xref: /freebsd/sys/kern/vfs_vnops.c (revision 88e852c0b5c872b1a3234515623104ae61b60773)
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_BENEATH) != 0)
203 		res |= BENEATH;
204 	if ((fmode & O_RESOLVE_BENEATH) != 0)
205 		res |= RBENEATH;
206 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
207 		res |= AUDITVNODE1;
208 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
209 		res |= NOCAPCHECK;
210 	return (res);
211 }
212 
213 /*
214  * Common code for vnode open operations via a name lookup.
215  * Lookup the vnode and invoke VOP_CREATE if needed.
216  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
217  *
218  * Note that this does NOT free nameidata for the successful case,
219  * due to the NDINIT being done elsewhere.
220  */
221 int
222 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
223     struct ucred *cred, struct file *fp)
224 {
225 	struct vnode *vp;
226 	struct mount *mp;
227 	struct thread *td = ndp->ni_cnd.cn_thread;
228 	struct vattr vat;
229 	struct vattr *vap = &vat;
230 	int fmode, error;
231 
232 restart:
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 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &ndp->ni_vp :
279 			    NULL, false);
280 			vn_finished_write(mp);
281 			if (error) {
282 				NDFREE(ndp, NDF_ONLY_PNBUF);
283 				if (error == ERELOOKUP) {
284 					NDREINIT(ndp);
285 					goto restart;
286 				}
287 				return (error);
288 			}
289 			fmode &= ~O_TRUNC;
290 			vp = ndp->ni_vp;
291 		} else {
292 			if (ndp->ni_dvp == ndp->ni_vp)
293 				vrele(ndp->ni_dvp);
294 			else
295 				vput(ndp->ni_dvp);
296 			ndp->ni_dvp = NULL;
297 			vp = ndp->ni_vp;
298 			if (fmode & O_EXCL) {
299 				error = EEXIST;
300 				goto bad;
301 			}
302 			if (vp->v_type == VDIR) {
303 				error = EISDIR;
304 				goto bad;
305 			}
306 			fmode &= ~O_CREAT;
307 		}
308 	} else {
309 		ndp->ni_cnd.cn_nameiop = LOOKUP;
310 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
311 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
312 		    FOLLOW;
313 		if ((fmode & FWRITE) == 0)
314 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
315 		if ((error = namei(ndp)) != 0)
316 			return (error);
317 		vp = ndp->ni_vp;
318 	}
319 	error = vn_open_vnode(vp, fmode, cred, td, fp);
320 	if (error)
321 		goto bad;
322 	*flagp = fmode;
323 	return (0);
324 bad:
325 	NDFREE(ndp, NDF_ONLY_PNBUF);
326 	vput(vp);
327 	*flagp = fmode;
328 	ndp->ni_vp = NULL;
329 	return (error);
330 }
331 
332 static int
333 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
334 {
335 	struct flock lf;
336 	int error, lock_flags, type;
337 
338 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
339 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
340 		return (0);
341 	KASSERT(fp != NULL, ("open with flock requires fp"));
342 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
343 		return (EOPNOTSUPP);
344 
345 	lock_flags = VOP_ISLOCKED(vp);
346 	VOP_UNLOCK(vp);
347 
348 	lf.l_whence = SEEK_SET;
349 	lf.l_start = 0;
350 	lf.l_len = 0;
351 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
352 	type = F_FLOCK;
353 	if ((fmode & FNONBLOCK) == 0)
354 		type |= F_WAIT;
355 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
356 	if (error == 0)
357 		fp->f_flag |= FHASLOCK;
358 
359 	vn_lock(vp, lock_flags | LK_RETRY);
360 	return (error);
361 }
362 
363 /*
364  * Common code for vnode open operations once a vnode is located.
365  * Check permissions, and call the VOP_OPEN routine.
366  */
367 int
368 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
369     struct thread *td, struct file *fp)
370 {
371 	accmode_t accmode;
372 	int error;
373 
374 	if (vp->v_type == VLNK)
375 		return (EMLINK);
376 	if (vp->v_type == VSOCK)
377 		return (EOPNOTSUPP);
378 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
379 		return (ENOTDIR);
380 	accmode = 0;
381 	if (fmode & (FWRITE | O_TRUNC)) {
382 		if (vp->v_type == VDIR)
383 			return (EISDIR);
384 		accmode |= VWRITE;
385 	}
386 	if (fmode & FREAD)
387 		accmode |= VREAD;
388 	if (fmode & FEXEC)
389 		accmode |= VEXEC;
390 	if ((fmode & O_APPEND) && (fmode & FWRITE))
391 		accmode |= VAPPEND;
392 #ifdef MAC
393 	if (fmode & O_CREAT)
394 		accmode |= VCREAT;
395 	if (fmode & O_VERIFY)
396 		accmode |= VVERIFY;
397 	error = mac_vnode_check_open(cred, vp, accmode);
398 	if (error)
399 		return (error);
400 
401 	accmode &= ~(VCREAT | VVERIFY);
402 #endif
403 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
404 		error = VOP_ACCESS(vp, accmode, cred, td);
405 		if (error != 0)
406 			return (error);
407 	}
408 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
409 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
410 	error = VOP_OPEN(vp, fmode, cred, td, fp);
411 	if (error != 0)
412 		return (error);
413 
414 	error = vn_open_vnode_advlock(vp, fmode, fp);
415 	if (error == 0 && (fmode & FWRITE) != 0) {
416 		error = VOP_ADD_WRITECOUNT(vp, 1);
417 		if (error == 0) {
418 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
419 			     __func__, vp, vp->v_writecount);
420 		}
421 	}
422 
423 	/*
424 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
425 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
426 	 * Arrange for that by having fdrop() to use vn_closefile().
427 	 */
428 	if (error != 0) {
429 		fp->f_flag |= FOPENFAILED;
430 		fp->f_vnode = vp;
431 		if (fp->f_ops == &badfileops) {
432 			fp->f_type = DTYPE_VNODE;
433 			fp->f_ops = &vnops;
434 		}
435 		vref(vp);
436 	}
437 
438 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
439 	return (error);
440 
441 }
442 
443 /*
444  * Check for write permissions on the specified vnode.
445  * Prototype text segments cannot be written.
446  * It is racy.
447  */
448 int
449 vn_writechk(struct vnode *vp)
450 {
451 
452 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
453 	/*
454 	 * If there's shared text associated with
455 	 * the vnode, try to free it up once.  If
456 	 * we fail, we can't allow writing.
457 	 */
458 	if (VOP_IS_TEXT(vp))
459 		return (ETXTBSY);
460 
461 	return (0);
462 }
463 
464 /*
465  * Vnode close call
466  */
467 static int
468 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
469     struct thread *td, bool keep_ref)
470 {
471 	struct mount *mp;
472 	int error, lock_flags;
473 
474 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
475 	    MNT_EXTENDED_SHARED(vp->v_mount))
476 		lock_flags = LK_SHARED;
477 	else
478 		lock_flags = LK_EXCLUSIVE;
479 
480 	vn_start_write(vp, &mp, V_WAIT);
481 	vn_lock(vp, lock_flags | LK_RETRY);
482 	AUDIT_ARG_VNODE1(vp);
483 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
484 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
485 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
486 		    __func__, vp, vp->v_writecount);
487 	}
488 	error = VOP_CLOSE(vp, flags, file_cred, td);
489 	if (keep_ref)
490 		VOP_UNLOCK(vp);
491 	else
492 		vput(vp);
493 	vn_finished_write(mp);
494 	return (error);
495 }
496 
497 int
498 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
499     struct thread *td)
500 {
501 
502 	return (vn_close1(vp, flags, file_cred, td, false));
503 }
504 
505 /*
506  * Heuristic to detect sequential operation.
507  */
508 static int
509 sequential_heuristic(struct uio *uio, struct file *fp)
510 {
511 	enum uio_rw rw;
512 
513 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
514 
515 	rw = uio->uio_rw;
516 	if (fp->f_flag & FRDAHEAD)
517 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
518 
519 	/*
520 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
521 	 * that the first I/O is normally considered to be slightly
522 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
523 	 * unless previous seeks have reduced f_seqcount to 0, in which
524 	 * case offset 0 is not special.
525 	 */
526 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
527 	    uio->uio_offset == fp->f_nextoff[rw]) {
528 		/*
529 		 * f_seqcount is in units of fixed-size blocks so that it
530 		 * depends mainly on the amount of sequential I/O and not
531 		 * much on the number of sequential I/O's.  The fixed size
532 		 * of 16384 is hard-coded here since it is (not quite) just
533 		 * a magic size that works well here.  This size is more
534 		 * closely related to the best I/O size for real disks than
535 		 * to any block size used by software.
536 		 */
537 		if (uio->uio_resid >= IO_SEQMAX * 16384)
538 			fp->f_seqcount[rw] = IO_SEQMAX;
539 		else {
540 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
541 			if (fp->f_seqcount[rw] > IO_SEQMAX)
542 				fp->f_seqcount[rw] = IO_SEQMAX;
543 		}
544 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
545 	}
546 
547 	/* Not sequential.  Quickly draw-down sequentiality. */
548 	if (fp->f_seqcount[rw] > 1)
549 		fp->f_seqcount[rw] = 1;
550 	else
551 		fp->f_seqcount[rw] = 0;
552 	return (0);
553 }
554 
555 /*
556  * Package up an I/O request on a vnode into a uio and do it.
557  */
558 int
559 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
560     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
561     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
562 {
563 	struct uio auio;
564 	struct iovec aiov;
565 	struct mount *mp;
566 	struct ucred *cred;
567 	void *rl_cookie;
568 	struct vn_io_fault_args args;
569 	int error, lock_flags;
570 
571 	if (offset < 0 && vp->v_type != VCHR)
572 		return (EINVAL);
573 	auio.uio_iov = &aiov;
574 	auio.uio_iovcnt = 1;
575 	aiov.iov_base = base;
576 	aiov.iov_len = len;
577 	auio.uio_resid = len;
578 	auio.uio_offset = offset;
579 	auio.uio_segflg = segflg;
580 	auio.uio_rw = rw;
581 	auio.uio_td = td;
582 	error = 0;
583 
584 	if ((ioflg & IO_NODELOCKED) == 0) {
585 		if ((ioflg & IO_RANGELOCKED) == 0) {
586 			if (rw == UIO_READ) {
587 				rl_cookie = vn_rangelock_rlock(vp, offset,
588 				    offset + len);
589 			} else if ((ioflg & IO_APPEND) != 0) {
590 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
591 			} else {
592 				rl_cookie = vn_rangelock_wlock(vp, offset,
593 				    offset + len);
594 			}
595 		} else
596 			rl_cookie = NULL;
597 		mp = NULL;
598 		if (rw == UIO_WRITE) {
599 			if (vp->v_type != VCHR &&
600 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
601 			    != 0)
602 				goto out;
603 			if (MNT_SHARED_WRITES(mp) ||
604 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
605 				lock_flags = LK_SHARED;
606 			else
607 				lock_flags = LK_EXCLUSIVE;
608 		} else
609 			lock_flags = LK_SHARED;
610 		vn_lock(vp, lock_flags | LK_RETRY);
611 	} else
612 		rl_cookie = NULL;
613 
614 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
615 #ifdef MAC
616 	if ((ioflg & IO_NOMACCHECK) == 0) {
617 		if (rw == UIO_READ)
618 			error = mac_vnode_check_read(active_cred, file_cred,
619 			    vp);
620 		else
621 			error = mac_vnode_check_write(active_cred, file_cred,
622 			    vp);
623 	}
624 #endif
625 	if (error == 0) {
626 		if (file_cred != NULL)
627 			cred = file_cred;
628 		else
629 			cred = active_cred;
630 		if (do_vn_io_fault(vp, &auio)) {
631 			args.kind = VN_IO_FAULT_VOP;
632 			args.cred = cred;
633 			args.flags = ioflg;
634 			args.args.vop_args.vp = vp;
635 			error = vn_io_fault1(vp, &auio, &args, td);
636 		} else if (rw == UIO_READ) {
637 			error = VOP_READ(vp, &auio, ioflg, cred);
638 		} else /* if (rw == UIO_WRITE) */ {
639 			error = VOP_WRITE(vp, &auio, ioflg, cred);
640 		}
641 	}
642 	if (aresid)
643 		*aresid = auio.uio_resid;
644 	else
645 		if (auio.uio_resid && error == 0)
646 			error = EIO;
647 	if ((ioflg & IO_NODELOCKED) == 0) {
648 		VOP_UNLOCK(vp);
649 		if (mp != NULL)
650 			vn_finished_write(mp);
651 	}
652  out:
653 	if (rl_cookie != NULL)
654 		vn_rangelock_unlock(vp, rl_cookie);
655 	return (error);
656 }
657 
658 /*
659  * Package up an I/O request on a vnode into a uio and do it.  The I/O
660  * request is split up into smaller chunks and we try to avoid saturating
661  * the buffer cache while potentially holding a vnode locked, so we
662  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
663  * to give other processes a chance to lock the vnode (either other processes
664  * core'ing the same binary, or unrelated processes scanning the directory).
