xref: /freebsd/sys/kern/vfs_vnops.c (revision 90b5fc95832da64a5f56295e687379732c33718f)
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 + resid > vsz)
954 		resid = vsz - uio->uio_offset;
955 
956 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
957 
958 out:
959 	for (j = 0; j < i; j++) {
960 		if (error == 0)
961 			vm_page_reference(ma[j]);
962 		vm_page_sunbusy(ma[j]);
963 	}
964 out_pip:
965 	vm_object_pip_wakeup(obj);
966 	if (error != 0)
967 		return (error);
968 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
969 }
970 
971 /*
972  * File table vnode read routine.
973  */
974 static int
975 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
976     struct thread *td)
977 {
978 	struct vnode *vp;
979 	off_t orig_offset;
980 	int error, ioflag;
981 	int advice;
982 
983 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
984 	    uio->uio_td, td));
985 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
986 	vp = fp->f_vnode;
987 	ioflag = 0;
988 	if (fp->f_flag & FNONBLOCK)
989 		ioflag |= IO_NDELAY;
990 	if (fp->f_flag & O_DIRECT)
991 		ioflag |= IO_DIRECT;
992 
993 	/*
994 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
995 	 * allows us to avoid unneeded work outright.
996 	 */
997 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
998 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
999 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1000 		if (error == 0) {
1001 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1002 			return (0);
1003 		}
1004 		if (error != EJUSTRETURN)
1005 			return (error);
1006 	}
1007 
1008 	advice = get_advice(fp, uio);
1009 	vn_lock(vp, LK_SHARED | LK_RETRY);
1010 
1011 	switch (advice) {
1012 	case POSIX_FADV_NORMAL:
1013 	case POSIX_FADV_SEQUENTIAL:
1014 	case POSIX_FADV_NOREUSE:
1015 		ioflag |= sequential_heuristic(uio, fp);
1016 		break;
1017 	case POSIX_FADV_RANDOM:
1018 		/* Disable read-ahead for random I/O. */
1019 		break;
1020 	}
1021 	orig_offset = uio->uio_offset;
1022 
1023 #ifdef MAC
1024 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1025 	if (error == 0)
1026 #endif
1027 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1028 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1029 	VOP_UNLOCK(vp);
1030 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1031 	    orig_offset != uio->uio_offset)
1032 		/*
1033 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1034 		 * for the backing file after a POSIX_FADV_NOREUSE
1035 		 * read(2).
1036 		 */
1037 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1038 		    POSIX_FADV_DONTNEED);
1039 	return (error);
1040 }
1041 
1042 /*
1043  * File table vnode write routine.
1044  */
1045 static int
1046 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1047     struct thread *td)
1048 {
1049 	struct vnode *vp;
1050 	struct mount *mp;
1051 	off_t orig_offset;
1052 	int error, ioflag, lock_flags;
1053 	int advice;
1054 
1055 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1056 	    uio->uio_td, td));
1057 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1058 	vp = fp->f_vnode;
1059 	if (vp->v_type == VREG)
1060 		bwillwrite();
1061 	ioflag = IO_UNIT;
1062 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1063 		ioflag |= IO_APPEND;
1064 	if (fp->f_flag & FNONBLOCK)
1065 		ioflag |= IO_NDELAY;
1066 	if (fp->f_flag & O_DIRECT)
1067 		ioflag |= IO_DIRECT;
1068 	if ((fp->f_flag & O_FSYNC) ||
1069 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1070 		ioflag |= IO_SYNC;
1071 	/*
1072 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1073 	 * implementations that don't understand IO_DATASYNC fall back to full
1074 	 * O_SYNC behavior.
1075 	 */
1076 	if (fp->f_flag & O_DSYNC)
1077 		ioflag |= IO_SYNC | IO_DATASYNC;
1078 	mp = NULL;
1079 	if (vp->v_type != VCHR &&
1080 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1081 		goto unlock;
1082 
1083 	advice = get_advice(fp, uio);
1084 
1085 	if (MNT_SHARED_WRITES(mp) ||
1086 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1087 		lock_flags = LK_SHARED;
1088 	} else {
1089 		lock_flags = LK_EXCLUSIVE;
1090 	}
1091 
1092 	vn_lock(vp, lock_flags | LK_RETRY);
1093 	switch (advice) {
1094 	case POSIX_FADV_NORMAL:
1095 	case POSIX_FADV_SEQUENTIAL:
1096 	case POSIX_FADV_NOREUSE:
1097 		ioflag |= sequential_heuristic(uio, fp);
1098 		break;
1099 	case POSIX_FADV_RANDOM:
1100 		/* XXX: Is this correct? */
1101 		break;
1102 	}
1103 	orig_offset = uio->uio_offset;
1104 
1105 #ifdef MAC
1106 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1107 	if (error == 0)
1108 #endif
1109 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1110 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1111 	VOP_UNLOCK(vp);
1112 	if (vp->v_type != VCHR)
1113 		vn_finished_write(mp);
1114 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1115 	    orig_offset != uio->uio_offset)
1116 		/*
1117 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1118 		 * for the backing file after a POSIX_FADV_NOREUSE
1119 		 * write(2).
1120 		 */
1121 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1122 		    POSIX_FADV_DONTNEED);
1123 unlock:
1124 	return (error);
1125 }
1126 
1127 /*
1128  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1129  * prevent the following deadlock:
1130  *
1131  * Assume that the thread A reads from the vnode vp1 into userspace
1132  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1133  * currently not resident, then system ends up with the call chain
1134  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1135  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1136  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1137  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1138  * backed by the pages of vnode vp1, and some page in buf2 is not
1139  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1140  *
1141  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1142  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1143  * Instead, it first tries to do the whole range i/o with pagefaults
1144  * disabled. If all pages in the i/o buffer are resident and mapped,
1145  * VOP will succeed (ignoring the genuine filesystem errors).
1146  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1147  * i/o in chunks, with all pages in the chunk prefaulted and held
1148  * using vm_fault_quick_hold_pages().
1149  *
1150  * Filesystems using this deadlock avoidance scheme should use the
1151  * array of the held pages from uio, saved in the curthread->td_ma,
1152  * instead of doing uiomove().  A helper function
1153  * vn_io_fault_uiomove() converts uiomove request into
1154  * uiomove_fromphys() over td_ma array.
1155  *
1156  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1157  * make the current i/o request atomic with respect to other i/os and
1158  * truncations.
1159  */
1160 
1161 /*
1162  * Decode vn_io_fault_args and perform the corresponding i/o.
1163  */
1164 static int
1165 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1166     struct thread *td)
1167 {
1168 	int error, save;
1169 
1170 	error = 0;
1171 	save = vm_fault_disable_pagefaults();
1172 	switch (args->kind) {
1173 	case VN_IO_FAULT_FOP:
1174 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1175 		    uio, args->cred, args->flags, td);
1176 		break;
1177 	case VN_IO_FAULT_VOP:
1178 		if (uio->uio_rw == UIO_READ) {
1179 			error = VOP_READ(args->args.vop_args.vp, uio,
1180 			    args->flags, args->cred);
1181 		} else if (uio->uio_rw == UIO_WRITE) {
1182 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1183 			    args->flags, args->cred);
1184 		}
1185 		break;
1186 	default:
1187 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1188 		    args->kind, uio->uio_rw);
1189 	}
1190 	vm_fault_enable_pagefaults(save);
1191 	return (error);
1192 }
1193 
1194 static int
1195 vn_io_fault_touch(char *base, const struct uio *uio)
1196 {
1197 	int r;
1198 
1199 	r = fubyte(base);
1200 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1201 		return (EFAULT);
1202 	return (0);
1203 }
1204 
1205 static int
1206 vn_io_fault_prefault_user(const struct uio *uio)
1207 {
1208 	char *base;
1209 	const struct iovec *iov;
1210 	size_t len;
1211 	ssize_t resid;
1212 	int error, i;
1213 
1214 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1215 	    ("vn_io_fault_prefault userspace"));
1216 
1217 	error = i = 0;
1218 	iov = uio->uio_iov;
1219 	resid = uio->uio_resid;
1220 	base = iov->iov_base;
1221 	len = iov->iov_len;
1222 	while (resid > 0) {
1223 		error = vn_io_fault_touch(base, uio);
1224 		if (error != 0)
1225 			break;
1226 		if (len < PAGE_SIZE) {
1227 			if (len != 0) {
1228 				error = vn_io_fault_touch(base + len - 1, uio);
1229 				if (error != 0)
1230 					break;
1231 				resid -= len;
1232 			}
1233 			if (++i >= uio->uio_iovcnt)
1234 				break;
1235 			iov = uio->uio_iov + i;
1236 			base = iov->iov_base;
1237 			len = iov->iov_len;
1238 		} else {
1239 			len -= PAGE_SIZE;
1240 			base += PAGE_SIZE;
1241 			resid -= PAGE_SIZE;
1242 		}
1243 	}
1244 	return (error);
1245 }
1246 
1247 /*
1248  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1249  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1250  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1251  * into args and call vn_io_fault1() to handle faults during the user
1252  * mode buffer accesses.
