xref: /freebsd/sys/kern/vfs_vnops.c (revision 3cbb4cc200f8a0ad7ed08233425ea54524a21f1c)
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 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE;
245 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
246 			ndp->ni_cnd.cn_flags |= FOLLOW;
247 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
248 			bwillwrite();
249 		if ((error = namei(ndp)) != 0)
250 			return (error);
251 		if (ndp->ni_vp == NULL) {
252 			VATTR_NULL(vap);
253 			vap->va_type = VREG;
254 			vap->va_mode = cmode;
255 			if (fmode & O_EXCL)
256 				vap->va_vaflags |= VA_EXCLUSIVE;
257 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
258 				NDFREE(ndp, NDF_ONLY_PNBUF);
259 				vput(ndp->ni_dvp);
260 				if ((error = vn_start_write(NULL, &mp,
261 				    V_XSLEEP | PCATCH)) != 0)
262 					return (error);
263 				NDREINIT(ndp);
264 				goto restart;
265 			}
266 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
267 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
268 #ifdef MAC
269 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
270 			    &ndp->ni_cnd, vap);
271 			if (error == 0)
272 #endif
273 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
274 						   &ndp->ni_cnd, vap);
275 			vput(ndp->ni_dvp);
276 			vn_finished_write(mp);
277 			if (error) {
278 				NDFREE(ndp, NDF_ONLY_PNBUF);
279 				if (error == ERELOOKUP) {
280 					NDREINIT(ndp);
281 					goto restart;
282 				}
283 				return (error);
284 			}
285 			fmode &= ~O_TRUNC;
286 			vp = ndp->ni_vp;
287 		} else {
288 			if (ndp->ni_dvp == ndp->ni_vp)
289 				vrele(ndp->ni_dvp);
290 			else
291 				vput(ndp->ni_dvp);
292 			ndp->ni_dvp = NULL;
293 			vp = ndp->ni_vp;
294 			if (fmode & O_EXCL) {
295 				error = EEXIST;
296 				goto bad;
297 			}
298 			if (vp->v_type == VDIR) {
299 				error = EISDIR;
300 				goto bad;
301 			}
302 			fmode &= ~O_CREAT;
303 		}
304 	} else {
305 		ndp->ni_cnd.cn_nameiop = LOOKUP;
306 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
307 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
308 		    FOLLOW;
309 		if ((fmode & FWRITE) == 0)
310 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
311 		if ((error = namei(ndp)) != 0)
312 			return (error);
313 		vp = ndp->ni_vp;
314 	}
315 	error = vn_open_vnode(vp, fmode, cred, td, fp);
316 	if (error)
317 		goto bad;
318 	*flagp = fmode;
319 	return (0);
320 bad:
321 	NDFREE(ndp, NDF_ONLY_PNBUF);
322 	vput(vp);
323 	*flagp = fmode;
324 	ndp->ni_vp = NULL;
325 	return (error);
326 }
327 
328 static int
329 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
330 {
331 	struct flock lf;
332 	int error, lock_flags, type;
333 
334 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
335 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
336 		return (0);
337 	KASSERT(fp != NULL, ("open with flock requires fp"));
338 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
339 		return (EOPNOTSUPP);
340 
341 	lock_flags = VOP_ISLOCKED(vp);
342 	VOP_UNLOCK(vp);
343 
344 	lf.l_whence = SEEK_SET;
345 	lf.l_start = 0;
346 	lf.l_len = 0;
347 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
348 	type = F_FLOCK;
349 	if ((fmode & FNONBLOCK) == 0)
350 		type |= F_WAIT;
351 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
352 	if (error == 0)
353 		fp->f_flag |= FHASLOCK;
354 
355 	vn_lock(vp, lock_flags | LK_RETRY);
356 	if (error == 0 && VN_IS_DOOMED(vp))
357 		error = ENOENT;
358 	return (error);
359 }
360 
361 /*
362  * Common code for vnode open operations once a vnode is located.
363  * Check permissions, and call the VOP_OPEN routine.
364  */
365 int
366 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
367     struct thread *td, struct file *fp)
368 {
369 	accmode_t accmode;
370 	int error;
371 
372 	if (vp->v_type == VLNK)
373 		return (EMLINK);
374 	if (vp->v_type == VSOCK)
375 		return (EOPNOTSUPP);
376 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
377 		return (ENOTDIR);
378 	accmode = 0;
379 	if (fmode & (FWRITE | O_TRUNC)) {
380 		if (vp->v_type == VDIR)
381 			return (EISDIR);
382 		accmode |= VWRITE;
383 	}
384 	if (fmode & FREAD)
385 		accmode |= VREAD;
386 	if (fmode & FEXEC)
387 		accmode |= VEXEC;
388 	if ((fmode & O_APPEND) && (fmode & FWRITE))
389 		accmode |= VAPPEND;
390 #ifdef MAC
391 	if (fmode & O_CREAT)
392 		accmode |= VCREAT;
393 	if (fmode & O_VERIFY)
394 		accmode |= VVERIFY;
395 	error = mac_vnode_check_open(cred, vp, accmode);
396 	if (error)
397 		return (error);
398 
399 	accmode &= ~(VCREAT | VVERIFY);
400 #endif
401 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
402 		error = VOP_ACCESS(vp, accmode, cred, td);
403 		if (error != 0)
404 			return (error);
405 	}
406 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
407 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
408 	error = VOP_OPEN(vp, fmode, cred, td, fp);
409 	if (error != 0)
410 		return (error);
411 
412 	error = vn_open_vnode_advlock(vp, fmode, fp);
413 	if (error == 0 && (fmode & FWRITE) != 0) {
414 		error = VOP_ADD_WRITECOUNT(vp, 1);
415 		if (error == 0) {
416 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
417 			     __func__, vp, vp->v_writecount);
418 		}
419 	}
420 
421 	/*
422 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
423 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
424 	 * Arrange for that by having fdrop() to use vn_closefile().
425 	 */
426 	if (error != 0) {
427 		fp->f_flag |= FOPENFAILED;
428 		fp->f_vnode = vp;
429 		if (fp->f_ops == &badfileops) {
430 			fp->f_type = DTYPE_VNODE;
431 			fp->f_ops = &vnops;
432 		}
433 		vref(vp);
434 	}
435 
436 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
437 	return (error);
438 
439 }
440 
441 /*
442  * Check for write permissions on the specified vnode.
443  * Prototype text segments cannot be written.
444  * It is racy.
445  */
446 int
447 vn_writechk(struct vnode *vp)
448 {
449 
450 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
451 	/*
452 	 * If there's shared text associated with
453 	 * the vnode, try to free it up once.  If
454 	 * we fail, we can't allow writing.
455 	 */
456 	if (VOP_IS_TEXT(vp))
457 		return (ETXTBSY);
458 
459 	return (0);
460 }
461 
462 /*
463  * Vnode close call
464  */
465 static int
466 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
467     struct thread *td, bool keep_ref)
468 {
469 	struct mount *mp;
470 	int error, lock_flags;
471 
472 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
473 	    MNT_EXTENDED_SHARED(vp->v_mount))
474 		lock_flags = LK_SHARED;
475 	else
476 		lock_flags = LK_EXCLUSIVE;
477 
478 	vn_start_write(vp, &mp, V_WAIT);
479 	vn_lock(vp, lock_flags | LK_RETRY);
480 	AUDIT_ARG_VNODE1(vp);
481 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
482 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
483 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
484 		    __func__, vp, vp->v_writecount);
485 	}
486 	error = VOP_CLOSE(vp, flags, file_cred, td);
487 	if (keep_ref)
488 		VOP_UNLOCK(vp);
489 	else
490 		vput(vp);
491 	vn_finished_write(mp);
492 	return (error);
493 }
494 
495 int
496 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
497     struct thread *td)
498 {
499 
500 	return (vn_close1(vp, flags, file_cred, td, false));
501 }
502 
503 /*
504  * Heuristic to detect sequential operation.
505  */
506 static int
507 sequential_heuristic(struct uio *uio, struct file *fp)
508 {
509 	enum uio_rw rw;
510 
511 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
512 
513 	rw = uio->uio_rw;
514 	if (fp->f_flag & FRDAHEAD)
515 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
516 
517 	/*
518 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
519 	 * that the first I/O is normally considered to be slightly
520 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
521 	 * unless previous seeks have reduced f_seqcount to 0, in which
522 	 * case offset 0 is not special.
523 	 */
524 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
525 	    uio->uio_offset == fp->f_nextoff[rw]) {
526 		/*
527 		 * f_seqcount is in units of fixed-size blocks so that it
528 		 * depends mainly on the amount of sequential I/O and not
529 		 * much on the number of sequential I/O's.  The fixed size
530 		 * of 16384 is hard-coded here since it is (not quite) just
531 		 * a magic size that works well here.  This size is more
532 		 * closely related to the best I/O size for real disks than
533 		 * to any block size used by software.
534 		 */
535 		if (uio->uio_resid >= IO_SEQMAX * 16384)
536 			fp->f_seqcount[rw] = IO_SEQMAX;
537 		else {
538 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
539 			if (fp->f_seqcount[rw] > IO_SEQMAX)
540 				fp->f_seqcount[rw] = IO_SEQMAX;
541 		}
542 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
543 	}
544 
545 	/* Not sequential.  Quickly draw-down sequentiality. */
546 	if (fp->f_seqcount[rw] > 1)
547 		fp->f_seqcount[rw] = 1;
548 	else
549 		fp->f_seqcount[rw] = 0;
550 	return (0);
551 }
552 
553 /*
554  * Package up an I/O request on a vnode into a uio and do it.