665  */
666 int
667 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
668     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
669     struct ucred *file_cred, size_t *aresid, struct thread *td)
670 {
671 	int error = 0;
672 	ssize_t iaresid;
673 
674 	do {
675 		int chunk;
676 
677 		/*
678 		 * Force `offset' to a multiple of MAXBSIZE except possibly
679 		 * for the first chunk, so that filesystems only need to
680 		 * write full blocks except possibly for the first and last
681 		 * chunks.
682 		 */
683 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
684 
685 		if (chunk > len)
686 			chunk = len;
687 		if (rw != UIO_READ && vp->v_type == VREG)
688 			bwillwrite();
689 		iaresid = 0;
690 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
691 		    ioflg, active_cred, file_cred, &iaresid, td);
692 		len -= chunk;	/* aresid calc already includes length */
693 		if (error)
694 			break;
695 		offset += chunk;
696 		base = (char *)base + chunk;
697 		kern_yield(PRI_USER);
698 	} while (len);
699 	if (aresid)
700 		*aresid = len + iaresid;
701 	return (error);
702 }
703 
704 #if OFF_MAX <= LONG_MAX
705 off_t
706 foffset_lock(struct file *fp, int flags)
707 {
708 	volatile short *flagsp;
709 	off_t res;
710 	short state;
711 
712 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
713 
714 	if ((flags & FOF_NOLOCK) != 0)
715 		return (atomic_load_long(&fp->f_offset));
716 
717 	/*
718 	 * According to McKusick the vn lock was protecting f_offset here.
719 	 * It is now protected by the FOFFSET_LOCKED flag.
720 	 */
721 	flagsp = &fp->f_vnread_flags;
722 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
723 		return (atomic_load_long(&fp->f_offset));
724 
725 	sleepq_lock(&fp->f_vnread_flags);
726 	state = atomic_load_16(flagsp);
727 	for (;;) {
728 		if ((state & FOFFSET_LOCKED) == 0) {
729 			if (!atomic_fcmpset_acq_16(flagsp, &state,
730 			    FOFFSET_LOCKED))
731 				continue;
732 			break;
733 		}
734 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
735 			if (!atomic_fcmpset_acq_16(flagsp, &state,
736 			    state | FOFFSET_LOCK_WAITING))
737 				continue;
738 		}
739 		DROP_GIANT();
740 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
741 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
742 		PICKUP_GIANT();
743 		sleepq_lock(&fp->f_vnread_flags);
744 		state = atomic_load_16(flagsp);
745 	}
746 	res = atomic_load_long(&fp->f_offset);
747 	sleepq_release(&fp->f_vnread_flags);
748 	return (res);
749 }
750 
751 void
752 foffset_unlock(struct file *fp, off_t val, int flags)
753 {
754 	volatile short *flagsp;
755 	short state;
756 
757 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
758 
759 	if ((flags & FOF_NOUPDATE) == 0)
760 		atomic_store_long(&fp->f_offset, val);
761 	if ((flags & FOF_NEXTOFF_R) != 0)
762 		fp->f_nextoff[UIO_READ] = val;
763 	if ((flags & FOF_NEXTOFF_W) != 0)
764 		fp->f_nextoff[UIO_WRITE] = val;
765 
766 	if ((flags & FOF_NOLOCK) != 0)
767 		return;
768 
769 	flagsp = &fp->f_vnread_flags;
770 	state = atomic_load_16(flagsp);
771 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
772 	    atomic_cmpset_rel_16(flagsp, state, 0))
773 		return;
774 
775 	sleepq_lock(&fp->f_vnread_flags);
776 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
777 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
778 	fp->f_vnread_flags = 0;
779 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
780 	sleepq_release(&fp->f_vnread_flags);
781 }
782 #else
783 off_t
784 foffset_lock(struct file *fp, int flags)
785 {
786 	struct mtx *mtxp;
787 	off_t res;
788 
789 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
790 
791 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
792 	mtx_lock(mtxp);
793 	if ((flags & FOF_NOLOCK) == 0) {
794 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
795 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
796 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
797 			    "vofflock", 0);
798 		}
799 		fp->f_vnread_flags |= FOFFSET_LOCKED;
800 	}
801 	res = fp->f_offset;
802 	mtx_unlock(mtxp);
803 	return (res);
804 }
805 
806 void
807 foffset_unlock(struct file *fp, off_t val, int flags)
808 {
809 	struct mtx *mtxp;
810 
811 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
812 
813 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
814 	mtx_lock(mtxp);
815 	if ((flags & FOF_NOUPDATE) == 0)
816 		fp->f_offset = val;
817 	if ((flags & FOF_NEXTOFF_R) != 0)
818 		fp->f_nextoff[UIO_READ] = val;
819 	if ((flags & FOF_NEXTOFF_W) != 0)
820 		fp->f_nextoff[UIO_WRITE] = val;
821 	if ((flags & FOF_NOLOCK) == 0) {
822 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
823 		    ("Lost FOFFSET_LOCKED"));
824 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
825 			wakeup(&fp->f_vnread_flags);
826 		fp->f_vnread_flags = 0;
827 	}
828 	mtx_unlock(mtxp);
829 }
830 #endif
831 
832 void
833 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
834 {
835 
836 	if ((flags & FOF_OFFSET) == 0)
837 		uio->uio_offset = foffset_lock(fp, flags);
838 }
839 
840 void
841 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
842 {
843 
844 	if ((flags & FOF_OFFSET) == 0)
845 		foffset_unlock(fp, uio->uio_offset, flags);
846 }
847 
848 static int
849 get_advice(struct file *fp, struct uio *uio)
850 {
851 	struct mtx *mtxp;
852 	int ret;
853 
854 	ret = POSIX_FADV_NORMAL;
855 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
856 		return (ret);
857 
858 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
859 	mtx_lock(mtxp);
860 	if (fp->f_advice != NULL &&
861 	    uio->uio_offset >= fp->f_advice->fa_start &&
862 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
863 		ret = fp->f_advice->fa_advice;
864 	mtx_unlock(mtxp);
865 	return (ret);
866 }
867 
868 int
869 vn_read_from_obj(struct vnode *vp, struct uio *uio)
870 {
871 	vm_object_t obj;
872 	vm_page_t ma[io_hold_cnt + 2];
873 	off_t off, vsz;
874 	ssize_t resid;
875 	int error, i, j;
876 
877 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
878 	obj = atomic_load_ptr(&vp->v_object);
879 	if (obj == NULL)
880 		return (EJUSTRETURN);
881 
882 	/*
883 	 * Depends on type stability of vm_objects.
884 	 */
885 	vm_object_pip_add(obj, 1);
886 	if ((obj->flags & OBJ_DEAD) != 0) {
887 		/*
888 		 * Note that object might be already reused from the
889 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
890 		 * we recheck for DOOMED vnode state after all pages
891 		 * are busied, and retract then.
892 		 *
893 		 * But we check for OBJ_DEAD to ensure that we do not
894 		 * busy pages while vm_object_terminate_pages()
895 		 * processes the queue.
896 		 */
897 		error = EJUSTRETURN;
898 		goto out_pip;
899 	}
900 
901 	resid = uio->uio_resid;
902 	off = uio->uio_offset;
903 	for (i = 0; resid > 0; i++) {
904 		MPASS(i < io_hold_cnt + 2);
905 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
906 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
907 		    VM_ALLOC_NOWAIT);
908 		if (ma[i] == NULL)
909 			break;
910 
911 		/*
912 		 * Skip invalid pages.  Valid mask can be partial only
913 		 * at EOF, and we clip later.
914 		 */
915 		if (vm_page_none_valid(ma[i])) {
916 			vm_page_sunbusy(ma[i]);
917 			break;
918 		}
919 
920 		resid -= PAGE_SIZE;
921 		off += PAGE_SIZE;
922 	}
923 	if (i == 0) {
924 		error = EJUSTRETURN;
925 		goto out_pip;
926 	}
927 
928 	/*
929 	 * Check VIRF_DOOMED after we busied our pages.  Since
930 	 * vgonel() terminates the vnode' vm_object, it cannot
931 	 * process past pages busied by us.
932 	 */
933 	if (VN_IS_DOOMED(vp)) {
934 		error = EJUSTRETURN;
935 		goto out;
936 	}
937 
938 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
939 	if (resid > uio->uio_resid)
940 		resid = uio->uio_resid;
941 
942 	/*
943 	 * Unlocked read of vnp_size is safe because truncation cannot
944 	 * pass busied page.  But we load vnp_size into a local
945 	 * variable so that possible concurrent extension does not
946 	 * break calculation.
947 	 */
948 #if defined(__powerpc__) && !defined(__powerpc64__)
949 	vsz = obj->un_pager.vnp.vnp_size;
950 #else
951 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
952 #endif
953 	if (uio->uio_offset >= vsz) {
954 		error = EJUSTRETURN;
955 		goto out;
956 	}
957 	if (uio->uio_offset + resid > vsz)
958 		resid = vsz - uio->uio_offset;
959 
960 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
961 
962 out:
963 	for (j = 0; j < i; j++) {
964 		if (error == 0)
965 			vm_page_reference(ma[j]);
966 		vm_page_sunbusy(ma[j]);
967 	}
968 out_pip:
969 	vm_object_pip_wakeup(obj);
970 	if (error != 0)
971 		return (error);
972 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
973 }
974 
975 /*
976  * File table vnode read routine.
977  */
978 static int
979 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
980     struct thread *td)
981 {
982 	struct vnode *vp;
983 	off_t orig_offset;
984 	int error, ioflag;
985 	int advice;
986 
987 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
988 	    uio->uio_td, td));
989 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
990 	vp = fp->f_vnode;
991 	ioflag = 0;
992 	if (fp->f_flag & FNONBLOCK)
993 		ioflag |= IO_NDELAY;
994 	if (fp->f_flag & O_DIRECT)
995 		ioflag |= IO_DIRECT;
996 
997 	/*
998 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
999 	 * allows us to avoid unneeded work outright.
1000 	 */
1001 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1002 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1003 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1004 		if (error == 0) {
1005 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1006 			return (0);
1007 		}
1008 		if (error != EJUSTRETURN)
1009 			return (error);
1010 	}
1011 
1012 	advice = get_advice(fp, uio);
1013 	vn_lock(vp, LK_SHARED | LK_RETRY);
1014 
1015 	switch (advice) {
1016 	case POSIX_FADV_NORMAL:
1017 	case POSIX_FADV_SEQUENTIAL:
1018 	case POSIX_FADV_NOREUSE:
1019 		ioflag |= sequential_heuristic(uio, fp);
1020 		break;
1021 	case POSIX_FADV_RANDOM:
1022 		/* Disable read-ahead for random I/O. */
1023 		break;
1024 	}
1025 	orig_offset = uio->uio_offset;
1026 
1027 #ifdef MAC
1028 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1029 	if (error == 0)
1030 #endif
1031 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1032 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1033 	VOP_UNLOCK(vp);
1034 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1035 	    orig_offset != uio->uio_offset)
1036 		/*
1037 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1038 		 * for the backing file after a POSIX_FADV_NOREUSE
1039 		 * read(2).
1040 		 */
1041 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1042 		    POSIX_FADV_DONTNEED);
1043 	return (error);
1044 }
1045 
1046 /*
1047  * File table vnode write routine.