1253  */
1254 static int
1255 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1256     struct thread *td)
1257 {
1258 	vm_page_t ma[io_hold_cnt + 2];
1259 	struct uio *uio_clone, short_uio;
1260 	struct iovec short_iovec[1];
1261 	vm_page_t *prev_td_ma;
1262 	vm_prot_t prot;
1263 	vm_offset_t addr, end;
1264 	size_t len, resid;
1265 	ssize_t adv;
1266 	int error, cnt, saveheld, prev_td_ma_cnt;
1267 
1268 	if (vn_io_fault_prefault) {
1269 		error = vn_io_fault_prefault_user(uio);
1270 		if (error != 0)
1271 			return (error); /* Or ignore ? */
1272 	}
1273 
1274 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1275 
1276 	/*
1277 	 * The UFS follows IO_UNIT directive and replays back both
1278 	 * uio_offset and uio_resid if an error is encountered during the
1279 	 * operation.  But, since the iovec may be already advanced,
1280 	 * uio is still in an inconsistent state.
1281 	 *
1282 	 * Cache a copy of the original uio, which is advanced to the redo
1283 	 * point using UIO_NOCOPY below.
1284 	 */
1285 	uio_clone = cloneuio(uio);
1286 	resid = uio->uio_resid;
1287 
1288 	short_uio.uio_segflg = UIO_USERSPACE;
1289 	short_uio.uio_rw = uio->uio_rw;
1290 	short_uio.uio_td = uio->uio_td;
1291 
1292 	error = vn_io_fault_doio(args, uio, td);
1293 	if (error != EFAULT)
1294 		goto out;
1295 
1296 	atomic_add_long(&vn_io_faults_cnt, 1);
1297 	uio_clone->uio_segflg = UIO_NOCOPY;
1298 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1299 	uio_clone->uio_segflg = uio->uio_segflg;
1300 
1301 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1302 	prev_td_ma = td->td_ma;
1303 	prev_td_ma_cnt = td->td_ma_cnt;
1304 
1305 	while (uio_clone->uio_resid != 0) {
1306 		len = uio_clone->uio_iov->iov_len;
1307 		if (len == 0) {
1308 			KASSERT(uio_clone->uio_iovcnt >= 1,
1309 			    ("iovcnt underflow"));
1310 			uio_clone->uio_iov++;
1311 			uio_clone->uio_iovcnt--;
1312 			continue;
1313 		}
1314 		if (len > ptoa(io_hold_cnt))
1315 			len = ptoa(io_hold_cnt);
1316 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1317 		end = round_page(addr + len);
1318 		if (end < addr) {
1319 			error = EFAULT;
1320 			break;
1321 		}
1322 		cnt = atop(end - trunc_page(addr));
1323 		/*
1324 		 * A perfectly misaligned address and length could cause
1325 		 * both the start and the end of the chunk to use partial
1326 		 * page.  +2 accounts for such a situation.
1327 		 */
1328 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1329 		    addr, len, prot, ma, io_hold_cnt + 2);
1330 		if (cnt == -1) {
1331 			error = EFAULT;
1332 			break;
1333 		}
1334 		short_uio.uio_iov = &short_iovec[0];
1335 		short_iovec[0].iov_base = (void *)addr;
1336 		short_uio.uio_iovcnt = 1;
1337 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1338 		short_uio.uio_offset = uio_clone->uio_offset;
1339 		td->td_ma = ma;
1340 		td->td_ma_cnt = cnt;
1341 
1342 		error = vn_io_fault_doio(args, &short_uio, td);
1343 		vm_page_unhold_pages(ma, cnt);
1344 		adv = len - short_uio.uio_resid;
1345 
1346 		uio_clone->uio_iov->iov_base =
1347 		    (char *)uio_clone->uio_iov->iov_base + adv;
1348 		uio_clone->uio_iov->iov_len -= adv;
1349 		uio_clone->uio_resid -= adv;
1350 		uio_clone->uio_offset += adv;
1351 
1352 		uio->uio_resid -= adv;
1353 		uio->uio_offset += adv;
1354 
1355 		if (error != 0 || adv == 0)
1356 			break;
1357 	}
1358 	td->td_ma = prev_td_ma;
1359 	td->td_ma_cnt = prev_td_ma_cnt;
1360 	curthread_pflags_restore(saveheld);
1361 out:
1362 	free(uio_clone, M_IOV);
1363 	return (error);
1364 }
1365 
1366 static int
1367 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1368     int flags, struct thread *td)
1369 {
1370 	fo_rdwr_t *doio;
1371 	struct vnode *vp;
1372 	void *rl_cookie;
1373 	struct vn_io_fault_args args;
1374 	int error;
1375 
1376 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1377 	vp = fp->f_vnode;
1378 
1379 	/*
1380 	 * The ability to read(2) on a directory has historically been
1381 	 * allowed for all users, but this can and has been the source of
1382 	 * at least one security issue in the past.  As such, it is now hidden
1383 	 * away behind a sysctl for those that actually need it to use it, and
1384 	 * restricted to root when it's turned on to make it relatively safe to
1385 	 * leave on for longer sessions of need.
1386 	 */
1387 	if (vp->v_type == VDIR) {
1388 		KASSERT(uio->uio_rw == UIO_READ,
1389 		    ("illegal write attempted on a directory"));
1390 		if (!vfs_allow_read_dir)
1391 			return (EISDIR);
1392 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1393 			return (EISDIR);
1394 	}
1395 
1396 	foffset_lock_uio(fp, uio, flags);
1397 	if (do_vn_io_fault(vp, uio)) {
1398 		args.kind = VN_IO_FAULT_FOP;
1399 		args.args.fop_args.fp = fp;
1400 		args.args.fop_args.doio = doio;
1401 		args.cred = active_cred;
1402 		args.flags = flags | FOF_OFFSET;
1403 		if (uio->uio_rw == UIO_READ) {
1404 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1405 			    uio->uio_offset + uio->uio_resid);
1406 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1407 		    (flags & FOF_OFFSET) == 0) {
1408 			/* For appenders, punt and lock the whole range. */
1409 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1410 		} else {
1411 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1412 			    uio->uio_offset + uio->uio_resid);
1413 		}
1414 		error = vn_io_fault1(vp, uio, &args, td);
1415 		vn_rangelock_unlock(vp, rl_cookie);
1416 	} else {
1417 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1418 	}
1419 	foffset_unlock_uio(fp, uio, flags);
1420 	return (error);
1421 }
1422 
1423 /*
1424  * Helper function to perform the requested uiomove operation using
1425  * the held pages for io->uio_iov[0].iov_base buffer instead of
1426  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1427  * instead of iov_base prevents page faults that could occur due to
1428  * pmap_collect() invalidating the mapping created by
1429  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1430  * object cleanup revoking the write access from page mappings.
1431  *
1432  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1433  * instead of plain uiomove().
1434  */
1435 int
1436 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1437 {
1438 	struct uio transp_uio;
1439 	struct iovec transp_iov[1];
1440 	struct thread *td;
1441 	size_t adv;
1442 	int error, pgadv;
1443 
1444 	td = curthread;
1445 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1446 	    uio->uio_segflg != UIO_USERSPACE)
1447 		return (uiomove(data, xfersize, uio));
1448 
1449 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1450 	transp_iov[0].iov_base = data;
1451 	transp_uio.uio_iov = &transp_iov[0];
1452 	transp_uio.uio_iovcnt = 1;
1453 	if (xfersize > uio->uio_resid)
1454 		xfersize = uio->uio_resid;
1455 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1456 	transp_uio.uio_offset = 0;
1457 	transp_uio.uio_segflg = UIO_SYSSPACE;
1458 	/*
1459 	 * Since transp_iov points to data, and td_ma page array
1460 	 * corresponds to original uio->uio_iov, we need to invert the
1461 	 * direction of the i/o operation as passed to
1462 	 * uiomove_fromphys().
1463 	 */
1464 	switch (uio->uio_rw) {
1465 	case UIO_WRITE:
1466 		transp_uio.uio_rw = UIO_READ;
1467 		break;
1468 	case UIO_READ:
1469 		transp_uio.uio_rw = UIO_WRITE;
1470 		break;
1471 	}
1472 	transp_uio.uio_td = uio->uio_td;
1473 	error = uiomove_fromphys(td->td_ma,
1474 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1475 	    xfersize, &transp_uio);
1476 	adv = xfersize - transp_uio.uio_resid;
1477 	pgadv =
1478 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1479 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1480 	td->td_ma += pgadv;
1481 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1482 	    pgadv));
1483 	td->td_ma_cnt -= pgadv;
1484 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1485 	uio->uio_iov->iov_len -= adv;
1486 	uio->uio_resid -= adv;
1487 	uio->uio_offset += adv;
1488 	return (error);
1489 }
1490 
1491 int
1492 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1493     struct uio *uio)
1494 {
1495 	struct thread *td;
1496 	vm_offset_t iov_base;
1497 	int cnt, pgadv;
1498 
1499 	td = curthread;
1500 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1501 	    uio->uio_segflg != UIO_USERSPACE)
1502 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1503 
1504 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1505 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1506 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1507 	switch (uio->uio_rw) {
1508 	case UIO_WRITE:
1509 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1510 		    offset, cnt);
1511 		break;
1512 	case UIO_READ:
1513 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1514 		    cnt);
1515 		break;
1516 	}
1517 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1518 	td->td_ma += pgadv;
1519 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1520 	    pgadv));
1521 	td->td_ma_cnt -= pgadv;
1522 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1523 	uio->uio_iov->iov_len -= cnt;
1524 	uio->uio_resid -= cnt;
1525 	uio->uio_offset += cnt;
1526 	return (0);
1527 }
1528 
1529 /*
1530  * File table truncate routine.
1531  */
1532 static int
1533 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1534     struct thread *td)
1535 {
1536 	struct mount *mp;
1537 	struct vnode *vp;
1538 	void *rl_cookie;
1539 	int error;
1540 
1541 	vp = fp->f_vnode;
1542 
1543 retry:
1544 	/*
1545 	 * Lock the whole range for truncation.  Otherwise split i/o
1546 	 * might happen partly before and partly after the truncation.