555  */
556 int
557 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
558     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
559     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
560 {
561 	struct uio auio;
562 	struct iovec aiov;
563 	struct mount *mp;
564 	struct ucred *cred;
565 	void *rl_cookie;
566 	struct vn_io_fault_args args;
567 	int error, lock_flags;
568 
569 	if (offset < 0 && vp->v_type != VCHR)
570 		return (EINVAL);
571 	auio.uio_iov = &aiov;
572 	auio.uio_iovcnt = 1;
573 	aiov.iov_base = base;
574 	aiov.iov_len = len;
575 	auio.uio_resid = len;
576 	auio.uio_offset = offset;
577 	auio.uio_segflg = segflg;
578 	auio.uio_rw = rw;
579 	auio.uio_td = td;
580 	error = 0;
581 
582 	if ((ioflg & IO_NODELOCKED) == 0) {
583 		if ((ioflg & IO_RANGELOCKED) == 0) {
584 			if (rw == UIO_READ) {
585 				rl_cookie = vn_rangelock_rlock(vp, offset,
586 				    offset + len);
587 			} else {
588 				rl_cookie = vn_rangelock_wlock(vp, offset,
589 				    offset + len);
590 			}
591 		} else
592 			rl_cookie = NULL;
593 		mp = NULL;
594 		if (rw == UIO_WRITE) {
595 			if (vp->v_type != VCHR &&
596 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
597 			    != 0)
598 				goto out;
599 			if (MNT_SHARED_WRITES(mp) ||
600 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
601 				lock_flags = LK_SHARED;
602 			else
603 				lock_flags = LK_EXCLUSIVE;
604 		} else
605 			lock_flags = LK_SHARED;
606 		vn_lock(vp, lock_flags | LK_RETRY);
607 	} else
608 		rl_cookie = NULL;
609 
610 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
611 #ifdef MAC
612 	if ((ioflg & IO_NOMACCHECK) == 0) {
613 		if (rw == UIO_READ)
614 			error = mac_vnode_check_read(active_cred, file_cred,
615 			    vp);
616 		else
617 			error = mac_vnode_check_write(active_cred, file_cred,
618 			    vp);
619 	}
620 #endif
621 	if (error == 0) {
622 		if (file_cred != NULL)
623 			cred = file_cred;
624 		else
625 			cred = active_cred;
626 		if (do_vn_io_fault(vp, &auio)) {
627 			args.kind = VN_IO_FAULT_VOP;
628 			args.cred = cred;
629 			args.flags = ioflg;
630 			args.args.vop_args.vp = vp;
631 			error = vn_io_fault1(vp, &auio, &args, td);
632 		} else if (rw == UIO_READ) {
633 			error = VOP_READ(vp, &auio, ioflg, cred);
634 		} else /* if (rw == UIO_WRITE) */ {
635 			error = VOP_WRITE(vp, &auio, ioflg, cred);
636 		}
637 	}
638 	if (aresid)
639 		*aresid = auio.uio_resid;
640 	else
641 		if (auio.uio_resid && error == 0)
642 			error = EIO;
643 	if ((ioflg & IO_NODELOCKED) == 0) {
644 		VOP_UNLOCK(vp);
645 		if (mp != NULL)
646 			vn_finished_write(mp);
647 	}
648  out:
649 	if (rl_cookie != NULL)
650 		vn_rangelock_unlock(vp, rl_cookie);
651 	return (error);
652 }
653 
654 /*
655  * Package up an I/O request on a vnode into a uio and do it.  The I/O
656  * request is split up into smaller chunks and we try to avoid saturating
657  * the buffer cache while potentially holding a vnode locked, so we
658  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
659  * to give other processes a chance to lock the vnode (either other processes
660  * core'ing the same binary, or unrelated processes scanning the directory).
661  */
662 int
663 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
664     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
665     struct ucred *file_cred, size_t *aresid, struct thread *td)
666 {
667 	int error = 0;
668 	ssize_t iaresid;
669 
670 	do {
671 		int chunk;
672 
673 		/*
674 		 * Force `offset' to a multiple of MAXBSIZE except possibly
675 		 * for the first chunk, so that filesystems only need to
676 		 * write full blocks except possibly for the first and last
677 		 * chunks.
678 		 */
679 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
680 
681 		if (chunk > len)
682 			chunk = len;
683 		if (rw != UIO_READ && vp->v_type == VREG)
684 			bwillwrite();
685 		iaresid = 0;
686 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
687 		    ioflg, active_cred, file_cred, &iaresid, td);
688 		len -= chunk;	/* aresid calc already includes length */
689 		if (error)
690 			break;
691 		offset += chunk;
692 		base = (char *)base + chunk;
693 		kern_yield(PRI_USER);
694 	} while (len);
695 	if (aresid)
696 		*aresid = len + iaresid;
697 	return (error);
698 }
699 
700 #if OFF_MAX <= LONG_MAX
701 off_t
702 foffset_lock(struct file *fp, int flags)
703 {
704 	volatile short *flagsp;
705 	off_t res;
706 	short state;
707 
708 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
709 
710 	if ((flags & FOF_NOLOCK) != 0)
711 		return (atomic_load_long(&fp->f_offset));
712 
713 	/*
714 	 * According to McKusick the vn lock was protecting f_offset here.
715 	 * It is now protected by the FOFFSET_LOCKED flag.
716 	 */
717 	flagsp = &fp->f_vnread_flags;
718 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
719 		return (atomic_load_long(&fp->f_offset));
720 
721 	sleepq_lock(&fp->f_vnread_flags);
722 	state = atomic_load_16(flagsp);
723 	for (;;) {
724 		if ((state & FOFFSET_LOCKED) == 0) {
725 			if (!atomic_fcmpset_acq_16(flagsp, &state,
726 			    FOFFSET_LOCKED))
727 				continue;
728 			break;
729 		}
730 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
731 			if (!atomic_fcmpset_acq_16(flagsp, &state,
732 			    state | FOFFSET_LOCK_WAITING))
733 				continue;
734 		}
735 		DROP_GIANT();
736 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
737 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
738 		PICKUP_GIANT();
739 		sleepq_lock(&fp->f_vnread_flags);
740 		state = atomic_load_16(flagsp);
741 	}
742 	res = atomic_load_long(&fp->f_offset);
743 	sleepq_release(&fp->f_vnread_flags);
744 	return (res);
745 }
746 
747 void
748 foffset_unlock(struct file *fp, off_t val, int flags)
749 {
750 	volatile short *flagsp;
751 	short state;
752 
753 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
754 
755 	if ((flags & FOF_NOUPDATE) == 0)
756 		atomic_store_long(&fp->f_offset, val);
757 	if ((flags & FOF_NEXTOFF_R) != 0)
758 		fp->f_nextoff[UIO_READ] = val;
759 	if ((flags & FOF_NEXTOFF_W) != 0)
760 		fp->f_nextoff[UIO_WRITE] = val;
761 
762 	if ((flags & FOF_NOLOCK) != 0)
763 		return;
764 
765 	flagsp = &fp->f_vnread_flags;
766 	state = atomic_load_16(flagsp);
767 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
768 	    atomic_cmpset_rel_16(flagsp, state, 0))
769 		return;
770 
771 	sleepq_lock(&fp->f_vnread_flags);
772 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
773 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
774 	fp->f_vnread_flags = 0;
775 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
776 	sleepq_release(&fp->f_vnread_flags);
777 }
778 #else
779 off_t
780 foffset_lock(struct file *fp, int flags)
781 {
782 	struct mtx *mtxp;
783 	off_t res;
784 
785 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
786 
787 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
788 	mtx_lock(mtxp);
789 	if ((flags & FOF_NOLOCK) == 0) {
790 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
791 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
792 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
793 			    "vofflock", 0);
794 		}
795 		fp->f_vnread_flags |= FOFFSET_LOCKED;
796 	}
797 	res = fp->f_offset;
798 	mtx_unlock(mtxp);
799 	return (res);
800 }
801 
802 void
803 foffset_unlock(struct file *fp, off_t val, int flags)
804 {
805 	struct mtx *mtxp;
806 
807 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
808 
809 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
810 	mtx_lock(mtxp);
811 	if ((flags & FOF_NOUPDATE) == 0)
812 		fp->f_offset = val;
813 	if ((flags & FOF_NEXTOFF_R) != 0)
814 		fp->f_nextoff[UIO_READ] = val;
815 	if ((flags & FOF_NEXTOFF_W) != 0)
816 		fp->f_nextoff[UIO_WRITE] = val;
817 	if ((flags & FOF_NOLOCK) == 0) {
818 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
819 		    ("Lost FOFFSET_LOCKED"));
820 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
821 			wakeup(&fp->f_vnread_flags);
822 		fp->f_vnread_flags = 0;
823 	}
824 	mtx_unlock(mtxp);
825 }
826 #endif
827 
828 void
829 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
830 {
831 
832 	if ((flags & FOF_OFFSET) == 0)
833 		uio->uio_offset = foffset_lock(fp, flags);
834 }
835 
836 void
837 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
838 {
839 
840 	if ((flags & FOF_OFFSET) == 0)
841 		foffset_unlock(fp, uio->uio_offset, flags);
842 }
843 
844 static int
845 get_advice(struct file *fp, struct uio *uio)
846 {
847 	struct mtx *mtxp;
848 	int ret;
849 
850 	ret = POSIX_FADV_NORMAL;
851 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
852 		return (ret);
853 
854 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
855 	mtx_lock(mtxp);
856 	if (fp->f_advice != NULL &&
857 	    uio->uio_offset >= fp->f_advice->fa_start &&
858 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
859 		ret = fp->f_advice->fa_advice;
860 	mtx_unlock(mtxp);
861 	return (ret);
862 }
863 
864 int
865 vn_read_from_obj(struct vnode *vp, struct uio *uio)
866 {
867 	vm_object_t obj;
868 	vm_page_t ma[io_hold_cnt + 2];
869 	off_t off, vsz;
870 	ssize_t resid;
871 	int error, i, j;
872 
873 	obj = vp->v_object;
874 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
875 	MPASS(obj != NULL);
876 	MPASS(obj->type == OBJT_VNODE);
877 
878 	/*
879 	 * Depends on type stability of vm_objects.
880 	 */
881 	vm_object_pip_add(obj, 1);
882 	if ((obj->flags & OBJ_DEAD) != 0) {
883 		/*
884 		 * Note that object might be already reused from the
885 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
886 		 * we recheck for DOOMED vnode state after all pages
887 		 * are busied, and retract then.
888 		 *
889 		 * But we check for OBJ_DEAD to ensure that we do not
890 		 * busy pages while vm_object_terminate_pages()
891 		 * processes the queue.
892 		 */
893 		error = EJUSTRETURN;
894 		goto out_pip;
895 	}
896 
897 	resid = uio->uio_resid;
898 	off = uio->uio_offset;
899 	for (i = 0; resid > 0; i++) {
900 		MPASS(i < io_hold_cnt + 2);
901 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
902 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
903 		    VM_ALLOC_NOWAIT);
904 		if (ma[i] == NULL)
905 			break;
906 
907 		/*
908 		 * Skip invalid pages.  Valid mask can be partial only
909 		 * at EOF, and we clip later.
910 		 */
911 		if (vm_page_none_valid(ma[i])) {
912 			vm_page_sunbusy(ma[i]);
913 			break;
914 		}
915 
916 		resid -= PAGE_SIZE;
917 		off += PAGE_SIZE;
918 	}
919 	if (i == 0) {
920 		error = EJUSTRETURN;
921 		goto out_pip;
922 	}
923 
924 	/*
925 	 * Check VIRF_DOOMED after we busied our pages.  Since
926 	 * vgonel() terminates the vnode' vm_object, it cannot
927 	 * process past pages busied by us.
928 	 */
929 	if (VN_IS_DOOMED(vp)) {
930 		error = EJUSTRETURN;
931 		goto out;
932 	}
933 
934 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
935 	if (resid > uio->uio_resid)
936 		resid = uio->uio_resid;
937 
938 	/*
939 	 * Unlocked read of vnp_size is safe because truncation cannot
940 	 * pass busied page.  But we load vnp_size into a local
941 	 * variable so that possible concurrent extension does not
942 	 * break calculation.