1048  */
1049 static int
1050 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1051     struct thread *td)
1052 {
1053 	struct vnode *vp;
1054 	struct mount *mp;
1055 	off_t orig_offset;
1056 	int error, ioflag, lock_flags;
1057 	int advice;
1058 
1059 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1060 	    uio->uio_td, td));
1061 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1062 	vp = fp->f_vnode;
1063 	if (vp->v_type == VREG)
1064 		bwillwrite();
1065 	ioflag = IO_UNIT;
1066 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1067 		ioflag |= IO_APPEND;
1068 	if (fp->f_flag & FNONBLOCK)
1069 		ioflag |= IO_NDELAY;
1070 	if (fp->f_flag & O_DIRECT)
1071 		ioflag |= IO_DIRECT;
1072 	if ((fp->f_flag & O_FSYNC) ||
1073 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1074 		ioflag |= IO_SYNC;
1075 	/*
1076 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1077 	 * implementations that don't understand IO_DATASYNC fall back to full
1078 	 * O_SYNC behavior.
1079 	 */
1080 	if (fp->f_flag & O_DSYNC)
1081 		ioflag |= IO_SYNC | IO_DATASYNC;
1082 	mp = NULL;
1083 	if (vp->v_type != VCHR &&
1084 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1085 		goto unlock;
1086 
1087 	advice = get_advice(fp, uio);
1088 
1089 	if (MNT_SHARED_WRITES(mp) ||
1090 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1091 		lock_flags = LK_SHARED;
1092 	} else {
1093 		lock_flags = LK_EXCLUSIVE;
1094 	}
1095 
1096 	vn_lock(vp, lock_flags | LK_RETRY);
1097 	switch (advice) {
1098 	case POSIX_FADV_NORMAL:
1099 	case POSIX_FADV_SEQUENTIAL:
1100 	case POSIX_FADV_NOREUSE:
1101 		ioflag |= sequential_heuristic(uio, fp);
1102 		break;
1103 	case POSIX_FADV_RANDOM:
1104 		/* XXX: Is this correct? */
1105 		break;
1106 	}
1107 	orig_offset = uio->uio_offset;
1108 
1109 #ifdef MAC
1110 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1111 	if (error == 0)
1112 #endif
1113 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1114 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1115 	VOP_UNLOCK(vp);
1116 	if (vp->v_type != VCHR)
1117 		vn_finished_write(mp);
1118 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1119 	    orig_offset != uio->uio_offset)
1120 		/*
1121 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1122 		 * for the backing file after a POSIX_FADV_NOREUSE
1123 		 * write(2).
1124 		 */
1125 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1126 		    POSIX_FADV_DONTNEED);
1127 unlock:
1128 	return (error);
1129 }
1130 
1131 /*
1132  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1133  * prevent the following deadlock:
1134  *
1135  * Assume that the thread A reads from the vnode vp1 into userspace
1136  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1137  * currently not resident, then system ends up with the call chain
1138  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1139  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1140  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1141  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1142  * backed by the pages of vnode vp1, and some page in buf2 is not
1143  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1144  *
1145  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1146  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1147  * Instead, it first tries to do the whole range i/o with pagefaults
1148  * disabled. If all pages in the i/o buffer are resident and mapped,
1149  * VOP will succeed (ignoring the genuine filesystem errors).
1150  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1151  * i/o in chunks, with all pages in the chunk prefaulted and held
1152  * using vm_fault_quick_hold_pages().
1153  *
1154  * Filesystems using this deadlock avoidance scheme should use the
1155  * array of the held pages from uio, saved in the curthread->td_ma,
1156  * instead of doing uiomove().  A helper function
1157  * vn_io_fault_uiomove() converts uiomove request into
1158  * uiomove_fromphys() over td_ma array.
1159  *
1160  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1161  * make the current i/o request atomic with respect to other i/os and
1162  * truncations.
1163  */
1164 
1165 /*
1166  * Decode vn_io_fault_args and perform the corresponding i/o.
1167  */
1168 static int
1169 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1170     struct thread *td)
1171 {
1172 	int error, save;
1173 
1174 	error = 0;
1175 	save = vm_fault_disable_pagefaults();
1176 	switch (args->kind) {
1177 	case VN_IO_FAULT_FOP:
1178 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1179 		    uio, args->cred, args->flags, td);
1180 		break;
1181 	case VN_IO_FAULT_VOP:
1182 		if (uio->uio_rw == UIO_READ) {
1183 			error = VOP_READ(args->args.vop_args.vp, uio,
1184 			    args->flags, args->cred);
1185 		} else if (uio->uio_rw == UIO_WRITE) {
1186 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1187 			    args->flags, args->cred);
1188 		}
1189 		break;
1190 	default:
1191 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1192 		    args->kind, uio->uio_rw);
1193 	}
1194 	vm_fault_enable_pagefaults(save);
1195 	return (error);
1196 }
1197 
1198 static int
1199 vn_io_fault_touch(char *base, const struct uio *uio)
1200 {
1201 	int r;
1202 
1203 	r = fubyte(base);
1204 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1205 		return (EFAULT);
1206 	return (0);
1207 }
1208 
1209 static int
1210 vn_io_fault_prefault_user(const struct uio *uio)
1211 {
1212 	char *base;
1213 	const struct iovec *iov;
1214 	size_t len;
1215 	ssize_t resid;
1216 	int error, i;
1217 
1218 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1219 	    ("vn_io_fault_prefault userspace"));
1220 
1221 	error = i = 0;
1222 	iov = uio->uio_iov;
1223 	resid = uio->uio_resid;
1224 	base = iov->iov_base;
1225 	len = iov->iov_len;
1226 	while (resid > 0) {
1227 		error = vn_io_fault_touch(base, uio);
1228 		if (error != 0)
1229 			break;
1230 		if (len < PAGE_SIZE) {
1231 			if (len != 0) {
1232 				error = vn_io_fault_touch(base + len - 1, uio);
1233 				if (error != 0)
1234 					break;
1235 				resid -= len;
1236 			}
1237 			if (++i >= uio->uio_iovcnt)
1238 				break;
1239 			iov = uio->uio_iov + i;
1240 			base = iov->iov_base;
1241 			len = iov->iov_len;
1242 		} else {
1243 			len -= PAGE_SIZE;
1244 			base += PAGE_SIZE;
1245 			resid -= PAGE_SIZE;
1246 		}
1247 	}
1248 	return (error);
1249 }
1250 
1251 /*
1252  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1253  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1254  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1255  * into args and call vn_io_fault1() to handle faults during the user
1256  * mode buffer accesses.
1257  */
1258 static int
1259 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1260     struct thread *td)
1261 {
1262 	vm_page_t ma[io_hold_cnt + 2];
1263 	struct uio *uio_clone, short_uio;
1264 	struct iovec short_iovec[1];
1265 	vm_page_t *prev_td_ma;
1266 	vm_prot_t prot;
1267 	vm_offset_t addr, end;
1268 	size_t len, resid;
1269 	ssize_t adv;
1270 	int error, cnt, saveheld, prev_td_ma_cnt;
1271 
1272 	if (vn_io_fault_prefault) {
1273 		error = vn_io_fault_prefault_user(uio);
1274 		if (error != 0)
1275 			return (error); /* Or ignore ? */
1276 	}
1277 
1278 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1279 
1280 	/*
1281 	 * The UFS follows IO_UNIT directive and replays back both
1282 	 * uio_offset and uio_resid if an error is encountered during the
1283 	 * operation.  But, since the iovec may be already advanced,
1284 	 * uio is still in an inconsistent state.
1285 	 *
1286 	 * Cache a copy of the original uio, which is advanced to the redo
1287 	 * point using UIO_NOCOPY below.
1288 	 */
1289 	uio_clone = cloneuio(uio);
1290 	resid = uio->uio_resid;
1291 
1292 	short_uio.uio_segflg = UIO_USERSPACE;
1293 	short_uio.uio_rw = uio->uio_rw;
1294 	short_uio.uio_td = uio->uio_td;
1295 
1296 	error = vn_io_fault_doio(args, uio, td);
1297 	if (error != EFAULT)
1298 		goto out;
1299 
1300 	atomic_add_long(&vn_io_faults_cnt, 1);
1301 	uio_clone->uio_segflg = UIO_NOCOPY;
1302 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1303 	uio_clone->uio_segflg = uio->uio_segflg;
1304 
1305 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1306 	prev_td_ma = td->td_ma;
1307 	prev_td_ma_cnt = td->td_ma_cnt;
1308 
1309 	while (uio_clone->uio_resid != 0) {
1310 		len = uio_clone->uio_iov->iov_len;
1311 		if (len == 0) {
1312 			KASSERT(uio_clone->uio_iovcnt >= 1,
1313 			    ("iovcnt underflow"));
1314 			uio_clone->uio_iov++;
1315 			uio_clone->uio_iovcnt--;
1316 			continue;
1317 		}
1318 		if (len > ptoa(io_hold_cnt))
1319 			len = ptoa(io_hold_cnt);
1320 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1321 		end = round_page(addr + len);
1322 		if (end < addr) {
1323 			error = EFAULT;
1324 			break;
1325 		}
1326 		cnt = atop(end - trunc_page(addr));
1327 		/*
1328 		 * A perfectly misaligned address and length could cause
1329 		 * both the start and the end of the chunk to use partial
1330 		 * page.  +2 accounts for such a situation.
1331 		 */
1332 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1333 		    addr, len, prot, ma, io_hold_cnt + 2);
1334 		if (cnt == -1) {
1335 			error = EFAULT;
1336 			break;
1337 		}
1338 		short_uio.uio_iov = &short_iovec[0];
1339 		short_iovec[0].iov_base = (void *)addr;
1340 		short_uio.uio_iovcnt = 1;
1341 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1342 		short_uio.uio_offset = uio_clone->uio_offset;
1343 		td->td_ma = ma;
1344 		td->td_ma_cnt = cnt;
1345 
1346 		error = vn_io_fault_doio(args, &short_uio, td);
1347 		vm_page_unhold_pages(ma, cnt);
1348 		adv = len - short_uio.uio_resid;
1349 
1350 		uio_clone->uio_iov->iov_base =
1351 		    (char *)uio_clone->uio_iov->iov_base + adv;
1352 		uio_clone->uio_iov->iov_len -= adv;
1353 		uio_clone->uio_resid -= adv;
1354 		uio_clone->uio_offset += adv;
1355 
1356 		uio->uio_resid -= adv;
1357 		uio->uio_offset += adv;
1358 
1359 		if (error != 0 || adv == 0)
1360 			break;
1361 	}
1362 	td->td_ma = prev_td_ma;
1363 	td->td_ma_cnt = prev_td_ma_cnt;
1364 	curthread_pflags_restore(saveheld);
1365 out:
1366 	free(uio_clone, M_IOV);
1367 	return (error);
1368 }
1369 
1370 static int
1371 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1372     int flags, struct thread *td)
1373 {
1374 	fo_rdwr_t *doio;
1375 	struct vnode *vp;
1376 	void *rl_cookie;
1377 	struct vn_io_fault_args args;
1378 	int error;
1379 
1380 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1381 	vp = fp->f_vnode;
1382 
1383 	/*
1384 	 * The ability to read(2) on a directory has historically been
1385 	 * allowed for all users, but this can and has been the source of
1386 	 * at least one security issue in the past.  As such, it is now hidden
1387 	 * away behind a sysctl for those that actually need it to use it, and
1388 	 * restricted to root when it's turned on to make it relatively safe to
1389 	 * leave on for longer sessions of need.
