va.va_size = *outoffp;
3122 		}
3123 		VOP_UNLOCK(outvp);
3124 	}
3125 	if (mp != NULL)
3126 		vn_finished_write(mp);
3127 	if (error != 0)
3128 		goto out;
3129 
3130 	/*
3131 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3132 	 * If hole sizes aren't available, set the blksize to the larger
3133 	 * f_iosize of invp and outvp.
3134 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3135 	 * This value is clipped at 4Kbytes and 1Mbyte.
3136 	 */
3137 	blksize = MAX(holein, holeout);
3138 
3139 	/* Clip len to end at an exact multiple of hole size. */
3140 	if (blksize > 1) {
3141 		rem = *inoffp % blksize;
3142 		if (rem > 0)
3143 			rem = blksize - rem;
3144 		if (len > rem && len - rem > blksize)
3145 			len = savlen = rounddown(len - rem, blksize) + rem;
3146 	}
3147 
3148 	if (blksize <= 1)
3149 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3150 		    outvp->v_mount->mnt_stat.f_iosize);
3151 	if (blksize < 4096)
3152 		blksize = 4096;
3153 	else if (blksize > 1024 * 1024)
3154 		blksize = 1024 * 1024;
3155 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3156 
3157 	/*
3158 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3159 	 * to find holes.  Otherwise, just scan the read block for all 0s
3160 	 * in the inner loop where the data copying is done.
3161 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3162 	 * support holes on the server, but do not support FIOSEEKHOLE.
3163 	 */
3164 	eof = false;
3165 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3166 		endoff = 0;			/* To shut up compilers. */
3167 		cantseek = true;
3168 		startoff = *inoffp;
3169 		copylen = len;
3170 
3171 		/*
3172 		 * Find the next data area.  If there is just a hole to EOF,
3173 		 * FIOSEEKDATA should fail and then we drop down into the
3174 		 * inner loop and create the hole on the outvp file.
3175 		 * (I do not know if any file system will report a hole to
3176 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3177 		 *  will fail for those file systems.)
3178 		 *
3179 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3180 		 * the code just falls through to the inner copy loop.
3181 		 */
3182 		error = EINVAL;
3183 		if (holein > 0)
3184 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3185 			    incred, curthread);
3186 		if (error == 0) {
3187 			endoff = startoff;
3188 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3189 			    incred, curthread);
3190 			/*
3191 			 * Since invp is unlocked, it may be possible for
3192 			 * another thread to do a truncate(), lseek(), write()
3193 			 * creating a hole at startoff between the above
3194 			 * VOP_IOCTL() calls, if the other thread does not do
3195 			 * rangelocking.
3196 			 * If that happens, startoff == endoff and finding
3197 			 * the hole has failed, so set an error.
3198 			 */
3199 			if (error == 0 && startoff == endoff)
3200 				error = EINVAL; /* Any error. Reset to 0. */
3201 		}
3202 		if (error == 0) {
3203 			if (startoff > *inoffp) {
3204 				/* Found hole before data block. */
3205 				xfer = MIN(startoff - *inoffp, len);
3206 				if (*outoffp < va.va_size) {
3207 					/* Must write 0s to punch hole. */
3208 					xfer2 = MIN(va.va_size - *outoffp,
3209 					    xfer);
3210 					memset(dat, 0, MIN(xfer2, blksize));
3211 					error = vn_write_outvp(outvp, dat,
3212 					    *outoffp, xfer2, blksize, false,
3213 					    holeout > 0, outcred);
3214 				}
3215 
3216 				if (error == 0 && *outoffp + xfer >
3217 				    va.va_size && xfer == len)
3218 					/* Grow last block. */
3219 					error = vn_write_outvp(outvp, dat,
3220 					    *outoffp, xfer, blksize, true,
3221 					    false, outcred);
3222 				if (error == 0) {
3223 					*inoffp += xfer;
3224 					*outoffp += xfer;
3225 					len -= xfer;
3226 					if (len < savlen)
3227 						interrupted = sig_intr();
3228 				}
3229 			}
3230 			copylen = MIN(len, endoff - startoff);
3231 			cantseek = false;
3232 		} else {
3233 			cantseek = true;
3234 			startoff = *inoffp;
3235 			copylen = len;
3236 			error = 0;
3237 		}
3238 
3239 		xfer = blksize;
3240 		if (cantseek) {
3241 			/*
3242 			 * Set first xfer to end at a block boundary, so that
3243 			 * holes are more likely detected in the loop below via
3244 			 * the for all bytes 0 method.
3245 			 */
3246 			xfer -= (*inoffp % blksize);
3247 		}
3248 		/* Loop copying the data block. */
3249 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3250 			if (copylen < xfer)
3251 				xfer = copylen;
3252 			error = vn_lock(invp, LK_SHARED);
3253 			if (error != 0)
3254 				goto out;
3255 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3256 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3257 			    curthread->td_ucred, incred, &aresid,
3258 			    curthread);
3259 			VOP_UNLOCK(invp);
3260 			lastblock = false;
3261 			if (error == 0 && aresid > 0) {
3262 				/* Stop the copy at EOF on the input file. */
3263 				xfer -= aresid;
3264 				eof = true;
3265 				lastblock = true;
3266 			}
3267 			if (error == 0) {
3268 				/*
3269 				 * Skip the write for holes past the initial EOF
3270 				 * of the output file, unless this is the last
3271 				 * write of the output file at EOF.