1547 	 */
1548 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1549 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1550 	if (error)
1551 		goto out1;
1552 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1553 	AUDIT_ARG_VNODE1(vp);
1554 	if (vp->v_type == VDIR) {
1555 		error = EISDIR;
1556 		goto out;
1557 	}
1558 #ifdef MAC
1559 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1560 	if (error)
1561 		goto out;
1562 #endif
1563 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1564 	    fp->f_cred);
1565 out:
1566 	VOP_UNLOCK(vp);
1567 	vn_finished_write(mp);
1568 out1:
1569 	vn_rangelock_unlock(vp, rl_cookie);
1570 	if (error == ERELOOKUP)
1571 		goto retry;
1572 	return (error);
1573 }
1574 
1575 /*
1576  * Truncate a file that is already locked.
1577  */
1578 int
1579 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1580     struct ucred *cred)
1581 {
1582 	struct vattr vattr;
1583 	int error;
1584 
1585 	error = VOP_ADD_WRITECOUNT(vp, 1);
1586 	if (error == 0) {
1587 		VATTR_NULL(&vattr);
1588 		vattr.va_size = length;
1589 		if (sync)
1590 			vattr.va_vaflags |= VA_SYNC;
1591 		error = VOP_SETATTR(vp, &vattr, cred);
1592 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1593 	}
1594 	return (error);
1595 }
1596 
1597 /*
1598  * File table vnode stat routine.
1599  */
1600 static int
1601 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1602     struct thread *td)
1603 {
1604 	struct vnode *vp = fp->f_vnode;
1605 	int error;
1606 
1607 	vn_lock(vp, LK_SHARED | LK_RETRY);
1608 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1609 	VOP_UNLOCK(vp);
1610 
1611 	return (error);
1612 }
1613 
1614 /*
1615  * File table vnode ioctl routine.
1616  */
1617 static int
1618 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1619     struct thread *td)
1620 {
1621 	struct vattr vattr;
1622 	struct vnode *vp;
1623 	struct fiobmap2_arg *bmarg;
1624 	int error;
1625 
1626 	vp = fp->f_vnode;
1627 	switch (vp->v_type) {
1628 	case VDIR:
1629 	case VREG:
1630 		switch (com) {
1631 		case FIONREAD:
1632 			vn_lock(vp, LK_SHARED | LK_RETRY);
1633 			error = VOP_GETATTR(vp, &vattr, active_cred);
1634 			VOP_UNLOCK(vp);
1635 			if (error == 0)
1636 				*(int *)data = vattr.va_size - fp->f_offset;
1637 			return (error);
1638 		case FIOBMAP2:
1639 			bmarg = (struct fiobmap2_arg *)data;
1640 			vn_lock(vp, LK_SHARED | LK_RETRY);
1641 #ifdef MAC
1642 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1643 			    vp);
1644 			if (error == 0)
1645 #endif
1646 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1647 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1648 			VOP_UNLOCK(vp);
1649 			return (error);
1650 		case FIONBIO:
1651 		case FIOASYNC:
1652 			return (0);
1653 		default:
1654 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1655 			    active_cred, td));
1656 		}
1657 		break;
1658 	case VCHR:
1659 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1660 		    active_cred, td));
1661 	default:
1662 		return (ENOTTY);
1663 	}
1664 }
1665 
1666 /*
1667  * File table vnode poll routine.
1668  */
1669 static int
1670 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1671     struct thread *td)
1672 {
1673 	struct vnode *vp;
1674 	int error;
1675 
1676 	vp = fp->f_vnode;
1677 #if defined(MAC) || defined(AUDIT)
1678 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1679 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1680 		AUDIT_ARG_VNODE1(vp);
1681 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1682 		VOP_UNLOCK(vp);
1683 		if (error != 0)
1684 			return (error);
1685 	}
1686 #endif
1687 	error = VOP_POLL(vp, events, fp->f_cred, td);
1688 	return (error);
1689 }
1690 
1691 /*
1692  * Acquire the requested lock and then check for validity.  LK_RETRY
1693  * permits vn_lock to return doomed vnodes.
1694  */
1695 static int __noinline
1696 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1697     int error)
1698 {
1699 
1700 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1701 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1702 
1703 	if (error == 0)
1704 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1705 
1706 	if ((flags & LK_RETRY) == 0) {
1707 		if (error == 0) {
1708 			VOP_UNLOCK(vp);
1709 			error = ENOENT;
1710 		}
1711 		return (error);
1712 	}
1713 
1714 	/*
1715 	 * LK_RETRY case.
1716 	 *
1717 	 * Nothing to do if we got the lock.
1718 	 */
1719 	if (error == 0)
1720 		return (0);
1721 
1722 	/*
1723 	 * Interlock was dropped by the call in _vn_lock.
1724 	 */
1725 	flags &= ~LK_INTERLOCK;
1726 	do {
1727 		error = VOP_LOCK1(vp, flags, file, line);
1728 	} while (error != 0);
1729 	return (0);
1730 }
1731 
1732 int
1733 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1734 {
1735 	int error;
1736 
1737 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1738 	    ("vn_lock: no locktype (%d passed)", flags));
1739 	VNPASS(vp->v_holdcnt > 0, vp);
1740 	error = VOP_LOCK1(vp, flags, file, line);
1741 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1742 		return (_vn_lock_fallback(vp, flags, file, line, error));
1743 	return (0);
1744 }
1745 
1746 /*
1747  * File table vnode close routine.
1748  */
1749 static int
1750 vn_closefile(struct file *fp, struct thread *td)
1751 {
1752 	struct vnode *vp;
1753 	struct flock lf;
1754 	int error;
1755 	bool ref;
1756 
1757 	vp = fp->f_vnode;
1758 	fp->f_ops = &badfileops;
1759 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1760 
1761 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1762 
1763 	if (__predict_false(ref)) {
1764 		lf.l_whence = SEEK_SET;
1765 		lf.l_start = 0;
1766 		lf.l_len = 0;
1767 		lf.l_type = F_UNLCK;
1768 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1769 		vrele(vp);
1770 	}
1771 	return (error);
1772 }
1773 
1774 /*
1775  * Preparing to start a filesystem write operation. If the operation is
1776  * permitted, then we bump the count of operations in progress and
1777  * proceed. If a suspend request is in progress, we wait until the
1778  * suspension is over, and then proceed.
1779  */
1780 static int
1781 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1782 {
1783 	struct mount_pcpu *mpcpu;
1784 	int error, mflags;
1785 
1786 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1787 	    vfs_op_thread_enter(mp, mpcpu)) {
1788 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1789 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1790 		vfs_op_thread_exit(mp, mpcpu);
1791 		return (0);
1792 	}
1793 
1794 	if (mplocked)
1795 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1796 	else
1797 		MNT_ILOCK(mp);
1798 
1799 	error = 0;
1800 
1801 	/*
1802 	 * Check on status of suspension.
1803 	 */
1804 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1805 	    mp->mnt_susp_owner != curthread) {
1806 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1807 		    (flags & PCATCH) : 0) | (PUSER - 1);
1808 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1809 			if (flags & V_NOWAIT) {
1810 				error = EWOULDBLOCK;
1811 				goto unlock;
1812 			}
1813 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1814 			    "suspfs", 0);
1815 			if (error)
1816 				goto unlock;
1817 		}
1818 	}
1819 	if (flags & V_XSLEEP)
1820 		goto unlock;
1821 	mp->mnt_writeopcount++;
1822 unlock:
1823 	if (error != 0 || (flags & V_XSLEEP) != 0)
1824 		MNT_REL(mp);
1825 	MNT_IUNLOCK(mp);
1826 	return (error);
1827 }
1828 
1829 int
1830 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1831 {
1832 	struct mount *mp;
1833 	int error;
1834 
1835 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1836 	    ("V_MNTREF requires mp"));
1837 
1838 	error = 0;
1839 	/*
1840 	 * If a vnode is provided, get and return the mount point that
1841 	 * to which it will write.
1842 	 */
1843 	if (vp != NULL) {
1844 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1845 			*mpp = NULL;
1846 			if (error != EOPNOTSUPP)
1847 				return (error);
1848 			return (0);
1849 		}
1850 	}
1851 	if ((mp = *mpp) == NULL)
1852 		return (0);
1853 
1854 	/*
1855 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1856 	 * a vfs_ref().
1857 	 * As long as a vnode is not provided we need to acquire a
1858 	 * refcount for the provided mountpoint too, in order to
1859 	 * emulate a vfs_ref().
1860 	 */
1861 	if (vp == NULL && (flags & V_MNTREF) == 0)
1862 		vfs_ref(mp);
1863 
1864 	return (vn_start_write_refed(mp, flags, false));
1865 }
1866 
1867 /*
1868  * Secondary suspension. Used by operations such as vop_inactive
1869  * routines that are needed by the higher level functions. These
1870  * are allowed to proceed until all the higher level functions have
1871  * completed (indicated by mnt_writeopcount dropping to zero). At that
1872  * time, these operations are halted until the suspension is over.
1873  */
1874 int
1875 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1876 {
1877 	struct mount *mp;
1878 	int error;
1879 
1880 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1881 	    ("V_MNTREF requires mp"));
1882 
1883  retry:
1884 	if (vp != NULL) {
1885 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1886 			*mpp = NULL;
1887 			if (error != EOPNOTSUPP)
1888 				return (error);
1889 			return (0);
1890 		}
1891 	}
1892 	/*
1893 	 * If we are not suspended or have not yet reached suspended
1894 	 * mode, then let the operation proceed.