1390 	 */
1391 	if (vp->v_type == VDIR) {
1392 		KASSERT(uio->uio_rw == UIO_READ,
1393 		    ("illegal write attempted on a directory"));
1394 		if (!vfs_allow_read_dir)
1395 			return (EISDIR);
1396 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1397 			return (EISDIR);
1398 	}
1399 
1400 	foffset_lock_uio(fp, uio, flags);
1401 	if (do_vn_io_fault(vp, uio)) {
1402 		args.kind = VN_IO_FAULT_FOP;
1403 		args.args.fop_args.fp = fp;
1404 		args.args.fop_args.doio = doio;
1405 		args.cred = active_cred;
1406 		args.flags = flags | FOF_OFFSET;
1407 		if (uio->uio_rw == UIO_READ) {
1408 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1409 			    uio->uio_offset + uio->uio_resid);
1410 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1411 		    (flags & FOF_OFFSET) == 0) {
1412 			/* For appenders, punt and lock the whole range. */
1413 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1414 		} else {
1415 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1416 			    uio->uio_offset + uio->uio_resid);
1417 		}
1418 		error = vn_io_fault1(vp, uio, &args, td);
1419 		vn_rangelock_unlock(vp, rl_cookie);
1420 	} else {
1421 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1422 	}
1423 	foffset_unlock_uio(fp, uio, flags);
1424 	return (error);
1425 }
1426 
1427 /*
1428  * Helper function to perform the requested uiomove operation using
1429  * the held pages for io->uio_iov[0].iov_base buffer instead of
1430  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1431  * instead of iov_base prevents page faults that could occur due to
1432  * pmap_collect() invalidating the mapping created by
1433  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1434  * object cleanup revoking the write access from page mappings.
1435  *
1436  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1437  * instead of plain uiomove().
1438  */
1439 int
1440 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1441 {
1442 	struct uio transp_uio;
1443 	struct iovec transp_iov[1];
1444 	struct thread *td;
1445 	size_t adv;
1446 	int error, pgadv;
1447 
1448 	td = curthread;
1449 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1450 	    uio->uio_segflg != UIO_USERSPACE)
1451 		return (uiomove(data, xfersize, uio));
1452 
1453 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1454 	transp_iov[0].iov_base = data;
1455 	transp_uio.uio_iov = &transp_iov[0];
1456 	transp_uio.uio_iovcnt = 1;
1457 	if (xfersize > uio->uio_resid)
1458 		xfersize = uio->uio_resid;
1459 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1460 	transp_uio.uio_offset = 0;
1461 	transp_uio.uio_segflg = UIO_SYSSPACE;
1462 	/*
1463 	 * Since transp_iov points to data, and td_ma page array
1464 	 * corresponds to original uio->uio_iov, we need to invert the
1465 	 * direction of the i/o operation as passed to
1466 	 * uiomove_fromphys().
1467 	 */
1468 	switch (uio->uio_rw) {
1469 	case UIO_WRITE:
1470 		transp_uio.uio_rw = UIO_READ;
1471 		break;
1472 	case UIO_READ:
1473 		transp_uio.uio_rw = UIO_WRITE;
1474 		break;
1475 	}
1476 	transp_uio.uio_td = uio->uio_td;
1477 	error = uiomove_fromphys(td->td_ma,
1478 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1479 	    xfersize, &transp_uio);
1480 	adv = xfersize - transp_uio.uio_resid;
1481 	pgadv =
1482 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1483 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1484 	td->td_ma += pgadv;
1485 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1486 	    pgadv));
1487 	td->td_ma_cnt -= pgadv;
1488 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1489 	uio->uio_iov->iov_len -= adv;
1490 	uio->uio_resid -= adv;
1491 	uio->uio_offset += adv;
1492 	return (error);
1493 }
1494 
1495 int
1496 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1497     struct uio *uio)
1498 {
1499 	struct thread *td;
1500 	vm_offset_t iov_base;
1501 	int cnt, pgadv;
1502 
1503 	td = curthread;
1504 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1505 	    uio->uio_segflg != UIO_USERSPACE)
1506 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1507 
1508 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1509 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1510 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1511 	switch (uio->uio_rw) {
1512 	case UIO_WRITE:
1513 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1514 		    offset, cnt);
1515 		break;
1516 	case UIO_READ:
1517 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1518 		    cnt);
1519 		break;
1520 	}
1521 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1522 	td->td_ma += pgadv;
1523 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1524 	    pgadv));
1525 	td->td_ma_cnt -= pgadv;
1526 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1527 	uio->uio_iov->iov_len -= cnt;
1528 	uio->uio_resid -= cnt;
1529 	uio->uio_offset += cnt;
1530 	return (0);
1531 }
1532 
1533 /*
1534  * File table truncate routine.
1535  */
1536 static int
1537 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1538     struct thread *td)
1539 {
1540 	struct mount *mp;
1541 	struct vnode *vp;
1542 	void *rl_cookie;
1543 	int error;
1544 
1545 	vp = fp->f_vnode;
1546 
1547 retry:
1548 	/*
1549 	 * Lock the whole range for truncation.  Otherwise split i/o
1550 	 * might happen partly before and partly after the truncation.
1551 	 */
1552 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1553 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1554 	if (error)
1555 		goto out1;
1556 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1557 	AUDIT_ARG_VNODE1(vp);
1558 	if (vp->v_type == VDIR) {
1559 		error = EISDIR;
1560 		goto out;
1561 	}
1562 #ifdef MAC
1563 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1564 	if (error)
1565 		goto out;
1566 #endif
1567 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1568 	    fp->f_cred);
1569 out:
1570 	VOP_UNLOCK(vp);
1571 	vn_finished_write(mp);
1572 out1:
1573 	vn_rangelock_unlock(vp, rl_cookie);
1574 	if (error == ERELOOKUP)
1575 		goto retry;
1576 	return (error);
1577 }
1578 
1579 /*
1580  * Truncate a file that is already locked.
1581  */
1582 int
1583 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1584     struct ucred *cred)
1585 {
1586 	struct vattr vattr;
1587 	int error;
1588 
1589 	error = VOP_ADD_WRITECOUNT(vp, 1);
1590 	if (error == 0) {
1591 		VATTR_NULL(&vattr);
1592 		vattr.va_size = length;
1593 		if (sync)
1594 			vattr.va_vaflags |= VA_SYNC;
1595 		error = VOP_SETATTR(vp, &vattr, cred);
1596 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1597 	}
1598 	return (error);
1599 }
1600 
1601 /*
1602  * File table vnode stat routine.
1603  */
1604 static int
1605 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1606     struct thread *td)
1607 {
1608 	struct vnode *vp = fp->f_vnode;
1609 	int error;
1610 
1611 	vn_lock(vp, LK_SHARED | LK_RETRY);
1612 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1613 	VOP_UNLOCK(vp);
1614 
1615 	return (error);
1616 }
1617 
1618 /*
1619  * File table vnode ioctl routine.
1620  */
1621 static int
1622 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1623     struct thread *td)
1624 {
1625 	struct vattr vattr;
1626 	struct vnode *vp;
1627 	struct fiobmap2_arg *bmarg;
1628 	int error;
1629 
1630 	vp = fp->f_vnode;
1631 	switch (vp->v_type) {
1632 	case VDIR:
1633 	case VREG:
1634 		switch (com) {
1635 		case FIONREAD:
1636 			vn_lock(vp, LK_SHARED | LK_RETRY);
1637 			error = VOP_GETATTR(vp, &vattr, active_cred);
1638 			VOP_UNLOCK(vp);
1639 			if (error == 0)
1640 				*(int *)data = vattr.va_size - fp->f_offset;
1641 			return (error);
1642 		case FIOBMAP2:
1643 			bmarg = (struct fiobmap2_arg *)data;
1644 			vn_lock(vp, LK_SHARED | LK_RETRY);
1645 #ifdef MAC
1646 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1647 			    vp);
1648 			if (error == 0)
1649 #endif
1650 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1651 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1652 			VOP_UNLOCK(vp);
1653 			return (error);
1654 		case FIONBIO:
1655 		case FIOASYNC:
1656 			return (0);
1657 		default:
1658 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1659 			    active_cred, td));
1660 		}
1661 		break;
1662 	case VCHR:
1663 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1664 		    active_cred, td));
1665 	default:
1666 		return (ENOTTY);
1667 	}
1668 }
1669 
1670 /*
1671  * File table vnode poll routine.
1672  */
1673 static int
1674 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1675     struct thread *td)
1676 {
1677 	struct vnode *vp;
1678 	int error;
1679 
1680 	vp = fp->f_vnode;
1681 #if defined(MAC) || defined(AUDIT)
1682 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1683 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1684 		AUDIT_ARG_VNODE1(vp);
1685 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1686 		VOP_UNLOCK(vp);
1687 		if (error != 0)
1688 			return (error);
1689 	}
1690 #endif
1691 	error = VOP_POLL(vp, events, fp->f_cred, td);
1692 	return (error);
1693 }
1694 
1695 /*
1696  * Acquire the requested lock and then check for validity.  LK_RETRY
1697  * permits vn_lock to return doomed vnodes.
1698  */
1699 static int __noinline
1700 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1701     int error)
1702 {
1703 
1704 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1705 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1706 
1707 	if (error == 0)
1708 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1709 
1710 	if ((flags & LK_RETRY) == 0) {
1711 		if (error == 0) {
1712 			VOP_UNLOCK(vp);
1713 			error = ENOENT;
1714 		}
1715 		return (error);
1716 	}
1717 
1718 	/*
1719 	 * LK_RETRY case.
1720 	 *
1721 	 * Nothing to do if we got the lock.
1722 	 */
1723 	if (error == 0)
1724 		return (0);
1725 
1726 	/*
1727 	 * Interlock was dropped by the call in _vn_lock.
1728 	 */
1729 	flags &= ~LK_INTERLOCK;
1730 	do {
1731 		error = VOP_LOCK1(vp, flags, file, line);
1732 	} while (error != 0);
1733 	return (0);
1734 }
1735 
1736 int
1737 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1738 {
1739 	int error;
1740 
1741 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1742 	    ("vn_lock: no locktype (%d passed)", flags));
1743 	VNPASS(vp->v_holdcnt > 0, vp);
1744 	error = VOP_LOCK1(vp, flags, file, line);
1745 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1746 		return (_vn_lock_fallback(vp, flags, file, line, error));
1747 	return (0);
1748 }
1749 
1750 /*
1751  * File table vnode close routine.
1752  */
1753 static int
1754 vn_closefile(struct file *fp, struct thread *td)
1755 {
1756 	struct vnode *vp;
1757 	struct flock lf;
1758 	int error;
1759 	bool ref;
1760 
1761 	vp = fp->f_vnode;
1762 	fp->f_ops = &badfileops;
1763 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1764 
1765 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1766 
1767 	if (__predict_false(ref)) {
1768 		lf.l_whence = SEEK_SET;
1769 		lf.l_start = 0;
1770 		lf.l_len = 0;
1771 		lf.l_type = F_UNLCK;
1772 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1773 		vrele(vp);
1774 	}
1775 	return (error);
1776 }
1777 
1778 /*
1779  * Preparing to start a filesystem write operation. If the operation is
1780  * permitted, then we bump the count of operations in progress and
1781  * proceed. If a suspend request is in progress, we wait until the
1782  * suspension is over, and then proceed.
1783  */
1784 static int
1785 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1786 {
1787 	struct mount_pcpu *mpcpu;
1788 	int error, mflags;
1789 
1790 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1791 	    vfs_op_thread_enter(mp, mpcpu)) {
1792 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1793 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1794 		vfs_op_thread_exit(mp, mpcpu);
1795 		return (0);
1796 	}
1797 
1798 	if (mplocked)
1799 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1800 	else
1801 		MNT_ILOCK(mp);
1802 
1803 	error = 0;
1804 
1805 	/*
1806 	 * Check on status of suspension.
1807 	 */
1808 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1809 	    mp->mnt_susp_owner != curthread) {
1810 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1811 		    (flags & PCATCH) : 0) | (PUSER - 1);
1812 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1813 			if (flags & V_NOWAIT) {
1814 				error = EWOULDBLOCK;
1815 				goto unlock;
1816 			}
1817 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1818 			    "suspfs", 0);
1819 			if (error)
1820 				goto unlock;
1821 		}
1822 	}
1823 	if (flags & V_XSLEEP)
1824 		goto unlock;
1825 	mp->mnt_writeopcount++;
1826 unlock:
1827 	if (error != 0 || (flags & V_XSLEEP) != 0)
1828 		MNT_REL(mp);
1829 	MNT_IUNLOCK(mp);
1830 	return (error);
1831 }
1832 
1833 int
1834 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1835 {
1836 	struct mount *mp;
1837 	int error;
1838 
1839 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1840 	    ("V_MNTREF requires mp"));
1841 
1842 	error = 0;
1843 	/*
1844 	 * If a vnode is provided, get and return the mount point that
1845 	 * to which it will write.