943 	 */
944 #if defined(__powerpc__) && !defined(__powerpc64__)
945 	vsz = obj->un_pager.vnp.vnp_size;
946 #else
947 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
948 #endif
949 	if (uio->uio_offset + resid > vsz)
950 		resid = vsz - uio->uio_offset;
951 
952 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
953 
954 out:
955 	for (j = 0; j < i; j++) {
956 		if (error == 0)
957 			vm_page_reference(ma[j]);
958 		vm_page_sunbusy(ma[j]);
959 	}
960 out_pip:
961 	vm_object_pip_wakeup(obj);
962 	if (error != 0)
963 		return (error);
964 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
965 }
966 
967 /*
968  * File table vnode read routine.
969  */
970 static int
971 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
972     struct thread *td)
973 {
974 	struct vnode *vp;
975 	off_t orig_offset;
976 	int error, ioflag;
977 	int advice;
978 
979 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
980 	    uio->uio_td, td));
981 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
982 	vp = fp->f_vnode;
983 	ioflag = 0;
984 	if (fp->f_flag & FNONBLOCK)
985 		ioflag |= IO_NDELAY;
986 	if (fp->f_flag & O_DIRECT)
987 		ioflag |= IO_DIRECT;
988 
989 	/*
990 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
991 	 * allows us to avoid unneeded work outright.
992 	 */
993 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
994 	    (vp->v_irflag & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
995 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
996 		if (error == 0) {
997 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
998 			return (0);
999 		}
1000 		if (error != EJUSTRETURN)
1001 			return (error);
1002 	}
1003 
1004 	advice = get_advice(fp, uio);
1005 	vn_lock(vp, LK_SHARED | LK_RETRY);
1006 
1007 	switch (advice) {
1008 	case POSIX_FADV_NORMAL:
1009 	case POSIX_FADV_SEQUENTIAL:
1010 	case POSIX_FADV_NOREUSE:
1011 		ioflag |= sequential_heuristic(uio, fp);
1012 		break;
1013 	case POSIX_FADV_RANDOM:
1014 		/* Disable read-ahead for random I/O. */
1015 		break;
1016 	}
1017 	orig_offset = uio->uio_offset;
1018 
1019 #ifdef MAC
1020 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1021 	if (error == 0)
1022 #endif
1023 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1024 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1025 	VOP_UNLOCK(vp);
1026 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1027 	    orig_offset != uio->uio_offset)
1028 		/*
1029 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1030 		 * for the backing file after a POSIX_FADV_NOREUSE
1031 		 * read(2).
1032 		 */
1033 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1034 		    POSIX_FADV_DONTNEED);
1035 	return (error);
1036 }
1037 
1038 /*
1039  * File table vnode write routine.
1040  */
1041 static int
1042 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1043     struct thread *td)
1044 {
1045 	struct vnode *vp;
1046 	struct mount *mp;
1047 	off_t orig_offset;
1048 	int error, ioflag, lock_flags;
1049 	int advice;
1050 
1051 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1052 	    uio->uio_td, td));
1053 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1054 	vp = fp->f_vnode;
1055 	if (vp->v_type == VREG)
1056 		bwillwrite();
1057 	ioflag = IO_UNIT;
1058 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1059 		ioflag |= IO_APPEND;
1060 	if (fp->f_flag & FNONBLOCK)
1061 		ioflag |= IO_NDELAY;
1062 	if (fp->f_flag & O_DIRECT)
1063 		ioflag |= IO_DIRECT;
1064 	if ((fp->f_flag & O_FSYNC) ||
1065 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1066 		ioflag |= IO_SYNC;
1067 	mp = NULL;
1068 	if (vp->v_type != VCHR &&
1069 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1070 		goto unlock;
1071 
1072 	advice = get_advice(fp, uio);
1073 
1074 	if (MNT_SHARED_WRITES(mp) ||
1075 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1076 		lock_flags = LK_SHARED;
1077 	} else {
1078 		lock_flags = LK_EXCLUSIVE;
1079 	}
1080 
1081 	vn_lock(vp, lock_flags | LK_RETRY);
1082 	switch (advice) {
1083 	case POSIX_FADV_NORMAL:
1084 	case POSIX_FADV_SEQUENTIAL:
1085 	case POSIX_FADV_NOREUSE:
1086 		ioflag |= sequential_heuristic(uio, fp);
1087 		break;
1088 	case POSIX_FADV_RANDOM:
1089 		/* XXX: Is this correct? */
1090 		break;
1091 	}
1092 	orig_offset = uio->uio_offset;
1093 
1094 #ifdef MAC
1095 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1096 	if (error == 0)
1097 #endif
1098 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1099 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1100 	VOP_UNLOCK(vp);
1101 	if (vp->v_type != VCHR)
1102 		vn_finished_write(mp);
1103 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1104 	    orig_offset != uio->uio_offset)
1105 		/*
1106 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1107 		 * for the backing file after a POSIX_FADV_NOREUSE
1108 		 * write(2).
1109 		 */
1110 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1111 		    POSIX_FADV_DONTNEED);
1112 unlock:
1113 	return (error);
1114 }
1115 
1116 /*
1117  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1118  * prevent the following deadlock:
1119  *
1120  * Assume that the thread A reads from the vnode vp1 into userspace
1121  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1122  * currently not resident, then system ends up with the call chain
1123  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1124  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1125  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1126  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1127  * backed by the pages of vnode vp1, and some page in buf2 is not
1128  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1129  *
1130  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1131  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1132  * Instead, it first tries to do the whole range i/o with pagefaults
1133  * disabled. If all pages in the i/o buffer are resident and mapped,
1134  * VOP will succeed (ignoring the genuine filesystem errors).
1135  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1136  * i/o in chunks, with all pages in the chunk prefaulted and held
1137  * using vm_fault_quick_hold_pages().
1138  *
1139  * Filesystems using this deadlock avoidance scheme should use the
1140  * array of the held pages from uio, saved in the curthread->td_ma,
1141  * instead of doing uiomove().  A helper function
1142  * vn_io_fault_uiomove() converts uiomove request into
1143  * uiomove_fromphys() over td_ma array.
1144  *
1145  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1146  * make the current i/o request atomic with respect to other i/os and
1147  * truncations.
1148  */
1149 
1150 /*
1151  * Decode vn_io_fault_args and perform the corresponding i/o.
1152  */
1153 static int
1154 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1155     struct thread *td)
1156 {
1157 	int error, save;
1158 
1159 	error = 0;
1160 	save = vm_fault_disable_pagefaults();
1161 	switch (args->kind) {
1162 	case VN_IO_FAULT_FOP:
1163 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1164 		    uio, args->cred, args->flags, td);
1165 		break;
1166 	case VN_IO_FAULT_VOP:
1167 		if (uio->uio_rw == UIO_READ) {
1168 			error = VOP_READ(args->args.vop_args.vp, uio,
1169 			    args->flags, args->cred);
1170 		} else if (uio->uio_rw == UIO_WRITE) {
1171 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1172 			    args->flags, args->cred);
1173 		}
1174 		break;
1175 	default:
1176 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1177 		    args->kind, uio->uio_rw);
1178 	}
1179 	vm_fault_enable_pagefaults(save);
1180 	return (error);
1181 }
1182 
1183 static int
1184 vn_io_fault_touch(char *base, const struct uio *uio)
1185 {
1186 	int r;
1187 
1188 	r = fubyte(base);
1189 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1190 		return (EFAULT);
1191 	return (0);
1192 }
1193 
1194 static int
1195 vn_io_fault_prefault_user(const struct uio *uio)
1196 {
1197 	char *base;
1198 	const struct iovec *iov;
1199 	size_t len;
1200 	ssize_t resid;
1201 	int error, i;
1202 
1203 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1204 	    ("vn_io_fault_prefault userspace"));
1205 
1206 	error = i = 0;
1207 	iov = uio->uio_iov;
1208 	resid = uio->uio_resid;
1209 	base = iov->iov_base;
1210 	len = iov->iov_len;
1211 	while (resid > 0) {
1212 		error = vn_io_fault_touch(base, uio);
1213 		if (error != 0)
1214 			break;
1215 		if (len < PAGE_SIZE) {
1216 			if (len != 0) {
1217 				error = vn_io_fault_touch(base + len - 1, uio);
1218 				if (error != 0)
1219 					break;
1220 				resid -= len;
1221 			}
1222 			if (++i >= uio->uio_iovcnt)
1223 				break;
1224 			iov = uio->uio_iov + i;
1225 			base = iov->iov_base;
1226 			len = iov->iov_len;
1227 		} else {
1228 			len -= PAGE_SIZE;
1229 			base += PAGE_SIZE;
1230 			resid -= PAGE_SIZE;
1231 		}
1232 	}
1233 	return (error);
1234 }
1235 
1236 /*
1237  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1238  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1239  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1240  * into args and call vn_io_fault1() to handle faults during the user
1241  * mode buffer accesses.
1242  */
1243 static int
1244 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1245     struct thread *td)
1246 {
1247 	vm_page_t ma[io_hold_cnt + 2];
1248 	struct uio *uio_clone, short_uio;
1249 	struct iovec short_iovec[1];
1250 	vm_page_t *prev_td_ma;
1251 	vm_prot_t prot;
1252 	vm_offset_t addr, end;
1253 	size_t len, resid;
1254 	ssize_t adv;
1255 	int error, cnt, saveheld, prev_td_ma_cnt;
1256 
1257 	if (vn_io_fault_prefault) {
1258 		error = vn_io_fault_prefault_user(uio);
1259 		if (error != 0)
1260 			return (error); /* Or ignore ? */
1261 	}
1262 
1263 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1264 
1265 	/*
1266 	 * The UFS follows IO_UNIT directive and replays back both
1267 	 * uio_offset and uio_resid if an error is encountered during the
1268 	 * operation.  But, since the iovec may be already advanced,
1269 	 * uio is still in an inconsistent state.
1270 	 *
1271 	 * Cache a copy of the original uio, which is advanced to the redo
1272 	 * point using UIO_NOCOPY below.
1273 	 */
1274 	uio_clone = cloneuio(uio);
1275 	resid = uio->uio_resid;
1276 
1277 	short_uio.uio_segflg = UIO_USERSPACE;
1278 	short_uio.uio_rw = uio->uio_rw;
1279 	short_uio.uio_td = uio->uio_td;
1280 
1281 	error = vn_io_fault_doio(args, uio, td);
1282 	if (error != EFAULT)
1283 		goto out;
1284 
1285 	atomic_add_long(&vn_io_faults_cnt, 1);
1286 	uio_clone->uio_segflg = UIO_NOCOPY;
1287 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1288 	uio_clone->uio_segflg = uio->uio_segflg;
1289 
1290 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1291 	prev_td_ma = td->td_ma;
1292 	prev_td_ma_cnt = td->td_ma_cnt;
1293 
1294 	while (uio_clone->uio_resid != 0) {
1295 		len = uio_clone->uio_iov->iov_len;
1296 		if (len == 0) {
1297 			KASSERT(uio_clone->uio_iovcnt >= 1,
1298 			    ("iovcnt underflow"));
1299 			uio_clone->uio_iov++;
1300 			uio_clone->uio_iovcnt--;
1301 			continue;
1302 		}
1303 		if (len > ptoa(io_hold_cnt))
1304 			len = ptoa(io_hold_cnt);
1305 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1306 		end = round_page(addr + len);
1307 		if (end < addr) {
1308 			error = EFAULT;
1309 			break;
1310 		}
1311 		cnt = atop(end - trunc_page(addr));
1312 		/*
1313 		 * A perfectly misaligned address and length could cause
1314 		 * both the start and the end of the chunk to use partial
1315 		 * page.  +2 accounts for such a situation.