1895 	 */
1896 	if ((mp = *mpp) == NULL)
1897 		return (0);
1898 
1899 	/*
1900 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1901 	 * a vfs_ref().
1902 	 * As long as a vnode is not provided we need to acquire a
1903 	 * refcount for the provided mountpoint too, in order to
1904 	 * emulate a vfs_ref().
1905 	 */
1906 	MNT_ILOCK(mp);
1907 	if (vp == NULL && (flags & V_MNTREF) == 0)
1908 		MNT_REF(mp);
1909 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1910 		mp->mnt_secondary_writes++;
1911 		mp->mnt_secondary_accwrites++;
1912 		MNT_IUNLOCK(mp);
1913 		return (0);
1914 	}
1915 	if (flags & V_NOWAIT) {
1916 		MNT_REL(mp);
1917 		MNT_IUNLOCK(mp);
1918 		return (EWOULDBLOCK);
1919 	}
1920 	/*
1921 	 * Wait for the suspension to finish.
1922 	 */
1923 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1924 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1925 	    "suspfs", 0);
1926 	vfs_rel(mp);
1927 	if (error == 0)
1928 		goto retry;
1929 	return (error);
1930 }
1931 
1932 /*
1933  * Filesystem write operation has completed. If we are suspending and this
1934  * operation is the last one, notify the suspender that the suspension is
1935  * now in effect.
1936  */
1937 void
1938 vn_finished_write(struct mount *mp)
1939 {
1940 	struct mount_pcpu *mpcpu;
1941 	int c;
1942 
1943 	if (mp == NULL)
1944 		return;
1945 
1946 	if (vfs_op_thread_enter(mp, mpcpu)) {
1947 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
1948 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
1949 		vfs_op_thread_exit(mp, mpcpu);
1950 		return;
1951 	}
1952 
1953 	MNT_ILOCK(mp);
1954 	vfs_assert_mount_counters(mp);
1955 	MNT_REL(mp);
1956 	c = --mp->mnt_writeopcount;
1957 	if (mp->mnt_vfs_ops == 0) {
1958 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1959 		MNT_IUNLOCK(mp);
1960 		return;
1961 	}
1962 	if (c < 0)
1963 		vfs_dump_mount_counters(mp);
1964 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1965 		wakeup(&mp->mnt_writeopcount);
1966 	MNT_IUNLOCK(mp);
1967 }
1968 
1969 /*
1970  * Filesystem secondary write operation has completed. If we are
1971  * suspending and this operation is the last one, notify the suspender
1972  * that the suspension is now in effect.
1973  */
1974 void
1975 vn_finished_secondary_write(struct mount *mp)
1976 {
1977 	if (mp == NULL)
1978 		return;
1979 	MNT_ILOCK(mp);
1980 	MNT_REL(mp);
1981 	mp->mnt_secondary_writes--;
1982 	if (mp->mnt_secondary_writes < 0)
1983 		panic("vn_finished_secondary_write: neg cnt");
1984 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1985 	    mp->mnt_secondary_writes <= 0)
1986 		wakeup(&mp->mnt_secondary_writes);
1987 	MNT_IUNLOCK(mp);
1988 }
1989 
1990 /*
1991  * Request a filesystem to suspend write operations.
1992  */
1993 int
1994 vfs_write_suspend(struct mount *mp, int flags)
1995 {
1996 	int error;
1997 
1998 	vfs_op_enter(mp);
1999 
2000 	MNT_ILOCK(mp);
2001 	vfs_assert_mount_counters(mp);
2002 	if (mp->mnt_susp_owner == curthread) {
2003 		vfs_op_exit_locked(mp);
2004 		MNT_IUNLOCK(mp);
2005 		return (EALREADY);
2006 	}
2007 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2008 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2009 
2010 	/*
2011 	 * Unmount holds a write reference on the mount point.  If we
2012 	 * own busy reference and drain for writers, we deadlock with
2013 	 * the reference draining in the unmount path.  Callers of
2014 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2015 	 * vfs_busy() reference is owned and caller is not in the
2016 	 * unmount context.
2017 	 */
2018 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2019 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2020 		vfs_op_exit_locked(mp);
2021 		MNT_IUNLOCK(mp);
2022 		return (EBUSY);
2023 	}
2024 
2025 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2026 	mp->mnt_susp_owner = curthread;
2027 	if (mp->mnt_writeopcount > 0)
2028 		(void) msleep(&mp->mnt_writeopcount,
2029 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2030 	else
2031 		MNT_IUNLOCK(mp);
2032 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2033 		vfs_write_resume(mp, 0);
2034 		/* vfs_write_resume does vfs_op_exit() for us */
2035 	}
2036 	return (error);
2037 }
2038 
2039 /*
2040  * Request a filesystem to resume write operations.
2041  */
2042 void
2043 vfs_write_resume(struct mount *mp, int flags)
2044 {
2045 
2046 	MNT_ILOCK(mp);
2047 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2048 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2049 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2050 				       MNTK_SUSPENDED);
2051 		mp->mnt_susp_owner = NULL;
2052 		wakeup(&mp->mnt_writeopcount);
2053 		wakeup(&mp->mnt_flag);
2054 		curthread->td_pflags &= ~TDP_IGNSUSP;
2055 		if ((flags & VR_START_WRITE) != 0) {
2056 			MNT_REF(mp);
2057 			mp->mnt_writeopcount++;
2058 		}
2059 		MNT_IUNLOCK(mp);
2060 		if ((flags & VR_NO_SUSPCLR) == 0)
2061 			VFS_SUSP_CLEAN(mp);
2062 		vfs_op_exit(mp);
2063 	} else if ((flags & VR_START_WRITE) != 0) {
2064 		MNT_REF(mp);
2065 		vn_start_write_refed(mp, 0, true);
2066 	} else {
2067 		MNT_IUNLOCK(mp);
2068 	}
2069 }
2070 
2071 /*
2072  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2073  * methods.
2074  */
2075 int
2076 vfs_write_suspend_umnt(struct mount *mp)
2077 {
2078 	int error;
2079 
2080 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2081 	    ("vfs_write_suspend_umnt: recursed"));
2082 
2083 	/* dounmount() already called vn_start_write(). */
2084 	for (;;) {
2085 		vn_finished_write(mp);
2086 		error = vfs_write_suspend(mp, 0);
2087 		if (error != 0) {
2088 			vn_start_write(NULL, &mp, V_WAIT);
2089 			return (error);
2090 		}
2091 		MNT_ILOCK(mp);
2092 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2093 			break;
2094 		MNT_IUNLOCK(mp);
2095 		vn_start_write(NULL, &mp, V_WAIT);
2096 	}
2097 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2098 	wakeup(&mp->mnt_flag);
2099 	MNT_IUNLOCK(mp);
2100 	curthread->td_pflags |= TDP_IGNSUSP;
2101 	return (0);
2102 }
2103 
2104 /*
2105  * Implement kqueues for files by translating it to vnode operation.
2106  */
2107 static int
2108 vn_kqfilter(struct file *fp, struct knote *kn)
2109 {
2110 
2111 	return (VOP_KQFILTER(fp->f_vnode, kn));
2112 }
2113 
2114 /*
2115  * Simplified in-kernel wrapper calls for extended attribute access.
2116  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2117  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2118  */
2119 int
2120 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2121     const char *attrname, int *buflen, char *buf, struct thread *td)
2122 {
2123 	struct uio	auio;
2124 	struct iovec	iov;
2125 	int	error;
2126 
2127 	iov.iov_len = *buflen;
2128 	iov.iov_base = buf;
2129 
2130 	auio.uio_iov = &iov;
2131 	auio.uio_iovcnt = 1;
2132 	auio.uio_rw = UIO_READ;
2133 	auio.uio_segflg = UIO_SYSSPACE;
2134 	auio.uio_td = td;
2135 	auio.uio_offset = 0;
2136 	auio.uio_resid = *buflen;
2137 
2138 	if ((ioflg & IO_NODELOCKED) == 0)
2139 		vn_lock(vp, LK_SHARED | LK_RETRY);
2140 
2141 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2142 
2143 	/* authorize attribute retrieval as kernel */
2144 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2145 	    td);
2146 
2147 	if ((ioflg & IO_NODELOCKED) == 0)
2148 		VOP_UNLOCK(vp);
2149 
2150 	if (error == 0) {
2151 		*buflen = *buflen - auio.uio_resid;
2152 	}
2153 
2154 	return (error);
2155 }
2156 
2157 /*
2158  * XXX failure mode if partially written?