1846 	 */
1847 	if (vp != NULL) {
1848 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1849 			*mpp = NULL;
1850 			if (error != EOPNOTSUPP)
1851 				return (error);
1852 			return (0);
1853 		}
1854 	}
1855 	if ((mp = *mpp) == NULL)
1856 		return (0);
1857 
1858 	/*
1859 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1860 	 * a vfs_ref().
1861 	 * As long as a vnode is not provided we need to acquire a
1862 	 * refcount for the provided mountpoint too, in order to
1863 	 * emulate a vfs_ref().
1864 	 */
1865 	if (vp == NULL && (flags & V_MNTREF) == 0)
1866 		vfs_ref(mp);
1867 
1868 	return (vn_start_write_refed(mp, flags, false));
1869 }
1870 
1871 /*
1872  * Secondary suspension. Used by operations such as vop_inactive
1873  * routines that are needed by the higher level functions. These
1874  * are allowed to proceed until all the higher level functions have
1875  * completed (indicated by mnt_writeopcount dropping to zero). At that
1876  * time, these operations are halted until the suspension is over.
1877  */
1878 int
1879 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1880 {
1881 	struct mount *mp;
1882 	int error;
1883 
1884 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1885 	    ("V_MNTREF requires mp"));
1886 
1887  retry:
1888 	if (vp != NULL) {
1889 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1890 			*mpp = NULL;
1891 			if (error != EOPNOTSUPP)
1892 				return (error);
1893 			return (0);
1894 		}
1895 	}
1896 	/*
1897 	 * If we are not suspended or have not yet reached suspended
1898 	 * mode, then let the operation proceed.
1899 	 */
1900 	if ((mp = *mpp) == NULL)
1901 		return (0);
1902 
1903 	/*
1904 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1905 	 * a vfs_ref().
1906 	 * As long as a vnode is not provided we need to acquire a
1907 	 * refcount for the provided mountpoint too, in order to
1908 	 * emulate a vfs_ref().
1909 	 */
1910 	MNT_ILOCK(mp);
1911 	if (vp == NULL && (flags & V_MNTREF) == 0)
1912 		MNT_REF(mp);
1913 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1914 		mp->mnt_secondary_writes++;
1915 		mp->mnt_secondary_accwrites++;
1916 		MNT_IUNLOCK(mp);
1917 		return (0);
1918 	}
1919 	if (flags & V_NOWAIT) {
1920 		MNT_REL(mp);
1921 		MNT_IUNLOCK(mp);
1922 		return (EWOULDBLOCK);
1923 	}
1924 	/*
1925 	 * Wait for the suspension to finish.
1926 	 */
1927 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1928 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1929 	    "suspfs", 0);
1930 	vfs_rel(mp);
1931 	if (error == 0)
1932 		goto retry;
1933 	return (error);
1934 }
1935 
1936 /*
1937  * Filesystem write operation has completed. If we are suspending and this
1938  * operation is the last one, notify the suspender that the suspension is
1939  * now in effect.
1940  */
1941 void
1942 vn_finished_write(struct mount *mp)
1943 {
1944 	struct mount_pcpu *mpcpu;
1945 	int c;
1946 
1947 	if (mp == NULL)
1948 		return;
1949 
1950 	if (vfs_op_thread_enter(mp, mpcpu)) {
1951 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
1952 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
1953 		vfs_op_thread_exit(mp, mpcpu);
1954 		return;
1955 	}
1956 
1957 	MNT_ILOCK(mp);
1958 	vfs_assert_mount_counters(mp);
1959 	MNT_REL(mp);
1960 	c = --mp->mnt_writeopcount;
1961 	if (mp->mnt_vfs_ops == 0) {
1962 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1963 		MNT_IUNLOCK(mp);
1964 		return;
1965 	}
1966 	if (c < 0)
1967 		vfs_dump_mount_counters(mp);
1968 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1969 		wakeup(&mp->mnt_writeopcount);
1970 	MNT_IUNLOCK(mp);
1971 }
1972 
1973 /*
1974  * Filesystem secondary write operation has completed. If we are
1975  * suspending and this operation is the last one, notify the suspender
1976  * that the suspension is now in effect.
1977  */
1978 void
1979 vn_finished_secondary_write(struct mount *mp)
1980 {
1981 	if (mp == NULL)
1982 		return;
1983 	MNT_ILOCK(mp);
1984 	MNT_REL(mp);
1985 	mp->mnt_secondary_writes--;
1986 	if (mp->mnt_secondary_writes < 0)
1987 		panic("vn_finished_secondary_write: neg cnt");
1988 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1989 	    mp->mnt_secondary_writes <= 0)
1990 		wakeup(&mp->mnt_secondary_writes);
1991 	MNT_IUNLOCK(mp);
1992 }
1993 
1994 /*
1995  * Request a filesystem to suspend write operations.
1996  */
1997 int
1998 vfs_write_suspend(struct mount *mp, int flags)
1999 {
2000 	int error;
2001 
2002 	vfs_op_enter(mp);
2003 
2004 	MNT_ILOCK(mp);
2005 	vfs_assert_mount_counters(mp);
2006 	if (mp->mnt_susp_owner == curthread) {
2007 		vfs_op_exit_locked(mp);
2008 		MNT_IUNLOCK(mp);
2009 		return (EALREADY);
2010 	}
2011 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2012 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2013 
2014 	/*
2015 	 * Unmount holds a write reference on the mount point.  If we
2016 	 * own busy reference and drain for writers, we deadlock with
2017 	 * the reference draining in the unmount path.  Callers of
2018 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2019 	 * vfs_busy() reference is owned and caller is not in the
2020 	 * unmount context.
2021 	 */
2022 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2023 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2024 		vfs_op_exit_locked(mp);
2025 		MNT_IUNLOCK(mp);
2026 		return (EBUSY);
2027 	}
2028 
2029 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2030 	mp->mnt_susp_owner = curthread;
2031 	if (mp->mnt_writeopcount > 0)
2032 		(void) msleep(&mp->mnt_writeopcount,
2033 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2034 	else
2035 		MNT_IUNLOCK(mp);
2036 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2037 		vfs_write_resume(mp, 0);
2038 		/* vfs_write_resume does vfs_op_exit() for us */
2039 	}
2040 	return (error);
2041 }
2042 
2043 /*
2044  * Request a filesystem to resume write operations.
2045  */
2046 void
2047 vfs_write_resume(struct mount *mp, int flags)
2048 {
2049 
2050 	MNT_ILOCK(mp);
2051 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2052 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2053 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2054 				       MNTK_SUSPENDED);
2055 		mp->mnt_susp_owner = NULL;
2056 		wakeup(&mp->mnt_writeopcount);
2057 		wakeup(&mp->mnt_flag);
2058 		curthread->td_pflags &= ~TDP_IGNSUSP;
2059 		if ((flags & VR_START_WRITE) != 0) {
2060 			MNT_REF(mp);
2061 			mp->mnt_writeopcount++;
2062 		}
2063 		MNT_IUNLOCK(mp);
2064 		if ((flags & VR_NO_SUSPCLR) == 0)
2065 			VFS_SUSP_CLEAN(mp);
2066 		vfs_op_exit(mp);
2067 	} else if ((flags & VR_START_WRITE) != 0) {
2068 		MNT_REF(mp);
2069 		vn_start_write_refed(mp, 0, true);
2070 	} else {
2071 		MNT_IUNLOCK(mp);
2072 	}
2073 }
2074 
2075 /*
2076  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2077  * methods.
2078  */
2079 int
2080 vfs_write_suspend_umnt(struct mount *mp)
2081 {
2082 	int error;
2083 
2084 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2085 	    ("vfs_write_suspend_umnt: recursed"));
2086 
2087 	/* dounmount() already called vn_start_write(). */
2088 	for (;;) {
2089 		vn_finished_write(mp);
2090 		error = vfs_write_suspend(mp, 0);
2091 		if (error != 0) {
2092 			vn_start_write(NULL, &mp, V_WAIT);
2093 			return (error);
2094 		}
2095 		MNT_ILOCK(mp);
2096 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2097 			break;
2098 		MNT_IUNLOCK(mp);
2099 		vn_start_write(NULL, &mp, V_WAIT);
2100 	}
2101 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2102 	wakeup(&mp->mnt_flag);
2103 	MNT_IUNLOCK(mp);
2104 	curthread->td_pflags |= TDP_IGNSUSP;
2105 	return (0);
2106 }
2107 
2108 /*
2109  * Implement kqueues for files by translating it to vnode operation.
2110  */
2111 static int
2112 vn_kqfilter(struct file *fp, struct knote *kn)
2113 {
2114 
2115 	return (VOP_KQFILTER(fp->f_vnode, kn));
2116 }
2117 
2118 /*
2119  * Simplified in-kernel wrapper calls for extended attribute access.
2120  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2121  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2122  */
2123 int
2124 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2125     const char *attrname, int *buflen, char *buf, struct thread *td)
2126 {
2127 	struct uio	auio;
2128 	struct iovec	iov;
2129 	int	error;
2130 
2131 	iov.iov_len = *buflen;
2132 	iov.iov_base = buf;
2133 
2134 	auio.uio_iov = &iov;
2135 	auio.uio_iovcnt = 1;
2136 	auio.uio_rw = UIO_READ;
2137 	auio.uio_segflg = UIO_SYSSPACE;
2138 	auio.uio_td = td;
2139 	auio.uio_offset = 0;
2140 	auio.uio_resid = *buflen;
2141 
2142 	if ((ioflg & IO_NODELOCKED) == 0)
2143 		vn_lock(vp, LK_SHARED | LK_RETRY);
2144 
2145 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2146 
2147 	/* authorize attribute retrieval as kernel */
2148 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2149 	    td);
2150 
2151 	if ((ioflg & IO_NODELOCKED) == 0)
2152 		VOP_UNLOCK(vp);
2153 
2154 	if (error == 0) {
2155 		*buflen = *buflen - auio.uio_resid;
2156 	}
2157 
2158 	return (error);
2159 }
2160 
2161 /*
2162  * XXX failure mode if partially written?