3272 				 */
3273 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3274 				    false;
3275 				if (xfer == len)
3276 					lastblock = true;
3277 				if (!cantseek || *outoffp < va.va_size ||
3278 				    lastblock || !readzeros)
3279 					error = vn_write_outvp(outvp, dat,
3280 					    *outoffp, xfer, blksize,
3281 					    readzeros && lastblock &&
3282 					    *outoffp >= va.va_size, false,
3283 					    outcred);
3284 				if (error == 0) {
3285 					*inoffp += xfer;
3286 					startoff += xfer;
3287 					*outoffp += xfer;
3288 					copylen -= xfer;
3289 					len -= xfer;
3290 					if (len < savlen)
3291 						interrupted = sig_intr();
3292 				}
3293 			}
3294 			xfer = blksize;
3295 		}
3296 	}
3297 out:
3298 	*lenp = savlen - len;
3299 	free(dat, M_TEMP);
3300 	return (error);
3301 }
3302 
3303 static int
3304 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3305 {
3306 	struct mount *mp;
3307 	struct vnode *vp;
3308 	off_t olen, ooffset;
3309 	int error;
3310 #ifdef AUDIT
3311 	int audited_vnode1 = 0;
3312 #endif
3313 
3314 	vp = fp->f_vnode;
3315 	if (vp->v_type != VREG)
3316 		return (ENODEV);
3317 
3318 	/* Allocating blocks may take a long time, so iterate. */
3319 	for (;;) {
3320 		olen = len;
3321 		ooffset = offset;
3322 
3323 		bwillwrite();
3324 		mp = NULL;
3325 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3326 		if (error != 0)
3327 			break;
3328 		error = vn_lock(vp, LK_EXCLUSIVE);
3329 		if (error != 0) {
3330 			vn_finished_write(mp);
3331 			break;
3332 		}
3333 #ifdef AUDIT
3334 		if (!audited_vnode1) {
3335 			AUDIT_ARG_VNODE1(vp);
3336 			audited_vnode1 = 1;
3337 		}
3338 #endif
3339 #ifdef MAC
3340 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3341 		if (error == 0)
3342 #endif
3343 			error = VOP_ALLOCATE(vp, &offset, &len);
3344 		VOP_UNLOCK(vp);
3345 		vn_finished_write(mp);
3346 
3347 		if (olen + ooffset != offset + len) {
3348 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3349 			    ooffset, olen, offset, len);
3350 		}
3351 		if (error != 0 || len == 0)
3352 			break;
3353 		KASSERT(olen > len, ("Iteration did not make progress?"));
3354 		maybe_yield();
3355 	}
3356 
3357 	return (error);
3358 }
3359 
3360 static u_long vn_lock_pair_pause_cnt;
3361 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3362     &vn_lock_pair_pause_cnt, 0,
3363     "Count of vn_lock_pair deadlocks");
3364 
3365 u_int vn_lock_pair_pause_max;
3366 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3367     &vn_lock_pair_pause_max, 0,
3368     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3369 
3370 static void
3371 vn_lock_pair_pause(const char *wmesg)
3372 {
3373 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3374 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3375 }
3376 
3377 /*
3378  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3379  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3380  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3381  * can be NULL.
3382  *
3383  * The function returns with both vnodes exclusively locked, and
3384  * guarantees that it does not create lock order reversal with other
3385  * threads during its execution.  Both vnodes could be unlocked
3386  * temporary (and reclaimed).
3387  */
3388 void
3389 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3390     bool vp2_locked)
3391 {
3392 	int error;
3393 
3394 	if (vp1 == NULL && vp2 == NULL)
3395 		return;
3396 	if (vp1 != NULL) {
3397 		if (vp1_locked)
3398 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3399 		else
3400 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3401 	} else {
3402 		vp1_locked = true;
3403 	}
3404 	if (vp2 != NULL) {
3405 		if (vp2_locked)
3406 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3407 		else
3408 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3409 	} else {
3410 		vp2_locked = true;
3411 	}
3412 	if (!vp1_locked && !vp2_locked) {
3413 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3414 		vp1_locked = true;
3415 	}
3416 
3417 	for (;;) {
3418 		if (vp1_locked && vp2_locked)
3419 			break;
3420 		if (vp1_locked && vp2 != NULL) {
3421 			if (vp1 != NULL) {
3422 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3423 				    __FILE__, __LINE__);
3424 				if (error == 0)
3425 					break;
3426 				VOP_UNLOCK(vp1);
3427 				vp1_locked = false;
3428 				vn_lock_pair_pause("vlp1");
3429 			}
3430 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3431 			vp2_locked = true;
3432 		}
3433 		if (vp2_locked && vp1 != NULL) {
3434 			if (vp2 != NULL) {
3435 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3436 				    __FILE__, __LINE__);
3437 				if (error == 0)
3438 					break;
3439 				VOP_UNLOCK(vp2);
3440 				vp2_locked = false;
3441 				vn_lock_pair_pause("vlp2");
3442 			}
3443 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3444 			vp1_locked = true;
3445 		}
3446 	}
3447 	if (vp1 != NULL)
3448 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3449 	if (vp2 != NULL)
3450 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3451 }
3452