2159  */
2160 int
2161 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2162     const char *attrname, int buflen, char *buf, struct thread *td)
2163 {
2164 	struct uio	auio;
2165 	struct iovec	iov;
2166 	struct mount	*mp;
2167 	int	error;
2168 
2169 	iov.iov_len = buflen;
2170 	iov.iov_base = buf;
2171 
2172 	auio.uio_iov = &iov;
2173 	auio.uio_iovcnt = 1;
2174 	auio.uio_rw = UIO_WRITE;
2175 	auio.uio_segflg = UIO_SYSSPACE;
2176 	auio.uio_td = td;
2177 	auio.uio_offset = 0;
2178 	auio.uio_resid = buflen;
2179 
2180 	if ((ioflg & IO_NODELOCKED) == 0) {
2181 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2182 			return (error);
2183 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2184 	}
2185 
2186 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2187 
2188 	/* authorize attribute setting as kernel */
2189 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2190 
2191 	if ((ioflg & IO_NODELOCKED) == 0) {
2192 		vn_finished_write(mp);
2193 		VOP_UNLOCK(vp);
2194 	}
2195 
2196 	return (error);
2197 }
2198 
2199 int
2200 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2201     const char *attrname, struct thread *td)
2202 {
2203 	struct mount	*mp;
2204 	int	error;
2205 
2206 	if ((ioflg & IO_NODELOCKED) == 0) {
2207 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2208 			return (error);
2209 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2210 	}
2211 
2212 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2213 
2214 	/* authorize attribute removal as kernel */
2215 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2216 	if (error == EOPNOTSUPP)
2217 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2218 		    NULL, td);
2219 
2220 	if ((ioflg & IO_NODELOCKED) == 0) {
2221 		vn_finished_write(mp);
2222 		VOP_UNLOCK(vp);
2223 	}
2224 
2225 	return (error);
2226 }
2227 
2228 static int
2229 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2230     struct vnode **rvp)
2231 {
2232 
2233 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2234 }
2235 
2236 int
2237 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2238 {
2239 
2240 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2241 	    lkflags, rvp));
2242 }
2243 
2244 int
2245 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2246     int lkflags, struct vnode **rvp)
2247 {
2248 	struct mount *mp;
2249 	int ltype, error;
2250 
2251 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2252 	mp = vp->v_mount;
2253 	ltype = VOP_ISLOCKED(vp);
2254 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2255 	    ("vn_vget_ino: vp not locked"));
2256 	error = vfs_busy(mp, MBF_NOWAIT);
2257 	if (error != 0) {
2258 		vfs_ref(mp);
2259 		VOP_UNLOCK(vp);
2260 		error = vfs_busy(mp, 0);
2261 		vn_lock(vp, ltype | LK_RETRY);
2262 		vfs_rel(mp);
2263 		if (error != 0)
2264 			return (ENOENT);
2265 		if (VN_IS_DOOMED(vp)) {
2266 			vfs_unbusy(mp);
2267 			return (ENOENT);
2268 		}
2269 	}
2270 	VOP_UNLOCK(vp);
2271 	error = alloc(mp, alloc_arg, lkflags, rvp);
2272 	vfs_unbusy(mp);
2273 	if (error != 0 || *rvp != vp)
2274 		vn_lock(vp, ltype | LK_RETRY);
2275 	if (VN_IS_DOOMED(vp)) {
2276 		if (error == 0) {
2277 			if (*rvp == vp)
2278 				vunref(vp);
2279 			else
2280 				vput(*rvp);
2281 		}
2282 		error = ENOENT;
2283 	}
2284 	return (error);
2285 }
2286 
2287 int
2288 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2289     struct thread *td)
2290 {
2291 
2292 	if (vp->v_type != VREG || td == NULL)
2293 		return (0);
2294 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2295 	    lim_cur(td, RLIMIT_FSIZE)) {
2296 		PROC_LOCK(td->td_proc);
2297 		kern_psignal(td->td_proc, SIGXFSZ);
2298 		PROC_UNLOCK(td->td_proc);
2299 		return (EFBIG);
2300 	}
2301 	return (0);
2302 }
2303 
2304 int
2305 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2306     struct thread *td)
2307 {
2308 	struct vnode *vp;
2309 
2310 	vp = fp->f_vnode;
2311 #ifdef AUDIT
2312 	vn_lock(vp, LK_SHARED | LK_RETRY);
2313 	AUDIT_ARG_VNODE1(vp);
2314 	VOP_UNLOCK(vp);
2315 #endif
2316 	return (setfmode(td, active_cred, vp, mode));
2317 }
2318 
2319 int
2320 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2321     struct thread *td)
2322 {
2323 	struct vnode *vp;
2324 
2325 	vp = fp->f_vnode;
2326 #ifdef AUDIT
2327 	vn_lock(vp, LK_SHARED | LK_RETRY);
2328 	AUDIT_ARG_VNODE1(vp);
2329 	VOP_UNLOCK(vp);
2330 #endif
2331 	return (setfown(td, active_cred, vp, uid, gid));
2332 }
2333 
2334 void
2335 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2336 {
2337 	vm_object_t object;
2338 
2339 	if ((object = vp->v_object) == NULL)
2340 		return;
2341 	VM_OBJECT_WLOCK(object);
2342 	vm_object_page_remove(object, start, end, 0);
2343 	VM_OBJECT_WUNLOCK(object);
2344 }
2345 
2346 int
2347 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2348 {
2349 	struct vattr va;
2350 	daddr_t bn, bnp;
2351 	uint64_t bsize;
2352 	off_t noff;
2353 	int error;
2354 
2355 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2356 	    ("Wrong command %lu", cmd));
2357 
2358 	if (vn_lock(vp, LK_SHARED) != 0)
2359 		return (EBADF);
2360 	if (vp->v_type != VREG) {
2361 		error = ENOTTY;
2362 		goto unlock;
2363 	}
2364 	error = VOP_GETATTR(vp, &va, cred);
2365 	if (error != 0)
2366 		goto unlock;
2367 	noff = *off;
2368 	if (noff >= va.va_size) {
2369 		error = ENXIO;
2370 		goto unlock;
2371 	}
2372 	bsize = vp->v_mount->mnt_stat.f_iosize;
2373 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2374 	    noff % bsize) {
2375 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2376 		if (error == EOPNOTSUPP) {
2377 			error = ENOTTY;
2378 			goto unlock;
2379 		}
2380 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2381 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2382 			noff = bn * bsize;
2383 			if (noff < *off)
2384 				noff = *off;
2385 			goto unlock;
2386 		}
2387 	}
2388 	if (noff > va.va_size)
2389 		noff = va.va_size;
2390 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2391 	if (cmd == FIOSEEKDATA)
2392 		error = ENXIO;
2393 unlock:
2394 	VOP_UNLOCK(vp);
2395 	if (error == 0)
2396 		*off = noff;
2397 	return (error);
2398 }
2399 
2400 int
2401 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2402 {
2403 	struct ucred *cred;
2404 	struct vnode *vp;
2405 	struct vattr vattr;
2406 	off_t foffset, size;
2407 	int error, noneg;
2408 
2409 	cred = td->td_ucred;
2410 	vp = fp->f_vnode;
2411 	foffset = foffset_lock(fp, 0);
2412 	noneg = (vp->v_type != VCHR);
2413 	error = 0;
2414 	switch (whence) {
2415 	case L_INCR:
2416 		if (noneg &&
2417 		    (foffset < 0 ||
2418 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2419 			error = EOVERFLOW;
2420 			break;
2421 		}
2422 		offset += foffset;
2423 		break;
2424 	case L_XTND:
2425 		vn_lock(vp, LK_SHARED | LK_RETRY);
2426 		error = VOP_GETATTR(vp, &vattr, cred);
2427 		VOP_UNLOCK(vp);
2428 		if (error)
2429 			break;
2430 
2431 		/*
2432 		 * If the file references a disk device, then fetch
2433 		 * the media size and use that to determine the ending
2434 		 * offset.
2435 		 */
2436 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2437 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2438 			vattr.va_size = size;
2439 		if (noneg &&
2440 		    (vattr.va_size > OFF_MAX ||
2441 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2442 			error = EOVERFLOW;
2443 			break;
2444 		}
2445 		offset += vattr.va_size;
2446 		break;
2447 	case L_SET:
2448 		break;
2449 	case SEEK_DATA:
2450 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2451 		if (error == ENOTTY)
2452 			error = EINVAL;
2453 		break;
2454 	case SEEK_HOLE:
2455 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2456 		if (error == ENOTTY)
2457 			error = EINVAL;
2458 		break;
2459 	default:
2460 		error = EINVAL;
2461 	}
2462 	if (error == 0 && noneg && offset < 0)
2463 		error = EINVAL;
2464 	if (error != 0)
2465 		goto drop;
2466 	VFS_KNOTE_UNLOCKED(vp, 0);
2467 	td->td_uretoff.tdu_off = offset;
2468 drop:
2469 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2470 	return (error);
2471 }
2472 
2473 int
2474 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2475     struct thread *td)
2476 {
2477 	int error;
2478 
2479 	/*
2480 	 * Grant permission if the caller is the owner of the file, or
2481 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2482 	 * on the file.  If the time pointer is null, then write
2483 	 * permission on the file is also sufficient.
2484 	 *
2485 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2486 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2487 	 * will be allowed to set the times [..] to the current
2488 	 * server time.
2489 	 */
2490 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2491 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2492 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2493 	return (error);
2494 }
2495 
2496 int
2497 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2498 {
2499 	struct vnode *vp;
2500 	int error;
2501 
2502 	if (fp->f_type == DTYPE_FIFO)
2503 		kif->kf_type = KF_TYPE_FIFO;
2504 	else
2505 		kif->kf_type = KF_TYPE_VNODE;
2506 	vp = fp->f_vnode;
2507 	vref(vp);
2508 	FILEDESC_SUNLOCK(fdp);
2509 	error = vn_fill_kinfo_vnode(vp, kif);
2510 	vrele(vp);
2511 	FILEDESC_SLOCK(fdp);
2512 	return (error);
2513 }
2514 
2515 static inline void
2516 vn_fill_junk(struct kinfo_file *kif)
2517 {
2518 	size_t len, olen;
2519 
2520 	/*
2521 	 * Simulate vn_fullpath returning changing values for a given
2522 	 * vp during e.g. coredump.