2163  */
2164 int
2165 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2166     const char *attrname, int buflen, char *buf, struct thread *td)
2167 {
2168 	struct uio	auio;
2169 	struct iovec	iov;
2170 	struct mount	*mp;
2171 	int	error;
2172 
2173 	iov.iov_len = buflen;
2174 	iov.iov_base = buf;
2175 
2176 	auio.uio_iov = &iov;
2177 	auio.uio_iovcnt = 1;
2178 	auio.uio_rw = UIO_WRITE;
2179 	auio.uio_segflg = UIO_SYSSPACE;
2180 	auio.uio_td = td;
2181 	auio.uio_offset = 0;
2182 	auio.uio_resid = buflen;
2183 
2184 	if ((ioflg & IO_NODELOCKED) == 0) {
2185 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2186 			return (error);
2187 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2188 	}
2189 
2190 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2191 
2192 	/* authorize attribute setting as kernel */
2193 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2194 
2195 	if ((ioflg & IO_NODELOCKED) == 0) {
2196 		vn_finished_write(mp);
2197 		VOP_UNLOCK(vp);
2198 	}
2199 
2200 	return (error);
2201 }
2202 
2203 int
2204 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2205     const char *attrname, struct thread *td)
2206 {
2207 	struct mount	*mp;
2208 	int	error;
2209 
2210 	if ((ioflg & IO_NODELOCKED) == 0) {
2211 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2212 			return (error);
2213 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2214 	}
2215 
2216 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2217 
2218 	/* authorize attribute removal as kernel */
2219 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2220 	if (error == EOPNOTSUPP)
2221 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2222 		    NULL, td);
2223 
2224 	if ((ioflg & IO_NODELOCKED) == 0) {
2225 		vn_finished_write(mp);
2226 		VOP_UNLOCK(vp);
2227 	}
2228 
2229 	return (error);
2230 }
2231 
2232 static int
2233 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2234     struct vnode **rvp)
2235 {
2236 
2237 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2238 }
2239 
2240 int
2241 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2242 {
2243 
2244 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2245 	    lkflags, rvp));
2246 }
2247 
2248 int
2249 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2250     int lkflags, struct vnode **rvp)
2251 {
2252 	struct mount *mp;
2253 	int ltype, error;
2254 
2255 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2256 	mp = vp->v_mount;
2257 	ltype = VOP_ISLOCKED(vp);
2258 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2259 	    ("vn_vget_ino: vp not locked"));
2260 	error = vfs_busy(mp, MBF_NOWAIT);
2261 	if (error != 0) {
2262 		vfs_ref(mp);
2263 		VOP_UNLOCK(vp);
2264 		error = vfs_busy(mp, 0);
2265 		vn_lock(vp, ltype | LK_RETRY);
2266 		vfs_rel(mp);
2267 		if (error != 0)
2268 			return (ENOENT);
2269 		if (VN_IS_DOOMED(vp)) {
2270 			vfs_unbusy(mp);
2271 			return (ENOENT);
2272 		}
2273 	}
2274 	VOP_UNLOCK(vp);
2275 	error = alloc(mp, alloc_arg, lkflags, rvp);
2276 	vfs_unbusy(mp);
2277 	if (error != 0 || *rvp != vp)
2278 		vn_lock(vp, ltype | LK_RETRY);
2279 	if (VN_IS_DOOMED(vp)) {
2280 		if (error == 0) {
2281 			if (*rvp == vp)
2282 				vunref(vp);
2283 			else
2284 				vput(*rvp);
2285 		}
2286 		error = ENOENT;
2287 	}
2288 	return (error);
2289 }
2290 
2291 int
2292 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2293     struct thread *td)
2294 {
2295 
2296 	if (vp->v_type != VREG || td == NULL)
2297 		return (0);
2298 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2299 	    lim_cur(td, RLIMIT_FSIZE)) {
2300 		PROC_LOCK(td->td_proc);
2301 		kern_psignal(td->td_proc, SIGXFSZ);
2302 		PROC_UNLOCK(td->td_proc);
2303 		return (EFBIG);
2304 	}
2305 	return (0);
2306 }
2307 
2308 int
2309 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2310     struct thread *td)
2311 {
2312 	struct vnode *vp;
2313 
2314 	vp = fp->f_vnode;
2315 #ifdef AUDIT
2316 	vn_lock(vp, LK_SHARED | LK_RETRY);
2317 	AUDIT_ARG_VNODE1(vp);
2318 	VOP_UNLOCK(vp);
2319 #endif
2320 	return (setfmode(td, active_cred, vp, mode));
2321 }
2322 
2323 int
2324 vn_chown(struct file *fp, uid_t uid, gid_t gid, 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 (setfown(td, active_cred, vp, uid, gid));
2336 }
2337 
2338 void
2339 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2340 {
2341 	vm_object_t object;
2342 
2343 	if ((object = vp->v_object) == NULL)
2344 		return;
2345 	VM_OBJECT_WLOCK(object);
2346 	vm_object_page_remove(object, start, end, 0);
2347 	VM_OBJECT_WUNLOCK(object);
2348 }
2349 
2350 int
2351 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2352 {
2353 	struct vattr va;
2354 	daddr_t bn, bnp;
2355 	uint64_t bsize;
2356 	off_t noff;
2357 	int error;
2358 
2359 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2360 	    ("Wrong command %lu", cmd));
2361 
2362 	if (vn_lock(vp, LK_SHARED) != 0)
2363 		return (EBADF);
2364 	if (vp->v_type != VREG) {
2365 		error = ENOTTY;
2366 		goto unlock;
2367 	}
2368 	error = VOP_GETATTR(vp, &va, cred);
2369 	if (error != 0)
2370 		goto unlock;
2371 	noff = *off;
2372 	if (noff >= va.va_size) {
2373 		error = ENXIO;
2374 		goto unlock;
2375 	}
2376 	bsize = vp->v_mount->mnt_stat.f_iosize;
2377 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2378 	    noff % bsize) {
2379 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2380 		if (error == EOPNOTSUPP) {
2381 			error = ENOTTY;
2382 			goto unlock;
2383 		}
2384 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2385 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2386 			noff = bn * bsize;
2387 			if (noff < *off)
2388 				noff = *off;
2389 			goto unlock;
2390 		}
2391 	}
2392 	if (noff > va.va_size)
2393 		noff = va.va_size;
2394 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2395 	if (cmd == FIOSEEKDATA)
2396 		error = ENXIO;
2397 unlock:
2398 	VOP_UNLOCK(vp);
2399 	if (error == 0)
2400 		*off = noff;
2401 	return (error);
2402 }
2403 
2404 int
2405 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2406 {
2407 	struct ucred *cred;
2408 	struct vnode *vp;
2409 	struct vattr vattr;
2410 	off_t foffset, size;
2411 	int error, noneg;
2412 
2413 	cred = td->td_ucred;
2414 	vp = fp->f_vnode;
2415 	foffset = foffset_lock(fp, 0);
2416 	noneg = (vp->v_type != VCHR);
2417 	error = 0;
2418 	switch (whence) {
2419 	case L_INCR:
2420 		if (noneg &&
2421 		    (foffset < 0 ||
2422 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2423 			error = EOVERFLOW;
2424 			break;
2425 		}
2426 		offset += foffset;
2427 		break;
2428 	case L_XTND:
2429 		vn_lock(vp, LK_SHARED | LK_RETRY);
2430 		error = VOP_GETATTR(vp, &vattr, cred);
2431 		VOP_UNLOCK(vp);
2432 		if (error)
2433 			break;
2434 
2435 		/*
2436 		 * If the file references a disk device, then fetch
2437 		 * the media size and use that to determine the ending
2438 		 * offset.
2439 		 */
2440 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2441 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2442 			vattr.va_size = size;
2443 		if (noneg &&
2444 		    (vattr.va_size > OFF_MAX ||
2445 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2446 			error = EOVERFLOW;
2447 			break;
2448 		}
2449 		offset += vattr.va_size;
2450 		break;
2451 	case L_SET:
2452 		break;
2453 	case SEEK_DATA:
2454 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2455 		if (error == ENOTTY)
2456 			error = EINVAL;
2457 		break;
2458 	case SEEK_HOLE:
2459 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2460 		if (error == ENOTTY)
2461 			error = EINVAL;
2462 		break;
2463 	default:
2464 		error = EINVAL;
2465 	}
2466 	if (error == 0 && noneg && offset < 0)
2467 		error = EINVAL;
2468 	if (error != 0)
2469 		goto drop;
2470 	VFS_KNOTE_UNLOCKED(vp, 0);
2471 	td->td_uretoff.tdu_off = offset;
2472 drop:
2473 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2474 	return (error);
2475 }
2476 
2477 int
2478 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2479     struct thread *td)
2480 {
2481 	int error;
2482 
2483 	/*
2484 	 * Grant permission if the caller is the owner of the file, or
2485 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2486 	 * on the file.  If the time pointer is null, then write
2487 	 * permission on the file is also sufficient.
2488 	 *
2489 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2490 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2491 	 * will be allowed to set the times [..] to the current
2492 	 * server time.
2493 	 */
2494 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2495 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2496 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2497 	return (error);
2498 }
2499 
2500 int
2501 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2502 {
2503 	struct vnode *vp;
2504 	int error;
2505 
2506 	if (fp->f_type == DTYPE_FIFO)
2507 		kif->kf_type = KF_TYPE_FIFO;
2508 	else
2509 		kif->kf_type = KF_TYPE_VNODE;
2510 	vp = fp->f_vnode;
2511 	vref(vp);
2512 	FILEDESC_SUNLOCK(fdp);
2513 	error = vn_fill_kinfo_vnode(vp, kif);
2514 	vrele(vp);
2515 	FILEDESC_SLOCK(fdp);
2516 	return (error);
2517 }
2518 
2519 static inline void
2520 vn_fill_junk(struct kinfo_file *kif)
2521 {
2522 	size_t len, olen;
2523 
2524 	/*
2525 	 * Simulate vn_fullpath returning changing values for a given
2526 	 * vp during e.g. coredump.
2527 	 */
2528 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2529 	olen = strlen(kif->kf_path);
2530 	if (len < olen)
2531 		strcpy(&kif->kf_path[len - 1], "$");
2532 	else
2533 		for (; olen < len; olen++)
2534 			strcpy(&kif->kf_path[olen], "A");
2535 }
2536 
2537 int
2538 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2539 {
2540 	struct vattr va;
2541 	char *fullpath, *freepath;
2542 	int error;
2543 
2544 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2545 	freepath = NULL;
2546 	fullpath = "-";
2547 	error = vn_fullpath(vp, &fullpath, &freepath);
2548 	if (error == 0) {
2549 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2550 	}
2551 	if (freepath != NULL)
2552 		free(freepath, M_TEMP);
2553 
2554 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2555 		vn_fill_junk(kif);
2556 	);
2557 
2558 	/*
2559 	 * Retrieve vnode attributes.
2560 	 */
2561 	va.va_fsid = VNOVAL;
2562 	va.va_rdev = NODEV;
2563 	vn_lock(vp, LK_SHARED | LK_RETRY);
2564 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2565 	VOP_UNLOCK(vp);
2566 	if (error != 0)
2567 		return (error);
2568 	if (va.va_fsid != VNOVAL)
2569 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2570 	else
2571 		kif->kf_un.kf_file.kf_file_fsid =
2572 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2573 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2574 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2575 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2576 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2577 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2578 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2579 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2580 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2581 	return (0);
2582 }
2583 
2584 int
2585 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2586     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2587     struct thread *td)
2588 {
2589 #ifdef HWPMC_HOOKS
2590 	struct pmckern_map_in pkm;
2591 #endif
2592 	struct mount *mp;
2593 	struct vnode *vp;
2594 	vm_object_t object;
2595 	vm_prot_t maxprot;
2596 	boolean_t writecounted;
2597 	int error;
2598 
2599 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2600     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2601 	/*
2602 	 * POSIX shared-memory objects are defined to have
2603 	 * kernel persistence, and are not defined to support
2604 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2605 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2606 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2607 	 * flag to request this behavior.
2608 	 */
2609 	if ((fp->f_flag & FPOSIXSHM) != 0)
2610 		flags |= MAP_NOSYNC;
2611 #endif
2612 	vp = fp->f_vnode;
2613 
2614 	/*
2615 	 * Ensure that file and memory protections are
2616 	 * compatible.  Note that we only worry about
2617 	 * writability if mapping is shared; in this case,
2618 	 * current and max prot are dictated by the open file.
2619 	 * XXX use the vnode instead?  Problem is: what
2620 	 * credentials do we use for determination? What if
2621 	 * proc does a setuid?
2622 	 */
2623 	mp = vp->v_mount;
2624 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2625 		maxprot = VM_PROT_NONE;
2626 		if ((prot & VM_PROT_EXECUTE) != 0)
2627 			return (EACCES);
2628 	} else
2629 		maxprot = VM_PROT_EXECUTE;
2630 	if ((fp->f_flag & FREAD) != 0)
2631 		maxprot |= VM_PROT_READ;
2632 	else if ((prot & VM_PROT_READ) != 0)
2633 		return (EACCES);
2634 
2635 	/*
2636 	 * If we are sharing potential changes via MAP_SHARED and we
2637 	 * are trying to get write permission although we opened it
2638 	 * without asking for it, bail out.
2639 	 */
2640 	if ((flags & MAP_SHARED) != 0) {
2641 		if ((fp->f_flag & FWRITE) != 0)
2642 			maxprot |= VM_PROT_WRITE;
2643 		else if ((prot & VM_PROT_WRITE) != 0)
2644 			return (EACCES);
2645 	} else {
2646 		maxprot |= VM_PROT_WRITE;
2647 		cap_maxprot |= VM_PROT_WRITE;
2648 	}
2649 	maxprot &= cap_maxprot;
2650 
2651 	/*
2652 	 * For regular files and shared memory, POSIX requires that
2653 	 * the value of foff be a legitimate offset within the data
2654 	 * object.  In particular, negative offsets are invalid.
2655 	 * Blocking negative offsets and overflows here avoids
2656 	 * possible wraparound or user-level access into reserved
2657 	 * ranges of the data object later.  In contrast, POSIX does
2658 	 * not dictate how offsets are used by device drivers, so in
2659 	 * the case of a device mapping a negative offset is passed
2660 	 * on.