1316 		 */
1317 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1318 		    addr, len, prot, ma, io_hold_cnt + 2);
1319 		if (cnt == -1) {
1320 			error = EFAULT;
1321 			break;
1322 		}
1323 		short_uio.uio_iov = &short_iovec[0];
1324 		short_iovec[0].iov_base = (void *)addr;
1325 		short_uio.uio_iovcnt = 1;
1326 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1327 		short_uio.uio_offset = uio_clone->uio_offset;
1328 		td->td_ma = ma;
1329 		td->td_ma_cnt = cnt;
1330 
1331 		error = vn_io_fault_doio(args, &short_uio, td);
1332 		vm_page_unhold_pages(ma, cnt);
1333 		adv = len - short_uio.uio_resid;
1334 
1335 		uio_clone->uio_iov->iov_base =
1336 		    (char *)uio_clone->uio_iov->iov_base + adv;
1337 		uio_clone->uio_iov->iov_len -= adv;
1338 		uio_clone->uio_resid -= adv;
1339 		uio_clone->uio_offset += adv;
1340 
1341 		uio->uio_resid -= adv;
1342 		uio->uio_offset += adv;
1343 
1344 		if (error != 0 || adv == 0)
1345 			break;
1346 	}
1347 	td->td_ma = prev_td_ma;
1348 	td->td_ma_cnt = prev_td_ma_cnt;
1349 	curthread_pflags_restore(saveheld);
1350 out:
1351 	free(uio_clone, M_IOV);
1352 	return (error);
1353 }
1354 
1355 static int
1356 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1357     int flags, struct thread *td)
1358 {
1359 	fo_rdwr_t *doio;
1360 	struct vnode *vp;
1361 	void *rl_cookie;
1362 	struct vn_io_fault_args args;
1363 	int error;
1364 
1365 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1366 	vp = fp->f_vnode;
1367 
1368 	/*
1369 	 * The ability to read(2) on a directory has historically been
1370 	 * allowed for all users, but this can and has been the source of
1371 	 * at least one security issue in the past.  As such, it is now hidden
1372 	 * away behind a sysctl for those that actually need it to use it, and
1373 	 * restricted to root when it's turned on to make it relatively safe to
1374 	 * leave on for longer sessions of need.
1375 	 */
1376 	if (vp->v_type == VDIR) {
1377 		KASSERT(uio->uio_rw == UIO_READ,
1378 		    ("illegal write attempted on a directory"));
1379 		if (!vfs_allow_read_dir)
1380 			return (EISDIR);
1381 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1382 			return (EISDIR);
1383 	}
1384 
1385 	foffset_lock_uio(fp, uio, flags);
1386 	if (do_vn_io_fault(vp, uio)) {
1387 		args.kind = VN_IO_FAULT_FOP;
1388 		args.args.fop_args.fp = fp;
1389 		args.args.fop_args.doio = doio;
1390 		args.cred = active_cred;
1391 		args.flags = flags | FOF_OFFSET;
1392 		if (uio->uio_rw == UIO_READ) {
1393 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1394 			    uio->uio_offset + uio->uio_resid);
1395 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1396 		    (flags & FOF_OFFSET) == 0) {
1397 			/* For appenders, punt and lock the whole range. */
1398 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1399 		} else {
1400 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1401 			    uio->uio_offset + uio->uio_resid);
1402 		}
1403 		error = vn_io_fault1(vp, uio, &args, td);
1404 		vn_rangelock_unlock(vp, rl_cookie);
1405 	} else {
1406 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1407 	}
1408 	foffset_unlock_uio(fp, uio, flags);
1409 	return (error);
1410 }
1411 
1412 /*
1413  * Helper function to perform the requested uiomove operation using
1414  * the held pages for io->uio_iov[0].iov_base buffer instead of
1415  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1416  * instead of iov_base prevents page faults that could occur due to
1417  * pmap_collect() invalidating the mapping created by
1418  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1419  * object cleanup revoking the write access from page mappings.
1420  *
1421  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1422  * instead of plain uiomove().
1423  */
1424 int
1425 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1426 {
1427 	struct uio transp_uio;
1428 	struct iovec transp_iov[1];
1429 	struct thread *td;
1430 	size_t adv;
1431 	int error, pgadv;
1432 
1433 	td = curthread;
1434 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1435 	    uio->uio_segflg != UIO_USERSPACE)
1436 		return (uiomove(data, xfersize, uio));
1437 
1438 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1439 	transp_iov[0].iov_base = data;
1440 	transp_uio.uio_iov = &transp_iov[0];
1441 	transp_uio.uio_iovcnt = 1;
1442 	if (xfersize > uio->uio_resid)
1443 		xfersize = uio->uio_resid;
1444 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1445 	transp_uio.uio_offset = 0;
1446 	transp_uio.uio_segflg = UIO_SYSSPACE;
1447 	/*
1448 	 * Since transp_iov points to data, and td_ma page array
1449 	 * corresponds to original uio->uio_iov, we need to invert the
1450 	 * direction of the i/o operation as passed to
1451 	 * uiomove_fromphys().
1452 	 */
1453 	switch (uio->uio_rw) {
1454 	case UIO_WRITE:
1455 		transp_uio.uio_rw = UIO_READ;
1456 		break;
1457 	case UIO_READ:
1458 		transp_uio.uio_rw = UIO_WRITE;
1459 		break;
1460 	}
1461 	transp_uio.uio_td = uio->uio_td;
1462 	error = uiomove_fromphys(td->td_ma,
1463 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1464 	    xfersize, &transp_uio);
1465 	adv = xfersize - transp_uio.uio_resid;
1466 	pgadv =
1467 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1468 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1469 	td->td_ma += pgadv;
1470 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1471 	    pgadv));
1472 	td->td_ma_cnt -= pgadv;
1473 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1474 	uio->uio_iov->iov_len -= adv;
1475 	uio->uio_resid -= adv;
1476 	uio->uio_offset += adv;
1477 	return (error);
1478 }
1479 
1480 int
1481 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1482     struct uio *uio)
1483 {
1484 	struct thread *td;
1485 	vm_offset_t iov_base;
1486 	int cnt, pgadv;
1487 
1488 	td = curthread;
1489 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1490 	    uio->uio_segflg != UIO_USERSPACE)
1491 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1492 
1493 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1494 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1495 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1496 	switch (uio->uio_rw) {
1497 	case UIO_WRITE:
1498 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1499 		    offset, cnt);
1500 		break;
1501 	case UIO_READ:
1502 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1503 		    cnt);
1504 		break;
1505 	}
1506 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1507 	td->td_ma += pgadv;
1508 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1509 	    pgadv));
1510 	td->td_ma_cnt -= pgadv;
1511 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1512 	uio->uio_iov->iov_len -= cnt;
1513 	uio->uio_resid -= cnt;
1514 	uio->uio_offset += cnt;
1515 	return (0);
1516 }
1517 
1518 /*
1519  * File table truncate routine.
1520  */
1521 static int
1522 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1523     struct thread *td)
1524 {
1525 	struct mount *mp;
1526 	struct vnode *vp;
1527 	void *rl_cookie;
1528 	int error;
1529 
1530 	vp = fp->f_vnode;
1531 
1532 retry:
1533 	/*
1534 	 * Lock the whole range for truncation.  Otherwise split i/o
1535 	 * might happen partly before and partly after the truncation.
1536 	 */
1537 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1538 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1539 	if (error)
1540 		goto out1;
1541 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1542 	AUDIT_ARG_VNODE1(vp);
1543 	if (vp->v_type == VDIR) {
1544 		error = EISDIR;
1545 		goto out;
1546 	}
1547 #ifdef MAC
1548 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1549 	if (error)
1550 		goto out;
1551 #endif
1552 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1553 	    fp->f_cred);
1554 out:
1555 	VOP_UNLOCK(vp);
1556 	vn_finished_write(mp);
1557 out1:
1558 	vn_rangelock_unlock(vp, rl_cookie);
1559 	if (error == ERELOOKUP)
1560 		goto retry;
1561 	return (error);
1562 }
1563 
1564 /*
1565  * Truncate a file that is already locked.
1566  */
1567 int
1568 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1569     struct ucred *cred)
1570 {
1571 	struct vattr vattr;
1572 	int error;
1573 
1574 	error = VOP_ADD_WRITECOUNT(vp, 1);
1575 	if (error == 0) {
1576 		VATTR_NULL(&vattr);
1577 		vattr.va_size = length;
1578 		if (sync)
1579 			vattr.va_vaflags |= VA_SYNC;
1580 		error = VOP_SETATTR(vp, &vattr, cred);
1581 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1582 	}
1583 	return (error);
1584 }
1585 
1586 /*
1587  * File table vnode stat routine.
1588  */
1589 static int
1590 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1591     struct thread *td)
1592 {
1593 	struct vnode *vp = fp->f_vnode;
1594 	int error;
1595 
1596 	vn_lock(vp, LK_SHARED | LK_RETRY);
1597 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1598 	VOP_UNLOCK(vp);
1599 
1600 	return (error);
1601 }
1602 
1603 /*
1604  * File table vnode ioctl routine.
1605  */
1606 static int
1607 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1608     struct thread *td)
1609 {
1610 	struct vattr vattr;
1611 	struct vnode *vp;
1612 	struct fiobmap2_arg *bmarg;
1613 	int error;
1614 
1615 	vp = fp->f_vnode;
1616 	switch (vp->v_type) {
1617 	case VDIR:
1618 	case VREG:
1619 		switch (com) {
1620 		case FIONREAD:
1621 			vn_lock(vp, LK_SHARED | LK_RETRY);
1622 			error = VOP_GETATTR(vp, &vattr, active_cred);
1623 			VOP_UNLOCK(vp);
1624 			if (error == 0)
1625 				*(int *)data = vattr.va_size - fp->f_offset;
1626 			return (error);
1627 		case FIOBMAP2:
1628 			bmarg = (struct fiobmap2_arg *)data;
1629 			vn_lock(vp, LK_SHARED | LK_RETRY);
1630 #ifdef MAC
1631 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1632 			    vp);
1633 			if (error == 0)
1634 #endif
1635 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1636 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1637 			VOP_UNLOCK(vp);
1638 			return (error);
1639 		case FIONBIO:
1640 		case FIOASYNC:
1641 			return (0);
1642 		default:
1643 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1644 			    active_cred, td));
1645 		}
1646 		break;
1647 	case VCHR:
1648 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1649 		    active_cred, td));
1650 	default:
1651 		return (ENOTTY);
1652 	}
1653 }
1654 
1655 /*
1656  * File table vnode poll routine.