2523 	 */
2524 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2525 	olen = strlen(kif->kf_path);
2526 	if (len < olen)
2527 		strcpy(&kif->kf_path[len - 1], "$");
2528 	else
2529 		for (; olen < len; olen++)
2530 			strcpy(&kif->kf_path[olen], "A");
2531 }
2532 
2533 int
2534 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2535 {
2536 	struct vattr va;
2537 	char *fullpath, *freepath;
2538 	int error;
2539 
2540 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2541 	freepath = NULL;
2542 	fullpath = "-";
2543 	error = vn_fullpath(vp, &fullpath, &freepath);
2544 	if (error == 0) {
2545 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2546 	}
2547 	if (freepath != NULL)
2548 		free(freepath, M_TEMP);
2549 
2550 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2551 		vn_fill_junk(kif);
2552 	);
2553 
2554 	/*
2555 	 * Retrieve vnode attributes.
2556 	 */
2557 	va.va_fsid = VNOVAL;
2558 	va.va_rdev = NODEV;
2559 	vn_lock(vp, LK_SHARED | LK_RETRY);
2560 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2561 	VOP_UNLOCK(vp);
2562 	if (error != 0)
2563 		return (error);
2564 	if (va.va_fsid != VNOVAL)
2565 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2566 	else
2567 		kif->kf_un.kf_file.kf_file_fsid =
2568 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2569 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2570 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2571 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2572 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2573 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2574 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2575 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2576 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2577 	return (0);
2578 }
2579 
2580 int
2581 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2582     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2583     struct thread *td)
2584 {
2585 #ifdef HWPMC_HOOKS
2586 	struct pmckern_map_in pkm;
2587 #endif
2588 	struct mount *mp;
2589 	struct vnode *vp;
2590 	vm_object_t object;
2591 	vm_prot_t maxprot;
2592 	boolean_t writecounted;
2593 	int error;
2594 
2595 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2596     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2597 	/*
2598 	 * POSIX shared-memory objects are defined to have
2599 	 * kernel persistence, and are not defined to support
2600 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2601 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2602 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2603 	 * flag to request this behavior.
2604 	 */
2605 	if ((fp->f_flag & FPOSIXSHM) != 0)
2606 		flags |= MAP_NOSYNC;
2607 #endif
2608 	vp = fp->f_vnode;
2609 
2610 	/*
2611 	 * Ensure that file and memory protections are
2612 	 * compatible.  Note that we only worry about
2613 	 * writability if mapping is shared; in this case,
2614 	 * current and max prot are dictated by the open file.
2615 	 * XXX use the vnode instead?  Problem is: what
2616 	 * credentials do we use for determination? What if
2617 	 * proc does a setuid?
2618 	 */
2619 	mp = vp->v_mount;
2620 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2621 		maxprot = VM_PROT_NONE;
2622 		if ((prot & VM_PROT_EXECUTE) != 0)
2623 			return (EACCES);
2624 	} else
2625 		maxprot = VM_PROT_EXECUTE;
2626 	if ((fp->f_flag & FREAD) != 0)
2627 		maxprot |= VM_PROT_READ;
2628 	else if ((prot & VM_PROT_READ) != 0)
2629 		return (EACCES);
2630 
2631 	/*
2632 	 * If we are sharing potential changes via MAP_SHARED and we
2633 	 * are trying to get write permission although we opened it
2634 	 * without asking for it, bail out.
2635 	 */
2636 	if ((flags & MAP_SHARED) != 0) {
2637 		if ((fp->f_flag & FWRITE) != 0)
2638 			maxprot |= VM_PROT_WRITE;
2639 		else if ((prot & VM_PROT_WRITE) != 0)
2640 			return (EACCES);
2641 	} else {
2642 		maxprot |= VM_PROT_WRITE;
2643 		cap_maxprot |= VM_PROT_WRITE;
2644 	}
2645 	maxprot &= cap_maxprot;
2646 
2647 	/*
2648 	 * For regular files and shared memory, POSIX requires that
2649 	 * the value of foff be a legitimate offset within the data
2650 	 * object.  In particular, negative offsets are invalid.
2651 	 * Blocking negative offsets and overflows here avoids
2652 	 * possible wraparound or user-level access into reserved
2653 	 * ranges of the data object later.  In contrast, POSIX does
2654 	 * not dictate how offsets are used by device drivers, so in
2655 	 * the case of a device mapping a negative offset is passed
2656 	 * on.
2657 	 */
2658 	if (
2659 #ifdef _LP64
2660 	    size > OFF_MAX ||
2661 #endif
2662 	    foff > OFF_MAX - size)
2663 		return (EINVAL);
2664 
2665 	writecounted = FALSE;
2666 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2667 	    &foff, &object, &writecounted);
2668 	if (error != 0)
2669 		return (error);
2670 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2671 	    foff, writecounted, td);
2672 	if (error != 0) {
2673 		/*
2674 		 * If this mapping was accounted for in the vnode's
2675 		 * writecount, then undo that now.
2676 		 */
2677 		if (writecounted)
2678 			vm_pager_release_writecount(object, 0, size);
2679 		vm_object_deallocate(object);
2680 	}
2681 #ifdef HWPMC_HOOKS
2682 	/* Inform hwpmc(4) if an executable is being mapped. */
2683 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2684 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2685 			pkm.pm_file = vp;
2686 			pkm.pm_address = (uintptr_t) *addr;
2687 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2688 		}
2689 	}
2690 #endif
2691 	return (error);
2692 }
2693 
2694 void
2695 vn_fsid(struct vnode *vp, struct vattr *va)
2696 {
2697 	fsid_t *f;
2698 
2699 	f = &vp->v_mount->mnt_stat.f_fsid;
2700 	va->va_fsid = (uint32_t)f->val[1];
2701 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2702 	va->va_fsid += (uint32_t)f->val[0];
2703 }
2704 
2705 int
2706 vn_fsync_buf(struct vnode *vp, int waitfor)
2707 {
2708 	struct buf *bp, *nbp;
2709 	struct bufobj *bo;
2710 	struct mount *mp;
2711 	int error, maxretry;
2712 
2713 	error = 0;
2714 	maxretry = 10000;     /* large, arbitrarily chosen */
2715 	mp = NULL;
2716 	if (vp->v_type == VCHR) {
2717 		VI_LOCK(vp);
2718 		mp = vp->v_rdev->si_mountpt;
2719 		VI_UNLOCK(vp);
2720 	}
2721 	bo = &vp->v_bufobj;
2722 	BO_LOCK(bo);
2723 loop1:
2724 	/*
2725 	 * MARK/SCAN initialization to avoid infinite loops.
2726 	 */
2727         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2728 		bp->b_vflags &= ~BV_SCANNED;
2729 		bp->b_error = 0;
2730 	}
2731 
2732 	/*
2733 	 * Flush all dirty buffers associated with a vnode.
2734 	 */
2735 loop2:
2736 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2737 		if ((bp->b_vflags & BV_SCANNED) != 0)
2738 			continue;
2739 		bp->b_vflags |= BV_SCANNED;
2740 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2741 			if (waitfor != MNT_WAIT)
2742 				continue;
2743 			if (BUF_LOCK(bp,
2744 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2745 			    BO_LOCKPTR(bo)) != 0) {
2746 				BO_LOCK(bo);
2747 				goto loop1;
2748 			}
2749 			BO_LOCK(bo);
2750 		}
2751 		BO_UNLOCK(bo);
2752 		KASSERT(bp->b_bufobj == bo,
2753 		    ("bp %p wrong b_bufobj %p should be %p",
2754 		    bp, bp->b_bufobj, bo));
2755 		if ((bp->b_flags & B_DELWRI) == 0)
2756 			panic("fsync: not dirty");
2757 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2758 			vfs_bio_awrite(bp);
2759 		} else {
2760 			bremfree(bp);
2761 			bawrite(bp);
2762 		}
2763 		if (maxretry < 1000)
2764 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2765 		BO_LOCK(bo);
2766 		goto loop2;
2767 	}
2768 
2769 	/*
2770 	 * If synchronous the caller expects us to completely resolve all
2771 	 * dirty buffers in the system.  Wait for in-progress I/O to
2772 	 * complete (which could include background bitmap writes), then
2773 	 * retry if dirty blocks still exist.
2774 	 */
2775 	if (waitfor == MNT_WAIT) {
2776 		bufobj_wwait(bo, 0, 0);
2777 		if (bo->bo_dirty.bv_cnt > 0) {
2778 			/*
2779 			 * If we are unable to write any of these buffers
2780 			 * then we fail now rather than trying endlessly
2781 			 * to write them out.
2782 			 */
2783 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2784 				if ((error = bp->b_error) != 0)
2785 					break;
2786 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2787 			    (error == 0 && --maxretry >= 0))
2788 				goto loop1;
2789 			if (error == 0)
2790 				error = EAGAIN;
2791 		}
2792 	}
2793 	BO_UNLOCK(bo);
2794 	if (error != 0)
2795 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2796 
2797 	return (error);
2798 }
2799 
2800 /*
2801  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2802  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2803  * to do the actual copy.
2804  * vn_generic_copy_file_range() is factored out, so it can be called
2805  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2806  * different file systems.