2661 	 */
2662 	if (
2663 #ifdef _LP64
2664 	    size > OFF_MAX ||
2665 #endif
2666 	    foff > OFF_MAX - size)
2667 		return (EINVAL);
2668 
2669 	writecounted = FALSE;
2670 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2671 	    &foff, &object, &writecounted);
2672 	if (error != 0)
2673 		return (error);
2674 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2675 	    foff, writecounted, td);
2676 	if (error != 0) {
2677 		/*
2678 		 * If this mapping was accounted for in the vnode's
2679 		 * writecount, then undo that now.
2680 		 */
2681 		if (writecounted)
2682 			vm_pager_release_writecount(object, 0, size);
2683 		vm_object_deallocate(object);
2684 	}
2685 #ifdef HWPMC_HOOKS
2686 	/* Inform hwpmc(4) if an executable is being mapped. */
2687 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2688 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2689 			pkm.pm_file = vp;
2690 			pkm.pm_address = (uintptr_t) *addr;
2691 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2692 		}
2693 	}
2694 #endif
2695 	return (error);
2696 }
2697 
2698 void
2699 vn_fsid(struct vnode *vp, struct vattr *va)
2700 {
2701 	fsid_t *f;
2702 
2703 	f = &vp->v_mount->mnt_stat.f_fsid;
2704 	va->va_fsid = (uint32_t)f->val[1];
2705 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2706 	va->va_fsid += (uint32_t)f->val[0];
2707 }
2708 
2709 int
2710 vn_fsync_buf(struct vnode *vp, int waitfor)
2711 {
2712 	struct buf *bp, *nbp;
2713 	struct bufobj *bo;
2714 	struct mount *mp;
2715 	int error, maxretry;
2716 
2717 	error = 0;
2718 	maxretry = 10000;     /* large, arbitrarily chosen */
2719 	mp = NULL;
2720 	if (vp->v_type == VCHR) {
2721 		VI_LOCK(vp);
2722 		mp = vp->v_rdev->si_mountpt;
2723 		VI_UNLOCK(vp);
2724 	}
2725 	bo = &vp->v_bufobj;
2726 	BO_LOCK(bo);
2727 loop1:
2728 	/*
2729 	 * MARK/SCAN initialization to avoid infinite loops.
2730 	 */
2731         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2732 		bp->b_vflags &= ~BV_SCANNED;
2733 		bp->b_error = 0;
2734 	}
2735 
2736 	/*
2737 	 * Flush all dirty buffers associated with a vnode.
2738 	 */
2739 loop2:
2740 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2741 		if ((bp->b_vflags & BV_SCANNED) != 0)
2742 			continue;
2743 		bp->b_vflags |= BV_SCANNED;
2744 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2745 			if (waitfor != MNT_WAIT)
2746 				continue;
2747 			if (BUF_LOCK(bp,
2748 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2749 			    BO_LOCKPTR(bo)) != 0) {
2750 				BO_LOCK(bo);
2751 				goto loop1;
2752 			}
2753 			BO_LOCK(bo);
2754 		}
2755 		BO_UNLOCK(bo);
2756 		KASSERT(bp->b_bufobj == bo,
2757 		    ("bp %p wrong b_bufobj %p should be %p",
2758 		    bp, bp->b_bufobj, bo));
2759 		if ((bp->b_flags & B_DELWRI) == 0)
2760 			panic("fsync: not dirty");
2761 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2762 			vfs_bio_awrite(bp);
2763 		} else {
2764 			bremfree(bp);
2765 			bawrite(bp);
2766 		}
2767 		if (maxretry < 1000)
2768 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2769 		BO_LOCK(bo);
2770 		goto loop2;
2771 	}
2772 
2773 	/*
2774 	 * If synchronous the caller expects us to completely resolve all
2775 	 * dirty buffers in the system.  Wait for in-progress I/O to
2776 	 * complete (which could include background bitmap writes), then
2777 	 * retry if dirty blocks still exist.
2778 	 */
2779 	if (waitfor == MNT_WAIT) {
2780 		bufobj_wwait(bo, 0, 0);
2781 		if (bo->bo_dirty.bv_cnt > 0) {
2782 			/*
2783 			 * If we are unable to write any of these buffers
2784 			 * then we fail now rather than trying endlessly
2785 			 * to write them out.
2786 			 */
2787 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2788 				if ((error = bp->b_error) != 0)
2789 					break;
2790 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2791 			    (error == 0 && --maxretry >= 0))
2792 				goto loop1;
2793 			if (error == 0)
2794 				error = EAGAIN;
2795 		}
2796 	}
2797 	BO_UNLOCK(bo);
2798 	if (error != 0)
2799 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2800 
2801 	return (error);
2802 }
2803 
2804 /*
2805  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2806  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2807  * to do the actual copy.
2808  * vn_generic_copy_file_range() is factored out, so it can be called
2809  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2810  * different file systems.
2811  */
2812 int
2813 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2814     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2815     struct ucred *outcred, struct thread *fsize_td)
2816 {
2817 	int error;
2818 	size_t len;
2819 	uint64_t uval;
2820 
2821 	len = *lenp;
2822 	*lenp = 0;		/* For error returns. */
2823 	error = 0;
2824 
2825 	/* Do some sanity checks on the arguments. */
2826 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2827 		error = EISDIR;
2828 	else if (*inoffp < 0 || *outoffp < 0 ||
2829 	    invp->v_type != VREG || outvp->v_type != VREG)
2830 		error = EINVAL;
2831 	if (error != 0)
2832 		goto out;
2833 
2834 	/* Ensure offset + len does not wrap around. */
2835 	uval = *inoffp;
2836 	uval += len;
2837 	if (uval > INT64_MAX)
2838 		len = INT64_MAX - *inoffp;
2839 	uval = *outoffp;
2840 	uval += len;
2841 	if (uval > INT64_MAX)
2842 		len = INT64_MAX - *outoffp;
2843 	if (len == 0)
2844 		goto out;
2845 
2846 	/*
2847 	 * If the two vnode are for the same file system, call
2848 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2849 	 * which can handle copies across multiple file systems.
2850 	 */
2851 	*lenp = len;
2852 	if (invp->v_mount == outvp->v_mount)
2853 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2854 		    lenp, flags, incred, outcred, fsize_td);
2855 	else
2856 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2857 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2858 out:
2859 	return (error);
2860 }
2861 
2862 /*
2863  * Test len bytes of data starting at dat for all bytes == 0.
2864  * Return true if all bytes are zero, false otherwise.
2865  * Expects dat to be well aligned.
2866  */
2867 static bool
2868 mem_iszero(void *dat, int len)
2869 {
2870 	int i;
2871 	const u_int *p;
2872 	const char *cp;
2873 
2874 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2875 		if (len >= sizeof(*p)) {
2876 			if (*p != 0)
2877 				return (false);
2878 		} else {
2879 			cp = (const char *)p;
2880 			for (i = 0; i < len; i++, cp++)
2881 				if (*cp != '\0')
2882 					return (false);
2883 		}
2884 	}
2885 	return (true);
2886 }
2887 
2888 /*
2889  * Look for a hole in the output file and, if found, adjust *outoffp
2890  * and *xferp to skip past the hole.
2891  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2892  * to be written as 0's upon return.
2893  */
2894 static off_t
2895 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2896     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2897 {
2898 	int error;
2899 	off_t delta;
2900 
2901 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2902 		*dataoffp = *outoffp;
2903 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2904 		    curthread);
2905 		if (error == 0) {
2906 			*holeoffp = *dataoffp;
2907 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2908 			    curthread);
2909 		}
2910 		if (error != 0 || *holeoffp == *dataoffp) {
2911 			/*
2912 			 * Since outvp is unlocked, it may be possible for
2913 			 * another thread to do a truncate(), lseek(), write()
2914 			 * creating a hole at startoff between the above
2915 			 * VOP_IOCTL() calls, if the other thread does not do
2916 			 * rangelocking.
2917 			 * If that happens, *holeoffp == *dataoffp and finding
2918 			 * the hole has failed, so disable vn_skip_hole().
2919 			 */
2920 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2921 			return (xfer2);
2922 		}
2923 		KASSERT(*dataoffp >= *outoffp,
2924 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2925 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2926 		KASSERT(*holeoffp > *dataoffp,
2927 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2928 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2929 	}
2930 
2931 	/*
2932 	 * If there is a hole before the data starts, advance *outoffp and
2933 	 * *xferp past the hole.
2934 	 */
2935 	if (*dataoffp > *outoffp) {
2936 		delta = *dataoffp - *outoffp;
2937 		if (delta >= *xferp) {
2938 			/* Entire *xferp is a hole. */
2939 			*outoffp += *xferp;
2940 			*xferp = 0;
2941 			return (0);
2942 		}
2943 		*xferp -= delta;
2944 		*outoffp += delta;
2945 		xfer2 = MIN(xfer2, *xferp);
2946 	}
2947 
2948 	/*
2949 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2950 	 * that the write ends at the start of the hole.
2951 	 * *holeoffp should always be greater than *outoffp, but for the
2952 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2953 	 * value.
2954 	 */
2955 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2956 		xfer2 = *holeoffp - *outoffp;
2957 	return (xfer2);
2958 }
2959 
2960 /*
2961  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2962  * dat is a maximum of blksize in length and can be written repeatedly in
2963  * the chunk.
2964  * If growfile == true, just grow the file via vn_truncate_locked() instead
2965  * of doing actual writes.
2966  * If checkhole == true, a hole is being punched, so skip over any hole
2967  * already in the output file.
2968  */
2969 static int
2970 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2971     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2972 {
2973 	struct mount *mp;
2974 	off_t dataoff, holeoff, xfer2;
2975 	int error, lckf;
2976 
2977 	/*
2978 	 * Loop around doing writes of blksize until write has been completed.
2979 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2980 	 * done for each iteration, since the xfer argument can be very
2981 	 * large if there is a large hole to punch in the output file.
2982 	 */
2983 	error = 0;
2984 	holeoff = 0;
2985 	do {
2986 		xfer2 = MIN(xfer, blksize);
2987 		if (checkhole) {
2988 			/*
2989 			 * Punching a hole.  Skip writing if there is
2990 			 * already a hole in the output file.
2991 			 */
2992 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2993 			    &dataoff, &holeoff, cred);
2994 			if (xfer == 0)
2995 				break;
2996 			if (holeoff < 0)
2997 				checkhole = false;
2998 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2999 			    (intmax_t)xfer2));
3000 		}
3001 		bwillwrite();
3002 		mp = NULL;
3003 		error = vn_start_write(outvp, &mp, V_WAIT);
3004 		if (error != 0)
3005 			break;
3006 		if (growfile) {
3007 			error = vn_lock(outvp, LK_EXCLUSIVE);
3008 			if (error == 0) {
3009 				error = vn_truncate_locked(outvp, outoff + xfer,
3010 				    false, cred);
3011 				VOP_UNLOCK(outvp);
3012 			}
3013 		} else {
3014 			if (MNT_SHARED_WRITES(mp))
3015 				lckf = LK_SHARED;
3016 			else
3017 				lckf = LK_EXCLUSIVE;
3018 			error = vn_lock(outvp, lckf);
3019 			if (error == 0) {
3020 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3021 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3022 				    curthread->td_ucred, cred, NULL, curthread);
3023 				outoff += xfer2;
3024 				xfer -= xfer2;
3025 				VOP_UNLOCK(outvp);
3026 			}
3027 		}
3028 		if (mp != NULL)
3029 			vn_finished_write(mp);
3030 	} while (!growfile && xfer > 0 && error == 0);
3031 	return (error);
3032 }
3033 
3034 /*
3035  * Copy a byte range of one file to another.  This function can handle the
3036  * case where invp and outvp are on different file systems.
3037  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3038  * is no better file system specific way to do it.