1657  */
1658 static int
1659 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1660     struct thread *td)
1661 {
1662 	struct vnode *vp;
1663 	int error;
1664 
1665 	vp = fp->f_vnode;
1666 #if defined(MAC) || defined(AUDIT)
1667 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1668 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1669 		AUDIT_ARG_VNODE1(vp);
1670 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1671 		VOP_UNLOCK(vp);
1672 		if (error != 0)
1673 			return (error);
1674 	}
1675 #endif
1676 	error = VOP_POLL(vp, events, fp->f_cred, td);
1677 	return (error);
1678 }
1679 
1680 /*
1681  * Acquire the requested lock and then check for validity.  LK_RETRY
1682  * permits vn_lock to return doomed vnodes.
1683  */
1684 static int __noinline
1685 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1686     int error)
1687 {
1688 
1689 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1690 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1691 
1692 	if (error == 0)
1693 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1694 
1695 	if ((flags & LK_RETRY) == 0) {
1696 		if (error == 0) {
1697 			VOP_UNLOCK(vp);
1698 			error = ENOENT;
1699 		}
1700 		return (error);
1701 	}
1702 
1703 	/*
1704 	 * LK_RETRY case.
1705 	 *
1706 	 * Nothing to do if we got the lock.
1707 	 */
1708 	if (error == 0)
1709 		return (0);
1710 
1711 	/*
1712 	 * Interlock was dropped by the call in _vn_lock.
1713 	 */
1714 	flags &= ~LK_INTERLOCK;
1715 	do {
1716 		error = VOP_LOCK1(vp, flags, file, line);
1717 	} while (error != 0);
1718 	return (0);
1719 }
1720 
1721 int
1722 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1723 {
1724 	int error;
1725 
1726 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1727 	    ("vn_lock: no locktype (%d passed)", flags));
1728 	VNPASS(vp->v_holdcnt > 0, vp);
1729 	error = VOP_LOCK1(vp, flags, file, line);
1730 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1731 		return (_vn_lock_fallback(vp, flags, file, line, error));
1732 	return (0);
1733 }
1734 
1735 /*
1736  * File table vnode close routine.
1737  */
1738 static int
1739 vn_closefile(struct file *fp, struct thread *td)
1740 {
1741 	struct vnode *vp;
1742 	struct flock lf;
1743 	int error;
1744 	bool ref;
1745 
1746 	vp = fp->f_vnode;
1747 	fp->f_ops = &badfileops;
1748 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1749 
1750 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1751 
1752 	if (__predict_false(ref)) {
1753 		lf.l_whence = SEEK_SET;
1754 		lf.l_start = 0;
1755 		lf.l_len = 0;
1756 		lf.l_type = F_UNLCK;
1757 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1758 		vrele(vp);
1759 	}
1760 	return (error);
1761 }
1762 
1763 /*
1764  * Preparing to start a filesystem write operation. If the operation is
1765  * permitted, then we bump the count of operations in progress and
1766  * proceed. If a suspend request is in progress, we wait until the
1767  * suspension is over, and then proceed.
1768  */
1769 static int
1770 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1771 {
1772 	struct mount_pcpu *mpcpu;
1773 	int error, mflags;
1774 
1775 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1776 	    vfs_op_thread_enter(mp, mpcpu)) {
1777 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1778 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1779 		vfs_op_thread_exit(mp, mpcpu);
1780 		return (0);
1781 	}
1782 
1783 	if (mplocked)
1784 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1785 	else
1786 		MNT_ILOCK(mp);
1787 
1788 	error = 0;
1789 
1790 	/*
1791 	 * Check on status of suspension.
1792 	 */
1793 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1794 	    mp->mnt_susp_owner != curthread) {
1795 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1796 		    (flags & PCATCH) : 0) | (PUSER - 1);
1797 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1798 			if (flags & V_NOWAIT) {
1799 				error = EWOULDBLOCK;
1800 				goto unlock;
1801 			}
1802 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1803 			    "suspfs", 0);
1804 			if (error)
1805 				goto unlock;
1806 		}
1807 	}
1808 	if (flags & V_XSLEEP)
1809 		goto unlock;
1810 	mp->mnt_writeopcount++;
1811 unlock:
1812 	if (error != 0 || (flags & V_XSLEEP) != 0)
1813 		MNT_REL(mp);
1814 	MNT_IUNLOCK(mp);
1815 	return (error);
1816 }
1817 
1818 int
1819 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1820 {
1821 	struct mount *mp;
1822 	int error;
1823 
1824 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1825 	    ("V_MNTREF requires mp"));
1826 
1827 	error = 0;
1828 	/*
1829 	 * If a vnode is provided, get and return the mount point that
1830 	 * to which it will write.
1831 	 */
1832 	if (vp != NULL) {
1833 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1834 			*mpp = NULL;
1835 			if (error != EOPNOTSUPP)
1836 				return (error);
1837 			return (0);
1838 		}
1839 	}
1840 	if ((mp = *mpp) == NULL)
1841 		return (0);
1842 
1843 	/*
1844 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1845 	 * a vfs_ref().
1846 	 * As long as a vnode is not provided we need to acquire a
1847 	 * refcount for the provided mountpoint too, in order to
1848 	 * emulate a vfs_ref().
1849 	 */
1850 	if (vp == NULL && (flags & V_MNTREF) == 0)
1851 		vfs_ref(mp);
1852 
1853 	return (vn_start_write_refed(mp, flags, false));
1854 }
1855 
1856 /*
1857  * Secondary suspension. Used by operations such as vop_inactive
1858  * routines that are needed by the higher level functions. These
1859  * are allowed to proceed until all the higher level functions have
1860  * completed (indicated by mnt_writeopcount dropping to zero). At that
1861  * time, these operations are halted until the suspension is over.
1862  */
1863 int
1864 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1865 {
1866 	struct mount *mp;
1867 	int error;
1868 
1869 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1870 	    ("V_MNTREF requires mp"));
1871 
1872  retry:
1873 	if (vp != NULL) {
1874 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1875 			*mpp = NULL;
1876 			if (error != EOPNOTSUPP)
1877 				return (error);
1878 			return (0);
1879 		}
1880 	}
1881 	/*
1882 	 * If we are not suspended or have not yet reached suspended
1883 	 * mode, then let the operation proceed.
1884 	 */
1885 	if ((mp = *mpp) == NULL)
1886 		return (0);
1887 
1888 	/*
1889 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1890 	 * a vfs_ref().
1891 	 * As long as a vnode is not provided we need to acquire a
1892 	 * refcount for the provided mountpoint too, in order to
1893 	 * emulate a vfs_ref().
1894 	 */
1895 	MNT_ILOCK(mp);
1896 	if (vp == NULL && (flags & V_MNTREF) == 0)
1897 		MNT_REF(mp);
1898 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1899 		mp->mnt_secondary_writes++;
1900 		mp->mnt_secondary_accwrites++;
1901 		MNT_IUNLOCK(mp);
1902 		return (0);
1903 	}
1904 	if (flags & V_NOWAIT) {
1905 		MNT_REL(mp);
1906 		MNT_IUNLOCK(mp);
1907 		return (EWOULDBLOCK);
1908 	}
1909 	/*
1910 	 * Wait for the suspension to finish.
1911 	 */
1912 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1913 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1914 	    "suspfs", 0);
1915 	vfs_rel(mp);
1916 	if (error == 0)
1917 		goto retry;
1918 	return (error);
1919 }
1920 
1921 /*
1922  * Filesystem write operation has completed. If we are suspending and this
1923  * operation is the last one, notify the suspender that the suspension is
1924  * now in effect.
1925  */
1926 void
1927 vn_finished_write(struct mount *mp)
1928 {
1929 	struct mount_pcpu *mpcpu;
1930 	int c;
1931 
1932 	if (mp == NULL)
1933 		return;
1934 
1935 	if (vfs_op_thread_enter(mp, mpcpu)) {
1936 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
1937 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
1938 		vfs_op_thread_exit(mp, mpcpu);
1939 		return;
1940 	}
1941 
1942 	MNT_ILOCK(mp);
1943 	vfs_assert_mount_counters(mp);
1944 	MNT_REL(mp);
1945 	c = --mp->mnt_writeopcount;
1946 	if (mp->mnt_vfs_ops == 0) {
1947 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1948 		MNT_IUNLOCK(mp);
1949 		return;
1950 	}
1951 	if (c < 0)
1952 		vfs_dump_mount_counters(mp);
1953 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1954 		wakeup(&mp->mnt_writeopcount);
1955 	MNT_IUNLOCK(mp);
1956 }
1957 
1958 /*
1959  * Filesystem secondary write operation has completed. If we are
1960  * suspending and this operation is the last one, notify the suspender
1961  * that the suspension is now in effect.
1962  */
1963 void
1964 vn_finished_secondary_write(struct mount *mp)
1965 {
1966 	if (mp == NULL)
1967 		return;
1968 	MNT_ILOCK(mp);
1969 	MNT_REL(mp);
1970 	mp->mnt_secondary_writes--;
1971 	if (mp->mnt_secondary_writes < 0)
1972 		panic("vn_finished_secondary_write: neg cnt");
1973 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1974 	    mp->mnt_secondary_writes <= 0)
1975 		wakeup(&mp->mnt_secondary_writes);
1976 	MNT_IUNLOCK(mp);
1977 }
1978 
1979 /*
1980  * Request a filesystem to suspend write operations.
1981  */
1982 int
1983 vfs_write_suspend(struct mount *mp, int flags)
1984 {
1985 	int error;
1986 
1987 	vfs_op_enter(mp);
1988 
1989 	MNT_ILOCK(mp);
1990 	vfs_assert_mount_counters(mp);
1991 	if (mp->mnt_susp_owner == curthread) {
1992 		vfs_op_exit_locked(mp);
1993 		MNT_IUNLOCK(mp);
1994 		return (EALREADY);
1995 	}
1996 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
1997 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1998 
1999 	/*
2000 	 * Unmount holds a write reference on the mount point.  If we
2001 	 * own busy reference and drain for writers, we deadlock with
2002 	 * the reference draining in the unmount path.  Callers of
2003 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2004 	 * vfs_busy() reference is owned and caller is not in the
2005 	 * unmount context.
2006 	 */
2007 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2008 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2009 		vfs_op_exit_locked(mp);
2010 		MNT_IUNLOCK(mp);
2011 		return (EBUSY);
2012 	}
2013 
2014 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2015 	mp->mnt_susp_owner = curthread;
2016 	if (mp->mnt_writeopcount > 0)
2017 		(void) msleep(&mp->mnt_writeopcount,
2018 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2019 	else
2020 		MNT_IUNLOCK(mp);
2021 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2022 		vfs_write_resume(mp, 0);
2023 		/* vfs_write_resume does vfs_op_exit() for us */
2024 	}
2025 	return (error);
2026 }
2027 
2028 /*
2029  * Request a filesystem to resume write operations.