2807  */
2808 int
2809 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2810     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2811     struct ucred *outcred, struct thread *fsize_td)
2812 {
2813 	int error;
2814 	size_t len;
2815 	uint64_t uval;
2816 
2817 	len = *lenp;
2818 	*lenp = 0;		/* For error returns. */
2819 	error = 0;
2820 
2821 	/* Do some sanity checks on the arguments. */
2822 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2823 		error = EISDIR;
2824 	else if (*inoffp < 0 || *outoffp < 0 ||
2825 	    invp->v_type != VREG || outvp->v_type != VREG)
2826 		error = EINVAL;
2827 	if (error != 0)
2828 		goto out;
2829 
2830 	/* Ensure offset + len does not wrap around. */
2831 	uval = *inoffp;
2832 	uval += len;
2833 	if (uval > INT64_MAX)
2834 		len = INT64_MAX - *inoffp;
2835 	uval = *outoffp;
2836 	uval += len;
2837 	if (uval > INT64_MAX)
2838 		len = INT64_MAX - *outoffp;
2839 	if (len == 0)
2840 		goto out;
2841 
2842 	/*
2843 	 * If the two vnode are for the same file system, call
2844 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2845 	 * which can handle copies across multiple file systems.
2846 	 */
2847 	*lenp = len;
2848 	if (invp->v_mount == outvp->v_mount)
2849 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2850 		    lenp, flags, incred, outcred, fsize_td);
2851 	else
2852 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2853 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2854 out:
2855 	return (error);
2856 }
2857 
2858 /*
2859  * Test len bytes of data starting at dat for all bytes == 0.
2860  * Return true if all bytes are zero, false otherwise.
2861  * Expects dat to be well aligned.
2862  */
2863 static bool
2864 mem_iszero(void *dat, int len)
2865 {
2866 	int i;
2867 	const u_int *p;
2868 	const char *cp;
2869 
2870 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2871 		if (len >= sizeof(*p)) {
2872 			if (*p != 0)
2873 				return (false);
2874 		} else {
2875 			cp = (const char *)p;
2876 			for (i = 0; i < len; i++, cp++)
2877 				if (*cp != '\0')
2878 					return (false);
2879 		}
2880 	}
2881 	return (true);
2882 }
2883 
2884 /*
2885  * Look for a hole in the output file and, if found, adjust *outoffp
2886  * and *xferp to skip past the hole.
2887  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2888  * to be written as 0's upon return.
2889  */
2890 static off_t
2891 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2892     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2893 {
2894 	int error;
2895 	off_t delta;
2896 
2897 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2898 		*dataoffp = *outoffp;
2899 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2900 		    curthread);
2901 		if (error == 0) {
2902 			*holeoffp = *dataoffp;
2903 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2904 			    curthread);
2905 		}
2906 		if (error != 0 || *holeoffp == *dataoffp) {
2907 			/*
2908 			 * Since outvp is unlocked, it may be possible for
2909 			 * another thread to do a truncate(), lseek(), write()
2910 			 * creating a hole at startoff between the above
2911 			 * VOP_IOCTL() calls, if the other thread does not do
2912 			 * rangelocking.
2913 			 * If that happens, *holeoffp == *dataoffp and finding
2914 			 * the hole has failed, so disable vn_skip_hole().
2915 			 */
2916 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2917 			return (xfer2);
2918 		}
2919 		KASSERT(*dataoffp >= *outoffp,
2920 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2921 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2922 		KASSERT(*holeoffp > *dataoffp,
2923 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2924 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2925 	}
2926 
2927 	/*
2928 	 * If there is a hole before the data starts, advance *outoffp and
2929 	 * *xferp past the hole.
2930 	 */
2931 	if (*dataoffp > *outoffp) {
2932 		delta = *dataoffp - *outoffp;
2933 		if (delta >= *xferp) {
2934 			/* Entire *xferp is a hole. */
2935 			*outoffp += *xferp;
2936 			*xferp = 0;
2937 			return (0);
2938 		}
2939 		*xferp -= delta;
2940 		*outoffp += delta;
2941 		xfer2 = MIN(xfer2, *xferp);
2942 	}
2943 
2944 	/*
2945 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2946 	 * that the write ends at the start of the hole.
2947 	 * *holeoffp should always be greater than *outoffp, but for the
2948 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2949 	 * value.
2950 	 */
2951 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2952 		xfer2 = *holeoffp - *outoffp;
2953 	return (xfer2);
2954 }
2955 
2956 /*
2957  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2958  * dat is a maximum of blksize in length and can be written repeatedly in
2959  * the chunk.
2960  * If growfile == true, just grow the file via vn_truncate_locked() instead
2961  * of doing actual writes.
2962  * If checkhole == true, a hole is being punched, so skip over any hole
2963  * already in the output file.
2964  */
2965 static int
2966 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2967     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2968 {
2969 	struct mount *mp;
2970 	off_t dataoff, holeoff, xfer2;
2971 	int error, lckf;
2972 
2973 	/*
2974 	 * Loop around doing writes of blksize until write has been completed.
2975 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2976 	 * done for each iteration, since the xfer argument can be very
2977 	 * large if there is a large hole to punch in the output file.
2978 	 */
2979 	error = 0;
2980 	holeoff = 0;
2981 	do {
2982 		xfer2 = MIN(xfer, blksize);
2983 		if (checkhole) {
2984 			/*
2985 			 * Punching a hole.  Skip writing if there is
2986 			 * already a hole in the output file.
2987 			 */
2988 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2989 			    &dataoff, &holeoff, cred);
2990 			if (xfer == 0)
2991 				break;
2992 			if (holeoff < 0)
2993 				checkhole = false;
2994 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2995 			    (intmax_t)xfer2));
2996 		}
2997 		bwillwrite();
2998 		mp = NULL;
2999 		error = vn_start_write(outvp, &mp, V_WAIT);
3000 		if (error != 0)
3001 			break;
3002 		if (growfile) {
3003 			error = vn_lock(outvp, LK_EXCLUSIVE);
3004 			if (error == 0) {
3005 				error = vn_truncate_locked(outvp, outoff + xfer,
3006 				    false, cred);
3007 				VOP_UNLOCK(outvp);
3008 			}
3009 		} else {
3010 			if (MNT_SHARED_WRITES(mp))
3011 				lckf = LK_SHARED;
3012 			else
3013 				lckf = LK_EXCLUSIVE;
3014 			error = vn_lock(outvp, lckf);
3015 			if (error == 0) {
3016 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3017 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3018 				    curthread->td_ucred, cred, NULL, curthread);
3019 				outoff += xfer2;
3020 				xfer -= xfer2;
3021 				VOP_UNLOCK(outvp);
3022 			}
3023 		}
3024 		if (mp != NULL)
3025 			vn_finished_write(mp);
3026 	} while (!growfile && xfer > 0 && error == 0);
3027 	return (error);
3028 }
3029 
3030 /*
3031  * Copy a byte range of one file to another.  This function can handle the
3032  * case where invp and outvp are on different file systems.
3033  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3034  * is no better file system specific way to do it.
3035  */
3036 int
3037 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3038     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3039     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3040 {
3041 	struct vattr va;
3042 	struct mount *mp;
3043 	struct uio io;
3044 	off_t startoff, endoff, xfer, xfer2;
3045 	u_long blksize;
3046 	int error, interrupted;
3047 	bool cantseek, readzeros, eof, lastblock;
3048 	ssize_t aresid;
3049 	size_t copylen, len, rem, savlen;
3050 	char *dat;
3051 	long holein, holeout;
3052 
3053 	holein = holeout = 0;
3054 	savlen = len = *lenp;
3055 	error = 0;
3056 	interrupted = 0;
3057 	dat = NULL;
3058 
3059 	error = vn_lock(invp, LK_SHARED);
3060 	if (error != 0)
3061 		goto out;
3062 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3063 		holein = 0;
3064 	VOP_UNLOCK(invp);
3065 
3066 	mp = NULL;
3067 	error = vn_start_write(outvp, &mp, V_WAIT);
3068 	if (error == 0)
3069 		error = vn_lock(outvp, LK_EXCLUSIVE);
3070 	if (error == 0) {
3071 		/*
3072 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3073 		 * now that outvp is locked.
3074 		 */
3075 		if (fsize_td != NULL) {
3076 			io.uio_offset = *outoffp;
3077 			io.uio_resid = len;
3078 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3079 			if (error != 0)
3080 				error = EFBIG;
3081 		}
3082 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3083 			holeout = 0;
3084 		/*
3085 		 * Holes that are past EOF do not need to be written as a block
3086 		 * of zero bytes.  So, truncate the output file as far as
3087 		 * possible and then use va.va_size to decide if writing 0
3088 		 * bytes is necessary in the loop below.
3089 		 */
3090 		if (error == 0)
3091 			error = VOP_GETATTR(outvp, &va, outcred);
3092 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3093 		    *outoffp + len) {
3094 #ifdef MAC
3095 			error = mac_vnode_check_write(curthread->td_ucred,
3096 			    outcred, outvp);
3097 			if (error == 0)
3098 #endif
3099 				error = vn_truncate_locked(outvp, *outoffp,
3100 				    false, outcred);
3101 			if (error == 0)
3102 				va.va_size = *outoffp;
3103 		}
3104 		VOP_UNLOCK(outvp);
3105 	}
3106 	if (mp != NULL)
3107 		vn_finished_write(mp);
3108 	if (error != 0)
3109 		goto out;
3110 
3111 	/*
3112 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3113 	 * If hole sizes aren't available, set the blksize to the larger
3114 	 * f_iosize of invp and outvp.
3115 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3116 	 * This value is clipped at 4Kbytes and 1Mbyte.