3039  */
3040 int
3041 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3042     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3043     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3044 {
3045 	struct vattr va;
3046 	struct mount *mp;
3047 	struct uio io;
3048 	off_t startoff, endoff, xfer, xfer2;
3049 	u_long blksize;
3050 	int error, interrupted;
3051 	bool cantseek, readzeros, eof, lastblock;
3052 	ssize_t aresid;
3053 	size_t copylen, len, rem, savlen;
3054 	char *dat;
3055 	long holein, holeout;
3056 
3057 	holein = holeout = 0;
3058 	savlen = len = *lenp;
3059 	error = 0;
3060 	interrupted = 0;
3061 	dat = NULL;
3062 
3063 	error = vn_lock(invp, LK_SHARED);
3064 	if (error != 0)
3065 		goto out;
3066 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3067 		holein = 0;
3068 	VOP_UNLOCK(invp);
3069 
3070 	mp = NULL;
3071 	error = vn_start_write(outvp, &mp, V_WAIT);
3072 	if (error == 0)
3073 		error = vn_lock(outvp, LK_EXCLUSIVE);
3074 	if (error == 0) {
3075 		/*
3076 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3077 		 * now that outvp is locked.
3078 		 */
3079 		if (fsize_td != NULL) {
3080 			io.uio_offset = *outoffp;
3081 			io.uio_resid = len;
3082 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3083 			if (error != 0)
3084 				error = EFBIG;
3085 		}
3086 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3087 			holeout = 0;
3088 		/*
3089 		 * Holes that are past EOF do not need to be written as a block
3090 		 * of zero bytes.  So, truncate the output file as far as
3091 		 * possible and then use va.va_size to decide if writing 0
3092 		 * bytes is necessary in the loop below.
3093 		 */
3094 		if (error == 0)
3095 			error = VOP_GETATTR(outvp, &va, outcred);
3096 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3097 		    *outoffp + len) {
3098 #ifdef MAC
3099 			error = mac_vnode_check_write(curthread->td_ucred,
3100 			    outcred, outvp);
3101 			if (error == 0)
3102 #endif
3103 				error = vn_truncate_locked(outvp, *outoffp,
3104 				    false, outcred);
3105 			if (error == 0)
3106 				va.va_size = *outoffp;
3107 		}
3108 		VOP_UNLOCK(outvp);
3109 	}
3110 	if (mp != NULL)
3111 		vn_finished_write(mp);
3112 	if (error != 0)
3113 		goto out;
3114 
3115 	/*
3116 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3117 	 * If hole sizes aren't available, set the blksize to the larger
3118 	 * f_iosize of invp and outvp.
3119 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3120 	 * This value is clipped at 4Kbytes and 1Mbyte.
3121 	 */
3122 	blksize = MAX(holein, holeout);
3123 
3124 	/* Clip len to end at an exact multiple of hole size. */
3125 	if (blksize > 1) {
3126 		rem = *inoffp % blksize;
3127 		if (rem > 0)
3128 			rem = blksize - rem;
3129 		if (len - rem > blksize)
3130 			len = savlen = rounddown(len - rem, blksize) + rem;
3131 	}
3132 
3133 	if (blksize <= 1)
3134 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3135 		    outvp->v_mount->mnt_stat.f_iosize);
3136 	if (blksize < 4096)
3137 		blksize = 4096;
3138 	else if (blksize > 1024 * 1024)
3139 		blksize = 1024 * 1024;
3140 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3141 
3142 	/*
3143 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3144 	 * to find holes.  Otherwise, just scan the read block for all 0s
3145 	 * in the inner loop where the data copying is done.
3146 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3147 	 * support holes on the server, but do not support FIOSEEKHOLE.
3148 	 */
3149 	eof = false;
3150 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3151 		endoff = 0;			/* To shut up compilers. */
3152 		cantseek = true;
3153 		startoff = *inoffp;
3154 		copylen = len;
3155 
3156 		/*
3157 		 * Find the next data area.  If there is just a hole to EOF,
3158 		 * FIOSEEKDATA should fail and then we drop down into the
3159 		 * inner loop and create the hole on the outvp file.
3160 		 * (I do not know if any file system will report a hole to
3161 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3162 		 *  will fail for those file systems.)
3163 		 *
3164 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3165 		 * the code just falls through to the inner copy loop.
3166 		 */
3167 		error = EINVAL;
3168 		if (holein > 0)
3169 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3170 			    incred, curthread);
3171 		if (error == 0) {
3172 			endoff = startoff;
3173 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3174 			    incred, curthread);
3175 			/*
3176 			 * Since invp is unlocked, it may be possible for
3177 			 * another thread to do a truncate(), lseek(), write()
3178 			 * creating a hole at startoff between the above
3179 			 * VOP_IOCTL() calls, if the other thread does not do
3180 			 * rangelocking.
3181 			 * If that happens, startoff == endoff and finding
3182 			 * the hole has failed, so set an error.
3183 			 */
3184 			if (error == 0 && startoff == endoff)
3185 				error = EINVAL; /* Any error. Reset to 0. */
3186 		}
3187 		if (error == 0) {
3188 			if (startoff > *inoffp) {
3189 				/* Found hole before data block. */
3190 				xfer = MIN(startoff - *inoffp, len);
3191 				if (*outoffp < va.va_size) {
3192 					/* Must write 0s to punch hole. */
3193 					xfer2 = MIN(va.va_size - *outoffp,
3194 					    xfer);
3195 					memset(dat, 0, MIN(xfer2, blksize));
3196 					error = vn_write_outvp(outvp, dat,
3197 					    *outoffp, xfer2, blksize, false,
3198 					    holeout > 0, outcred);
3199 				}
3200 
3201 				if (error == 0 && *outoffp + xfer >
3202 				    va.va_size && xfer == len)
3203 					/* Grow last block. */
3204 					error = vn_write_outvp(outvp, dat,
3205 					    *outoffp, xfer, blksize, true,
3206 					    false, outcred);
3207 				if (error == 0) {
3208 					*inoffp += xfer;
3209 					*outoffp += xfer;
3210 					len -= xfer;
3211 					if (len < savlen)
3212 						interrupted = sig_intr();
3213 				}
3214 			}
3215 			copylen = MIN(len, endoff - startoff);
3216 			cantseek = false;
3217 		} else {
3218 			cantseek = true;
3219 			startoff = *inoffp;
3220 			copylen = len;
3221 			error = 0;
3222 		}
3223 
3224 		xfer = blksize;
3225 		if (cantseek) {
3226 			/*
3227 			 * Set first xfer to end at a block boundary, so that
3228 			 * holes are more likely detected in the loop below via
3229 			 * the for all bytes 0 method.
3230 			 */
3231 			xfer -= (*inoffp % blksize);
3232 		}
3233 		/* Loop copying the data block. */
3234 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3235 			if (copylen < xfer)
3236 				xfer = copylen;
3237 			error = vn_lock(invp, LK_SHARED);
3238 			if (error != 0)
3239 				goto out;
3240 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3241 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3242 			    curthread->td_ucred, incred, &aresid,
3243 			    curthread);
3244 			VOP_UNLOCK(invp);
3245 			lastblock = false;
3246 			if (error == 0 && aresid > 0) {
3247 				/* Stop the copy at EOF on the input file. */
3248 				xfer -= aresid;
3249 				eof = true;
3250 				lastblock = true;
3251 			}
3252 			if (error == 0) {
3253 				/*
3254 				 * Skip the write for holes past the initial EOF
3255 				 * of the output file, unless this is the last
3256 				 * write of the output file at EOF.
3257 				 */
3258 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3259 				    false;
3260 				if (xfer == len)
3261 					lastblock = true;
3262 				if (!cantseek || *outoffp < va.va_size ||
3263 				    lastblock || !readzeros)
3264 					error = vn_write_outvp(outvp, dat,
3265 					    *outoffp, xfer, blksize,
3266 					    readzeros && lastblock &&
3267 					    *outoffp >= va.va_size, false,
3268 					    outcred);
3269 				if (error == 0) {
3270 					*inoffp += xfer;
3271 					startoff += xfer;
3272 					*outoffp += xfer;
3273 					copylen -= xfer;
3274 					len -= xfer;
3275 					if (len < savlen)
3276 						interrupted = sig_intr();
3277 				}
3278 			}
3279 			xfer = blksize;
3280 		}
3281 	}
3282 out:
3283 	*lenp = savlen - len;
3284 	free(dat, M_TEMP);
3285 	return (error);
3286 }
3287 
3288 static int
3289 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3290 {
3291 	struct mount *mp;
3292 	struct vnode *vp;
3293 	off_t olen, ooffset;
3294 	int error;
3295 #ifdef AUDIT
3296 	int audited_vnode1 = 0;
3297 #endif
3298 
3299 	vp = fp->f_vnode;
3300 	if (vp->v_type != VREG)
3301 		return (ENODEV);
3302 
3303 	/* Allocating blocks may take a long time, so iterate. */
3304 	for (;;) {
3305 		olen = len;
3306 		ooffset = offset;
3307 
3308 		bwillwrite();
3309 		mp = NULL;
3310 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3311 		if (error != 0)
3312 			break;
3313 		error = vn_lock(vp, LK_EXCLUSIVE);
3314 		if (error != 0) {
3315 			vn_finished_write(mp);
3316 			break;
3317 		}
3318 #ifdef AUDIT
3319 		if (!audited_vnode1) {
3320 			AUDIT_ARG_VNODE1(vp);
3321 			audited_vnode1 = 1;
3322 		}
3323 #endif
3324 #ifdef MAC
3325 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3326 		if (error == 0)
3327 #endif
3328 			error = VOP_ALLOCATE(vp, &offset, &len);
3329 		VOP_UNLOCK(vp);
3330 		vn_finished_write(mp);
3331 
3332 		if (olen + ooffset != offset + len) {
3333 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3334 			    ooffset, olen, offset, len);
3335 		}
3336 		if (error != 0 || len == 0)
3337 			break;
3338 		KASSERT(olen > len, ("Iteration did not make progress?"));
3339 		maybe_yield();
3340 	}
3341 
3342 	return (error);
3343 }
3344 
3345 static u_long vn_lock_pair_pause_cnt;
3346 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3347     &vn_lock_pair_pause_cnt, 0,
3348     "Count of vn_lock_pair deadlocks");
3349 
3350 u_int vn_lock_pair_pause_max;
3351 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3352     &vn_lock_pair_pause_max, 0,
3353     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3354 
3355 static void
3356 vn_lock_pair_pause(const char *wmesg)
3357 {
3358 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3359 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3360 }
3361 
3362 /*
3363  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3364  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3365  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3366  * can be NULL.
3367  *
3368  * The function returns with both vnodes exclusively locked, and
3369  * guarantees that it does not create lock order reversal with other
3370  * threads during its execution.  Both vnodes could be unlocked
3371  * temporary (and reclaimed).
3372  */
3373 void
3374 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3375     bool vp2_locked)
3376 {
3377 	int error;
3378 
3379 	if (vp1 == NULL && vp2 == NULL)
3380 		return;
3381 	if (vp1 != NULL) {
3382 		if (vp1_locked)
3383 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3384 		else
3385 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3386 	} else {
3387 		vp1_locked = true;
3388 	}
3389 	if (vp2 != NULL) {
3390 		if (vp2_locked)
3391 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3392 		else
3393 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3394 	} else {
3395 		vp2_locked = true;
3396 	}
3397 	if (!vp1_locked && !vp2_locked) {
3398 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3399 		vp1_locked = true;
3400 	}
3401 
3402 	for (;;) {
3403 		if (vp1_locked && vp2_locked)
3404 			break;
3405 		if (vp1_locked && vp2 != NULL) {
3406 			if (vp1 != NULL) {
3407 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3408 				    __FILE__, __LINE__);
3409 				if (error == 0)
3410 					break;
3411 				VOP_UNLOCK(vp1);
3412 				vp1_locked = false;
3413 				vn_lock_pair_pause("vlp1");
3414 			}
3415 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3416 			vp2_locked = true;
3417 		}
3418 		if (vp2_locked && vp1 != NULL) {
3419 			if (vp2 != NULL) {
3420 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3421 				    __FILE__, __LINE__);
3422 				if (error == 0)
3423 					break;
3424 				VOP_UNLOCK(vp2);
3425 				vp2_locked = false;
3426 				vn_lock_pair_pause("vlp2");
3427 			}
3428 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3429 			vp1_locked = true;
3430 		}
3431 	}
3432 	if (vp1 != NULL)
3433 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3434 	if (vp2 != NULL)
3435 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3436 }
3437