2030  */
2031 void
2032 vfs_write_resume(struct mount *mp, int flags)
2033 {
2034 
2035 	MNT_ILOCK(mp);
2036 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2037 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2038 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2039 				       MNTK_SUSPENDED);
2040 		mp->mnt_susp_owner = NULL;
2041 		wakeup(&mp->mnt_writeopcount);
2042 		wakeup(&mp->mnt_flag);
2043 		curthread->td_pflags &= ~TDP_IGNSUSP;
2044 		if ((flags & VR_START_WRITE) != 0) {
2045 			MNT_REF(mp);
2046 			mp->mnt_writeopcount++;
2047 		}
2048 		MNT_IUNLOCK(mp);
2049 		if ((flags & VR_NO_SUSPCLR) == 0)
2050 			VFS_SUSP_CLEAN(mp);
2051 		vfs_op_exit(mp);
2052 	} else if ((flags & VR_START_WRITE) != 0) {
2053 		MNT_REF(mp);
2054 		vn_start_write_refed(mp, 0, true);
2055 	} else {
2056 		MNT_IUNLOCK(mp);
2057 	}
2058 }
2059 
2060 /*
2061  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2062  * methods.
2063  */
2064 int
2065 vfs_write_suspend_umnt(struct mount *mp)
2066 {
2067 	int error;
2068 
2069 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2070 	    ("vfs_write_suspend_umnt: recursed"));
2071 
2072 	/* dounmount() already called vn_start_write(). */
2073 	for (;;) {
2074 		vn_finished_write(mp);
2075 		error = vfs_write_suspend(mp, 0);
2076 		if (error != 0) {
2077 			vn_start_write(NULL, &mp, V_WAIT);
2078 			return (error);
2079 		}
2080 		MNT_ILOCK(mp);
2081 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2082 			break;
2083 		MNT_IUNLOCK(mp);
2084 		vn_start_write(NULL, &mp, V_WAIT);
2085 	}
2086 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2087 	wakeup(&mp->mnt_flag);
2088 	MNT_IUNLOCK(mp);
2089 	curthread->td_pflags |= TDP_IGNSUSP;
2090 	return (0);
2091 }
2092 
2093 /*
2094  * Implement kqueues for files by translating it to vnode operation.
2095  */
2096 static int
2097 vn_kqfilter(struct file *fp, struct knote *kn)
2098 {
2099 
2100 	return (VOP_KQFILTER(fp->f_vnode, kn));
2101 }
2102 
2103 /*
2104  * Simplified in-kernel wrapper calls for extended attribute access.
2105  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2106  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2107  */
2108 int
2109 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2110     const char *attrname, int *buflen, char *buf, struct thread *td)
2111 {
2112 	struct uio	auio;
2113 	struct iovec	iov;
2114 	int	error;
2115 
2116 	iov.iov_len = *buflen;
2117 	iov.iov_base = buf;
2118 
2119 	auio.uio_iov = &iov;
2120 	auio.uio_iovcnt = 1;
2121 	auio.uio_rw = UIO_READ;
2122 	auio.uio_segflg = UIO_SYSSPACE;
2123 	auio.uio_td = td;
2124 	auio.uio_offset = 0;
2125 	auio.uio_resid = *buflen;
2126 
2127 	if ((ioflg & IO_NODELOCKED) == 0)
2128 		vn_lock(vp, LK_SHARED | LK_RETRY);
2129 
2130 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2131 
2132 	/* authorize attribute retrieval as kernel */
2133 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2134 	    td);
2135 
2136 	if ((ioflg & IO_NODELOCKED) == 0)
2137 		VOP_UNLOCK(vp);
2138 
2139 	if (error == 0) {
2140 		*buflen = *buflen - auio.uio_resid;
2141 	}
2142 
2143 	return (error);
2144 }
2145 
2146 /*
2147  * XXX failure mode if partially written?
2148  */
2149 int
2150 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2151     const char *attrname, int buflen, char *buf, struct thread *td)
2152 {
2153 	struct uio	auio;
2154 	struct iovec	iov;
2155 	struct mount	*mp;
2156 	int	error;
2157 
2158 	iov.iov_len = buflen;
2159 	iov.iov_base = buf;
2160 
2161 	auio.uio_iov = &iov;
2162 	auio.uio_iovcnt = 1;
2163 	auio.uio_rw = UIO_WRITE;
2164 	auio.uio_segflg = UIO_SYSSPACE;
2165 	auio.uio_td = td;
2166 	auio.uio_offset = 0;
2167 	auio.uio_resid = buflen;
2168 
2169 	if ((ioflg & IO_NODELOCKED) == 0) {
2170 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2171 			return (error);
2172 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2173 	}
2174 
2175 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2176 
2177 	/* authorize attribute setting as kernel */
2178 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2179 
2180 	if ((ioflg & IO_NODELOCKED) == 0) {
2181 		vn_finished_write(mp);
2182 		VOP_UNLOCK(vp);
2183 	}
2184 
2185 	return (error);
2186 }
2187 
2188 int
2189 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2190     const char *attrname, struct thread *td)
2191 {
2192 	struct mount	*mp;
2193 	int	error;
2194 
2195 	if ((ioflg & IO_NODELOCKED) == 0) {
2196 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2197 			return (error);
2198 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2199 	}
2200 
2201 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2202 
2203 	/* authorize attribute removal as kernel */
2204 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2205 	if (error == EOPNOTSUPP)
2206 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2207 		    NULL, td);
2208 
2209 	if ((ioflg & IO_NODELOCKED) == 0) {
2210 		vn_finished_write(mp);
2211 		VOP_UNLOCK(vp);
2212 	}
2213 
2214 	return (error);
2215 }
2216 
2217 static int
2218 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2219     struct vnode **rvp)
2220 {
2221 
2222 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2223 }
2224 
2225 int
2226 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2227 {
2228 
2229 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2230 	    lkflags, rvp));
2231 }
2232 
2233 int
2234 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2235     int lkflags, struct vnode **rvp)
2236 {
2237 	struct mount *mp;
2238 	int ltype, error;
2239 
2240 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2241 	mp = vp->v_mount;
2242 	ltype = VOP_ISLOCKED(vp);
2243 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2244 	    ("vn_vget_ino: vp not locked"));
2245 	error = vfs_busy(mp, MBF_NOWAIT);
2246 	if (error != 0) {
2247 		vfs_ref(mp);
2248 		VOP_UNLOCK(vp);
2249 		error = vfs_busy(mp, 0);
2250 		vn_lock(vp, ltype | LK_RETRY);
2251 		vfs_rel(mp);
2252 		if (error != 0)
2253 			return (ENOENT);
2254 		if (VN_IS_DOOMED(vp)) {
2255 			vfs_unbusy(mp);
2256 			return (ENOENT);
2257 		}
2258 	}
2259 	VOP_UNLOCK(vp);
2260 	error = alloc(mp, alloc_arg, lkflags, rvp);
2261 	vfs_unbusy(mp);
2262 	if (error != 0 || *rvp != vp)
2263 		vn_lock(vp, ltype | LK_RETRY);
2264 	if (VN_IS_DOOMED(vp)) {
2265 		if (error == 0) {
2266 			if (*rvp == vp)
2267 				vunref(vp);
2268 			else
2269 				vput(*rvp);
2270 		}
2271 		error = ENOENT;
2272 	}
2273 	return (error);
2274 }
2275 
2276 int
2277 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2278     struct thread *td)
2279 {
2280 
2281 	if (vp->v_type != VREG || td == NULL)
2282 		return (0);
2283 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2284 	    lim_cur(td, RLIMIT_FSIZE)) {
2285 		PROC_LOCK(td->td_proc);
2286 		kern_psignal(td->td_proc, SIGXFSZ);
2287 		PROC_UNLOCK(td->td_proc);
2288 		return (EFBIG);
2289 	}
2290 	return (0);
2291 }
2292 
2293 int
2294 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2295     struct thread *td)
2296 {
2297 	struct vnode *vp;
2298 
2299 	vp = fp->f_vnode;
2300 #ifdef AUDIT
2301 	vn_lock(vp, LK_SHARED | LK_RETRY);
2302 	AUDIT_ARG_VNODE1(vp);
2303 	VOP_UNLOCK(vp);
2304 #endif
2305 	return (setfmode(td, active_cred, vp, mode));
2306 }
2307 
2308 int
2309 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2310     struct thread *td)
2311 {
2312 	struct vnode *vp;
2313 
2314 	vp = fp->f_vnode;
2315 #ifdef AUDIT
2316 	vn_lock(vp, LK_SHARED | LK_RETRY);
2317 	AUDIT_ARG_VNODE1(vp);
2318 	VOP_UNLOCK(vp);
2319 #endif
2320 	return (setfown(td, active_cred, vp, uid, gid));
2321 }
2322 
2323 void
2324 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2325 {
2326 	vm_object_t object;
2327 
2328 	if ((object = vp->v_object) == NULL)
2329 		return;
2330 	VM_OBJECT_WLOCK(object);
2331 	vm_object_page_remove(object, start, end, 0);
2332 	VM_OBJECT_WUNLOCK(object);
2333 }
2334 
2335 int
2336 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2337 {
2338 	struct vattr va;
2339 	daddr_t bn, bnp;
2340 	uint64_t bsize;
2341 	off_t noff;
2342 	int error;
2343 
2344 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2345 	    ("Wrong command %lu", cmd));
2346 
2347 	if (vn_lock(vp, LK_SHARED) != 0)
2348 		return (EBADF);
2349 	if (vp->v_type != VREG) {
2350 		error = ENOTTY;
2351 		goto unlock;
2352 	}
2353 	error = VOP_GETATTR(vp, &va, cred);
2354 	if (error != 0)
2355 		goto unlock;
2356 	noff = *off;
2357 	if (noff >= va.va_size) {
2358 		error = ENXIO;
2359 		goto unlock;
2360 	}
2361 	bsize = vp->v_mount->mnt_stat.f_iosize;
2362 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2363 	    noff % bsize) {
2364 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2365 		if (error == EOPNOTSUPP) {
2366 			error = ENOTTY;
2367 			goto unlock;
2368 		}
2369 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2370 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2371 			noff = bn * bsize;
2372 			if (noff < *off)
2373 				noff = *off;
2374 			goto unlock;
2375 		}
2376 	}
2377 	if (noff > va.va_size)
2378 		noff = va.va_size;
2379 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2380 	if (cmd == FIOSEEKDATA)
2381 		error = ENXIO;
2382 unlock:
2383 	VOP_UNLOCK(vp);
2384 	if (error == 0)
2385 		*off = noff;
2386 	return (error);
2387 }
2388 
2389 int
2390 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2391 {
2392 	struct ucred *cred;
2393 	struct vnode *vp;
2394 	struct vattr vattr;
2395 	off_t foffset, size;
2396 	int error, noneg;
2397 
2398 	cred = td->td_ucred;
2399 	vp = fp->f_vnode;
2400 	foffset = foffset_lock(fp, 0);
2401 	noneg = (vp->v_type != VCHR);
2402 	error = 0;
2403 	switch (whence) {
2404 	case L_INCR:
2405 		if (noneg &&
2406 		    (foffset < 0 ||
2407 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2408 			error = EOVERFLOW;
2409 			break;
2410 		}
2411 		offset += foffset;
2412 		break;
2413 	case L_XTND:
2414 		vn_lock(vp, LK_SHARED | LK_RETRY);
2415 		error = VOP_GETATTR(vp, &vattr, cred);
2416 		VOP_UNLOCK(vp);
2417 		if (error)
2418 			break;
2419 
2420 		/*
2421 		 * If the file references a disk device, then fetch
2422 		 * the media size and use that to determine the ending
2423 		 * offset.