3117 	 */
3118 	blksize = MAX(holein, holeout);
3119 
3120 	/* Clip len to end at an exact multiple of hole size. */
3121 	if (blksize > 1) {
3122 		rem = *inoffp % blksize;
3123 		if (rem > 0)
3124 			rem = blksize - rem;
3125 		if (len - rem > blksize)
3126 			len = savlen = rounddown(len - rem, blksize) + rem;
3127 	}
3128 
3129 	if (blksize <= 1)
3130 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3131 		    outvp->v_mount->mnt_stat.f_iosize);
3132 	if (blksize < 4096)
3133 		blksize = 4096;
3134 	else if (blksize > 1024 * 1024)
3135 		blksize = 1024 * 1024;
3136 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3137 
3138 	/*
3139 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3140 	 * to find holes.  Otherwise, just scan the read block for all 0s
3141 	 * in the inner loop where the data copying is done.
3142 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3143 	 * support holes on the server, but do not support FIOSEEKHOLE.
3144 	 */
3145 	eof = false;
3146 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3147 		endoff = 0;			/* To shut up compilers. */
3148 		cantseek = true;
3149 		startoff = *inoffp;
3150 		copylen = len;
3151 
3152 		/*
3153 		 * Find the next data area.  If there is just a hole to EOF,
3154 		 * FIOSEEKDATA should fail and then we drop down into the
3155 		 * inner loop and create the hole on the outvp file.
3156 		 * (I do not know if any file system will report a hole to
3157 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3158 		 *  will fail for those file systems.)
3159 		 *
3160 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3161 		 * the code just falls through to the inner copy loop.
3162 		 */
3163 		error = EINVAL;
3164 		if (holein > 0)
3165 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3166 			    incred, curthread);
3167 		if (error == 0) {
3168 			endoff = startoff;
3169 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3170 			    incred, curthread);
3171 			/*
3172 			 * Since invp is unlocked, it may be possible for
3173 			 * another thread to do a truncate(), lseek(), write()
3174 			 * creating a hole at startoff between the above
3175 			 * VOP_IOCTL() calls, if the other thread does not do
3176 			 * rangelocking.
3177 			 * If that happens, startoff == endoff and finding
3178 			 * the hole has failed, so set an error.
3179 			 */
3180 			if (error == 0 && startoff == endoff)
3181 				error = EINVAL; /* Any error. Reset to 0. */
3182 		}
3183 		if (error == 0) {
3184 			if (startoff > *inoffp) {
3185 				/* Found hole before data block. */
3186 				xfer = MIN(startoff - *inoffp, len);
3187 				if (*outoffp < va.va_size) {
3188 					/* Must write 0s to punch hole. */
3189 					xfer2 = MIN(va.va_size - *outoffp,
3190 					    xfer);
3191 					memset(dat, 0, MIN(xfer2, blksize));
3192 					error = vn_write_outvp(outvp, dat,
3193 					    *outoffp, xfer2, blksize, false,
3194 					    holeout > 0, outcred);
3195 				}
3196 
3197 				if (error == 0 && *outoffp + xfer >
3198 				    va.va_size && xfer == len)
3199 					/* Grow last block. */
3200 					error = vn_write_outvp(outvp, dat,
3201 					    *outoffp, xfer, blksize, true,
3202 					    false, outcred);
3203 				if (error == 0) {
3204 					*inoffp += xfer;
3205 					*outoffp += xfer;
3206 					len -= xfer;
3207 					if (len < savlen)
3208 						interrupted = sig_intr();
3209 				}
3210 			}
3211 			copylen = MIN(len, endoff - startoff);
3212 			cantseek = false;
3213 		} else {
3214 			cantseek = true;
3215 			startoff = *inoffp;
3216 			copylen = len;
3217 			error = 0;
3218 		}
3219 
3220 		xfer = blksize;
3221 		if (cantseek) {
3222 			/*
3223 			 * Set first xfer to end at a block boundary, so that
3224 			 * holes are more likely detected in the loop below via
3225 			 * the for all bytes 0 method.
3226 			 */
3227 			xfer -= (*inoffp % blksize);
3228 		}
3229 		/* Loop copying the data block. */
3230 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3231 			if (copylen < xfer)
3232 				xfer = copylen;
3233 			error = vn_lock(invp, LK_SHARED);
3234 			if (error != 0)
3235 				goto out;
3236 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3237 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3238 			    curthread->td_ucred, incred, &aresid,
3239 			    curthread);
3240 			VOP_UNLOCK(invp);
3241 			lastblock = false;
3242 			if (error == 0 && aresid > 0) {
3243 				/* Stop the copy at EOF on the input file. */
3244 				xfer -= aresid;
3245 				eof = true;
3246 				lastblock = true;
3247 			}
3248 			if (error == 0) {
3249 				/*
3250 				 * Skip the write for holes past the initial EOF
3251 				 * of the output file, unless this is the last
3252 				 * write of the output file at EOF.
3253 				 */
3254 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3255 				    false;
3256 				if (xfer == len)
3257 					lastblock = true;
3258 				if (!cantseek || *outoffp < va.va_size ||
3259 				    lastblock || !readzeros)
3260 					error = vn_write_outvp(outvp, dat,
3261 					    *outoffp, xfer, blksize,
3262 					    readzeros && lastblock &&
3263 					    *outoffp >= va.va_size, false,
3264 					    outcred);
3265 				if (error == 0) {
3266 					*inoffp += xfer;
3267 					startoff += xfer;
3268 					*outoffp += xfer;
3269 					copylen -= xfer;
3270 					len -= xfer;
3271 					if (len < savlen)
3272 						interrupted = sig_intr();
3273 				}
3274 			}
3275 			xfer = blksize;
3276 		}
3277 	}
3278 out:
3279 	*lenp = savlen - len;
3280 	free(dat, M_TEMP);
3281 	return (error);
3282 }
3283 
3284 static int
3285 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3286 {
3287 	struct mount *mp;
3288 	struct vnode *vp;
3289 	off_t olen, ooffset;
3290 	int error;
3291 #ifdef AUDIT
3292 	int audited_vnode1 = 0;
3293 #endif
3294 
3295 	vp = fp->f_vnode;
3296 	if (vp->v_type != VREG)
3297 		return (ENODEV);
3298 
3299 	/* Allocating blocks may take a long time, so iterate. */
3300 	for (;;) {
3301 		olen = len;
3302 		ooffset = offset;
3303 
3304 		bwillwrite();
3305 		mp = NULL;
3306 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3307 		if (error != 0)
3308 			break;
3309 		error = vn_lock(vp, LK_EXCLUSIVE);
3310 		if (error != 0) {
3311 			vn_finished_write(mp);
3312 			break;
3313 		}
3314 #ifdef AUDIT
3315 		if (!audited_vnode1) {
3316 			AUDIT_ARG_VNODE1(vp);
3317 			audited_vnode1 = 1;
3318 		}
3319 #endif
3320 #ifdef MAC
3321 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3322 		if (error == 0)
3323 #endif
3324 			error = VOP_ALLOCATE(vp, &offset, &len);
3325 		VOP_UNLOCK(vp);
3326 		vn_finished_write(mp);
3327 
3328 		if (olen + ooffset != offset + len) {
3329 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3330 			    ooffset, olen, offset, len);
3331 		}
3332 		if (error != 0 || len == 0)
3333 			break;
3334 		KASSERT(olen > len, ("Iteration did not make progress?"));
3335 		maybe_yield();
3336 	}
3337 
3338 	return (error);
3339 }
3340 
3341 static u_long vn_lock_pair_pause_cnt;
3342 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3343     &vn_lock_pair_pause_cnt, 0,
3344     "Count of vn_lock_pair deadlocks");
3345 
3346 u_int vn_lock_pair_pause_max;
3347 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3348     &vn_lock_pair_pause_max, 0,
3349     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3350 
3351 static void
3352 vn_lock_pair_pause(const char *wmesg)
3353 {
3354 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3355 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3356 }
3357 
3358 /*
3359  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3360  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3361  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3362  * can be NULL.
3363  *
3364  * The function returns with both vnodes exclusively locked, and
3365  * guarantees that it does not create lock order reversal with other
3366  * threads during its execution.  Both vnodes could be unlocked
3367  * temporary (and reclaimed).
3368  */
3369 void
3370 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3371     bool vp2_locked)
3372 {
3373 	int error;
3374 
3375 	if (vp1 == NULL && vp2 == NULL)
3376 		return;
3377 	if (vp1 != NULL) {
3378 		if (vp1_locked)
3379 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3380 		else
3381 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3382 	} else {
3383 		vp1_locked = true;
3384 	}
3385 	if (vp2 != NULL) {
3386 		if (vp2_locked)
3387 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3388 		else
3389 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3390 	} else {
3391 		vp2_locked = true;
3392 	}
3393 	if (!vp1_locked && !vp2_locked) {
3394 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3395 		vp1_locked = true;
3396 	}
3397 
3398 	for (;;) {
3399 		if (vp1_locked && vp2_locked)
3400 			break;
3401 		if (vp1_locked && vp2 != NULL) {
3402 			if (vp1 != NULL) {
3403 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3404 				    __FILE__, __LINE__);
3405 				if (error == 0)
3406 					break;
3407 				VOP_UNLOCK(vp1);
3408 				vp1_locked = false;
3409 				vn_lock_pair_pause("vlp1");
3410 			}
3411 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3412 			vp2_locked = true;
3413 		}
3414 		if (vp2_locked && vp1 != NULL) {
3415 			if (vp2 != NULL) {
3416 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3417 				    __FILE__, __LINE__);
3418 				if (error == 0)
3419 					break;
3420 				VOP_UNLOCK(vp2);
3421 				vp2_locked = false;
3422 				vn_lock_pair_pause("vlp2");
3423 			}
3424 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3425 			vp1_locked = true;
3426 		}
3427 	}
3428 	if (vp1 != NULL)
3429 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3430 	if (vp2 != NULL)
3431 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3432 }
3433