2424 		 */
2425 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2426 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2427 			vattr.va_size = size;
2428 		if (noneg &&
2429 		    (vattr.va_size > OFF_MAX ||
2430 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2431 			error = EOVERFLOW;
2432 			break;
2433 		}
2434 		offset += vattr.va_size;
2435 		break;
2436 	case L_SET:
2437 		break;
2438 	case SEEK_DATA:
2439 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2440 		if (error == ENOTTY)
2441 			error = EINVAL;
2442 		break;
2443 	case SEEK_HOLE:
2444 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2445 		if (error == ENOTTY)
2446 			error = EINVAL;
2447 		break;
2448 	default:
2449 		error = EINVAL;
2450 	}
2451 	if (error == 0 && noneg && offset < 0)
2452 		error = EINVAL;
2453 	if (error != 0)
2454 		goto drop;
2455 	VFS_KNOTE_UNLOCKED(vp, 0);
2456 	td->td_uretoff.tdu_off = offset;
2457 drop:
2458 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2459 	return (error);
2460 }
2461 
2462 int
2463 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2464     struct thread *td)
2465 {
2466 	int error;
2467 
2468 	/*
2469 	 * Grant permission if the caller is the owner of the file, or
2470 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2471 	 * on the file.  If the time pointer is null, then write
2472 	 * permission on the file is also sufficient.
2473 	 *
2474 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2475 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2476 	 * will be allowed to set the times [..] to the current
2477 	 * server time.
2478 	 */
2479 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2480 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2481 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2482 	return (error);
2483 }
2484 
2485 int
2486 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2487 {
2488 	struct vnode *vp;
2489 	int error;
2490 
2491 	if (fp->f_type == DTYPE_FIFO)
2492 		kif->kf_type = KF_TYPE_FIFO;
2493 	else
2494 		kif->kf_type = KF_TYPE_VNODE;
2495 	vp = fp->f_vnode;
2496 	vref(vp);
2497 	FILEDESC_SUNLOCK(fdp);
2498 	error = vn_fill_kinfo_vnode(vp, kif);
2499 	vrele(vp);
2500 	FILEDESC_SLOCK(fdp);
2501 	return (error);
2502 }
2503 
2504 static inline void
2505 vn_fill_junk(struct kinfo_file *kif)
2506 {
2507 	size_t len, olen;
2508 
2509 	/*
2510 	 * Simulate vn_fullpath returning changing values for a given
2511 	 * vp during e.g. coredump.
2512 	 */
2513 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2514 	olen = strlen(kif->kf_path);
2515 	if (len < olen)
2516 		strcpy(&kif->kf_path[len - 1], "$");
2517 	else
2518 		for (; olen < len; olen++)
2519 			strcpy(&kif->kf_path[olen], "A");
2520 }
2521 
2522 int
2523 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2524 {
2525 	struct vattr va;
2526 	char *fullpath, *freepath;
2527 	int error;
2528 
2529 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2530 	freepath = NULL;
2531 	fullpath = "-";
2532 	error = vn_fullpath(vp, &fullpath, &freepath);
2533 	if (error == 0) {
2534 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2535 	}
2536 	if (freepath != NULL)
2537 		free(freepath, M_TEMP);
2538 
2539 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2540 		vn_fill_junk(kif);
2541 	);
2542 
2543 	/*
2544 	 * Retrieve vnode attributes.
2545 	 */
2546 	va.va_fsid = VNOVAL;
2547 	va.va_rdev = NODEV;
2548 	vn_lock(vp, LK_SHARED | LK_RETRY);
2549 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2550 	VOP_UNLOCK(vp);
2551 	if (error != 0)
2552 		return (error);
2553 	if (va.va_fsid != VNOVAL)
2554 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2555 	else
2556 		kif->kf_un.kf_file.kf_file_fsid =
2557 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2558 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2559 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2560 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2561 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2562 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2563 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2564 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2565 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2566 	return (0);
2567 }
2568 
2569 int
2570 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2571     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2572     struct thread *td)
2573 {
2574 #ifdef HWPMC_HOOKS
2575 	struct pmckern_map_in pkm;
2576 #endif
2577 	struct mount *mp;
2578 	struct vnode *vp;
2579 	vm_object_t object;
2580 	vm_prot_t maxprot;
2581 	boolean_t writecounted;
2582 	int error;
2583 
2584 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2585     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2586 	/*
2587 	 * POSIX shared-memory objects are defined to have
2588 	 * kernel persistence, and are not defined to support
2589 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2590 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2591 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2592 	 * flag to request this behavior.
2593 	 */
2594 	if ((fp->f_flag & FPOSIXSHM) != 0)
2595 		flags |= MAP_NOSYNC;
2596 #endif
2597 	vp = fp->f_vnode;
2598 
2599 	/*
2600 	 * Ensure that file and memory protections are
2601 	 * compatible.  Note that we only worry about
2602 	 * writability if mapping is shared; in this case,
2603 	 * current and max prot are dictated by the open file.
2604 	 * XXX use the vnode instead?  Problem is: what
2605 	 * credentials do we use for determination? What if
2606 	 * proc does a setuid?
2607 	 */
2608 	mp = vp->v_mount;
2609 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2610 		maxprot = VM_PROT_NONE;
2611 		if ((prot & VM_PROT_EXECUTE) != 0)
2612 			return (EACCES);
2613 	} else
2614 		maxprot = VM_PROT_EXECUTE;
2615 	if ((fp->f_flag & FREAD) != 0)
2616 		maxprot |= VM_PROT_READ;
2617 	else if ((prot & VM_PROT_READ) != 0)
2618 		return (EACCES);
2619 
2620 	/*
2621 	 * If we are sharing potential changes via MAP_SHARED and we
2622 	 * are trying to get write permission although we opened it
2623 	 * without asking for it, bail out.
2624 	 */
2625 	if ((flags & MAP_SHARED) != 0) {
2626 		if ((fp->f_flag & FWRITE) != 0)
2627 			maxprot |= VM_PROT_WRITE;
2628 		else if ((prot & VM_PROT_WRITE) != 0)
2629 			return (EACCES);
2630 	} else {
2631 		maxprot |= VM_PROT_WRITE;
2632 		cap_maxprot |= VM_PROT_WRITE;
2633 	}
2634 	maxprot &= cap_maxprot;
2635 
2636 	/*
2637 	 * For regular files and shared memory, POSIX requires that
2638 	 * the value of foff be a legitimate offset within the data
2639 	 * object.  In particular, negative offsets are invalid.
2640 	 * Blocking negative offsets and overflows here avoids
2641 	 * possible wraparound or user-level access into reserved
2642 	 * ranges of the data object later.  In contrast, POSIX does
2643 	 * not dictate how offsets are used by device drivers, so in
2644 	 * the case of a device mapping a negative offset is passed
2645 	 * on.
2646 	 */
2647 	if (
2648 #ifdef _LP64
2649 	    size > OFF_MAX ||
2650 #endif
2651 	    foff > OFF_MAX - size)
2652 		return (EINVAL);
2653 
2654 	writecounted = FALSE;
2655 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2656 	    &foff, &object, &writecounted);
2657 	if (error != 0)
2658 		return (error);
2659 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2660 	    foff, writecounted, td);
2661 	if (error != 0) {
2662 		/*
2663 		 * If this mapping was accounted for in the vnode's
2664 		 * writecount, then undo that now.
2665 		 */
2666 		if (writecounted)
2667 			vm_pager_release_writecount(object, 0, size);
2668 		vm_object_deallocate(object);
2669 	}
2670 #ifdef HWPMC_HOOKS
2671 	/* Inform hwpmc(4) if an executable is being mapped. */
2672 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2673 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2674 			pkm.pm_file = vp;
2675 			pkm.pm_address = (uintptr_t) *addr;
2676 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2677 		}
2678 	}
2679 #endif
2680 	return (error);
2681 }
2682 
2683 void
2684 vn_fsid(struct vnode *vp, struct vattr *va)
2685 {
2686 	fsid_t *f;
2687 
2688 	f = &vp->v_mount->mnt_stat.f_fsid;
2689 	va->va_fsid = (uint32_t)f->val[1];
2690 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2691 	va->va_fsid += (uint32_t)f->val[0];
2692 }
2693 
2694 int
2695 vn_fsync_buf(struct vnode *vp, int waitfor)
2696 {
2697 	struct buf *bp, *nbp;
2698 	struct bufobj *bo;
2699 	struct mount *mp;
2700 	int error, maxretry;
2701 
2702 	error = 0;
2703 	maxretry = 10000;     /* large, arbitrarily chosen */
2704 	mp = NULL;
2705 	if (vp->v_type == VCHR) {
2706 		VI_LOCK(vp);
2707 		mp = vp->v_rdev->si_mountpt;
2708 		VI_UNLOCK(vp);
2709 	}
2710 	bo = &vp->v_bufobj;
2711 	BO_LOCK(bo);
2712 loop1:
2713 	/*
2714 	 * MARK/SCAN initialization to avoid infinite loops.
2715 	 */
2716         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2717 		bp->b_vflags &= ~BV_SCANNED;
2718 		bp->b_error = 0;
2719 	}
2720 
2721 	/*
2722 	 * Flush all dirty buffers associated with a vnode.
2723 	 */
2724 loop2:
2725 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2726 		if ((bp->b_vflags & BV_SCANNED) != 0)
2727 			continue;
2728 		bp->b_vflags |= BV_SCANNED;
2729 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2730 			if (waitfor != MNT_WAIT)
2731 				continue;
2732 			if (BUF_LOCK(bp,
2733 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2734 			    BO_LOCKPTR(bo)) != 0) {
2735 				BO_LOCK(bo);
2736 				goto loop1;
2737 			}
2738 			BO_LOCK(bo);
2739 		}
2740 		BO_UNLOCK(bo);
2741 		KASSERT(bp->b_bufobj == bo,
2742 		    ("bp %p wrong b_bufobj %p should be %p",
2743 		    bp, bp->b_bufobj, bo));
2744 		if ((bp->b_flags & B_DELWRI) == 0)
2745 			panic("fsync: not dirty");
2746 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2747 			vfs_bio_awrite(bp);
2748 		} else {
2749 			bremfree(bp);
2750 			bawrite(bp);
2751 		}
2752 		if (maxretry < 1000)
2753 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2754 		BO_LOCK(bo);
2755 		goto loop2;
2756 	}
2757 
2758 	/*
2759 	 * If synchronous the caller expects us to completely resolve all
2760 	 * dirty buffers in the system.  Wait for in-progress I/O to
2761 	 * complete (which could include background bitmap writes), then
2762 	 * retry if dirty blocks still exist.
2763 	 */
2764 	if (waitfor == MNT_WAIT) {
2765 		bufobj_wwait(bo, 0, 0);
2766 		if (bo->bo_dirty.bv_cnt > 0) {
2767 			/*
2768 			 * If we are unable to write any of these buffers
2769 			 * then we fail now rather than trying endlessly
2770 			 * to write them out.
2771 			 */
2772 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2773 				if ((error = bp->b_error) != 0)
2774 					break;
2775 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2776 			    (error == 0 && --maxretry >= 0))
2777 				goto loop1;
2778 			if (error == 0)
2779 				error = EAGAIN;
2780 		}
2781 	}
2782 	BO_UNLOCK(bo);
2783 	if (error != 0)
2784 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2785 
2786 	return (error);
2787 }
2788 
2789 /*
2790  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2791  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2792  * to do the actual copy.
2793  * vn_generic_copy_file_range() is factored out, so it can be called
2794  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2795  * different file systems.
2796  */
2797 int
2798 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2799     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2800     struct ucred *outcred, struct thread *fsize_td)
2801 {
2802 	int error;
2803 	size_t len;
2804 	uint64_t uval;
2805 
2806 	len = *lenp;
2807 	*lenp = 0;		/* For error returns. */
2808 	error = 0;
2809 
2810 	/* Do some sanity checks on the arguments. */
2811 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2812 		error = EISDIR;
2813 	else if (*inoffp < 0 || *outoffp < 0 ||
2814 	    invp->v_type != VREG || outvp->v_type != VREG)
2815 		error = EINVAL;
2816 	if (error != 0)
2817 		goto out;
2818 
2819 	/* Ensure offset + len does not wrap around. */
2820 	uval = *inoffp;
2821 	uval += len;
2822 	if (uval > INT64_MAX)
2823 		len = INT64_MAX - *inoffp;
2824 	uval = *outoffp;
2825 	uval += len;
2826 	if (uval > INT64_MAX)
2827 		len = INT64_MAX - *outoffp;
2828 	if (len == 0)
2829 		goto out;
2830 
2831 	/*
2832 	 * If the two vnode are for the same file system, call
2833 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2834 	 * which can handle copies across multiple file systems.
2835 	 */
2836 	*lenp = len;
2837 	if (invp->v_mount == outvp->v_mount)
2838 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2839 		    lenp, flags, incred, outcred, fsize_td);
2840 	else
2841 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2842 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2843 out:
2844 	return (error);
2845 }
2846 
2847 /*
2848  * Test len bytes of data starting at dat for all bytes == 0.
2849  * Return true if all bytes are zero, false otherwise.
2850  * Expects dat to be well aligned.
2851  */
2852 static bool
2853 mem_iszero(void *dat, int len)
2854 {
2855 	int i;
2856 	const u_int *p;
2857 	const char *cp;
2858 
2859 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2860 		if (len >= sizeof(*p)) {
2861 			if (*p != 0)
2862 				return (false);
2863 		} else {
2864 			cp = (const char *)p;
2865 			for (i = 0; i < len; i++, cp++)
2866 				if (*cp != '\0')
2867 					return (false);
2868 		}
2869 	}
2870 	return (true);
2871 }
2872 
2873 /*
2874  * Look for a hole in the output file and, if found, adjust *outoffp
2875  * and *xferp to skip past the hole.
2876  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2877  * to be written as 0's upon return.
2878  */
2879 static off_t
2880 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2881     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2882 {
2883 	int error;
2884 	off_t delta;
2885 
2886 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2887 		*dataoffp = *outoffp;
2888 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2889 		    curthread);
2890 		if (error == 0) {
2891 			*holeoffp = *dataoffp;
2892 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2893 			    curthread);
2894 		}
2895 		if (error != 0 || *holeoffp == *dataoffp) {
2896 			/*
2897 			 * Since outvp is unlocked, it may be possible for
2898 			 * another thread to do a truncate(), lseek(), write()
2899 			 * creating a hole at startoff between the above
2900 			 * VOP_IOCTL() calls, if the other thread does not do
2901 			 * rangelocking.
2902 			 * If that happens, *holeoffp == *dataoffp and finding
2903 			 * the hole has failed, so disable vn_skip_hole().
2904 			 */
2905 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2906 			return (xfer2);
2907 		}
2908 		KASSERT(*dataoffp >= *outoffp,
2909 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2910 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2911 		KASSERT(*holeoffp > *dataoffp,
2912 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2913 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2914 	}
2915 
2916 	/*
2917 	 * If there is a hole before the data starts, advance *outoffp and
2918 	 * *xferp past the hole.
2919 	 */
2920 	if (*dataoffp > *outoffp) {
2921 		delta = *dataoffp - *outoffp;
2922 		if (delta >= *xferp) {
2923 			/* Entire *xferp is a hole. */
2924 			*outoffp += *xferp;
2925 			*xferp = 0;
2926 			return (0);
2927 		}
2928 		*xferp -= delta;
2929 		*outoffp += delta;
2930 		xfer2 = MIN(xfer2, *xferp);
2931 	}
2932 
2933 	/*
2934 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2935 	 * that the write ends at the start of the hole.
2936 	 * *holeoffp should always be greater than *outoffp, but for the
2937 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2938 	 * value.
2939 	 */
2940 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2941 		xfer2 = *holeoffp - *outoffp;
2942 	return (xfer2);
2943 }
2944 
2945 /*
2946  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2947  * dat is a maximum of blksize in length and can be written repeatedly in
2948  * the chunk.
2949  * If growfile == true, just grow the file via vn_truncate_locked() instead
2950  * of doing actual writes.
2951  * If checkhole == true, a hole is being punched, so skip over any hole
2952  * already in the output file.
2953  */
2954 static int
2955 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2956     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2957 {
2958 	struct mount *mp;
2959 	off_t dataoff, holeoff, xfer2;
2960 	int error, lckf;
2961 
2962 	/*
2963 	 * Loop around doing writes of blksize until write has been completed.
2964 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2965 	 * done for each iteration, since the xfer argument can be very
2966 	 * large if there is a large hole to punch in the output file.
2967 	 */
2968 	error = 0;
2969 	holeoff = 0;
2970 	do {
2971 		xfer2 = MIN(xfer, blksize);
2972 		if (checkhole) {
2973 			/*
2974 			 * Punching a hole.  Skip writing if there is
2975 			 * already a hole in the output file.
2976 			 */
2977 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
2978 			    &dataoff, &holeoff, cred);
2979 			if (xfer == 0)
2980 				break;
2981 			if (holeoff < 0)
2982 				checkhole = false;
2983 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
2984 			    (intmax_t)xfer2));
2985 		}
2986 		bwillwrite();
2987 		mp = NULL;
2988 		error = vn_start_write(outvp, &mp, V_WAIT);
2989 		if (error != 0)
2990 			break;
2991 		if (growfile) {
2992 			error = vn_lock(outvp, LK_EXCLUSIVE);
2993 			if (error == 0) {
2994 				error = vn_truncate_locked(outvp, outoff + xfer,
2995 				    false, cred);
2996 				VOP_UNLOCK(outvp);
2997 			}
2998 		} else {
2999 			if (MNT_SHARED_WRITES(mp))
3000 				lckf = LK_SHARED;
3001 			else
3002 				lckf = LK_EXCLUSIVE;
3003 			error = vn_lock(outvp, lckf);
3004 			if (error == 0) {
3005 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3006 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3007 				    curthread->td_ucred, cred, NULL, curthread);
3008 				outoff += xfer2;
3009 				xfer -= xfer2;
3010 			}
3011 			VOP_UNLOCK(outvp);
3012 		}
3013 		if (mp != NULL)
3014 			vn_finished_write(mp);
3015 	} while (!growfile && xfer > 0 && error == 0);
3016 	return (error);
3017 }
3018 
3019 /*
3020  * Copy a byte range of one file to another.  This function can handle the
3021  * case where invp and outvp are on different file systems.
3022  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3023  * is no better file system specific way to do it.
3024  */
3025 int
3026 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3027     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3028     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3029 {
3030 	struct vattr va;
3031 	struct mount *mp;
3032 	struct uio io;
3033 	off_t startoff, endoff, xfer, xfer2;
3034 	u_long blksize;
3035 	int error, interrupted;
3036 	bool cantseek, readzeros, eof, lastblock;
3037 	ssize_t aresid;
3038 	size_t copylen, len, rem, savlen;
3039 	char *dat;
3040 	long holein, holeout;
3041 
3042 	holein = holeout = 0;
3043 	savlen = len = *lenp;
3044 	error = 0;
3045 	interrupted = 0;
3046 	dat = NULL;
3047 
3048 	error = vn_lock(invp, LK_SHARED);
3049 	if (error != 0)
3050 		goto out;
3051 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3052 		holein = 0;
3053 	VOP_UNLOCK(invp);
3054 
3055 	mp = NULL;
3056 	error = vn_start_write(outvp, &mp, V_WAIT);
3057 	if (error == 0)
3058 		error = vn_lock(outvp, LK_EXCLUSIVE);
3059 	if (error == 0) {
3060 		/*
3061 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3062 		 * now that outvp is locked.
3063 		 */
3064 		if (fsize_td != NULL) {
3065 			io.uio_offset = *outoffp;
3066 			io.uio_resid = len;
3067 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3068 			if (error != 0)
3069 				error = EFBIG;
3070 		}
3071 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3072 			holeout = 0;
3073 		/*
3074 		 * Holes that are past EOF do not need to be written as a block
3075 		 * of zero bytes.  So, truncate the output file as far as
3076 		 * possible and then use va.va_size to decide if writing 0
3077 		 * bytes is necessary in the loop below.
3078 		 */
3079 		if (error == 0)
3080 			error = VOP_GETATTR(outvp, &va, outcred);
3081 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3082 		    *outoffp + len) {
3083 #ifdef MAC
3084 			error = mac_vnode_check_write(curthread->td_ucred,
3085 			    outcred, outvp);
3086 			if (error == 0)
3087 #endif
3088 				error = vn_truncate_locked(outvp, *outoffp,
3089 				    false, outcred);
3090 			if (error == 0)
3091 				va.va_size = *outoffp;
3092 		}
3093 		VOP_UNLOCK(outvp);
3094 	}
3095 	if (mp != NULL)
3096 		vn_finished_write(mp);
3097 	if (error != 0)
3098 		goto out;
3099 
3100 	/*
3101 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3102 	 * If hole sizes aren't available, set the blksize to the larger
3103 	 * f_iosize of invp and outvp.
3104 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3105 	 * This value is clipped at 4Kbytes and 1Mbyte.
3106 	 */
3107 	blksize = MAX(holein, holeout);
3108 
3109 	/* Clip len to end at an exact multiple of hole size. */
3110 	if (blksize > 1) {
3111 		rem = *inoffp % blksize;
3112 		if (rem > 0)
3113 			rem = blksize - rem;
3114 		if (len - rem > blksize)
3115 			len = savlen = rounddown(len - rem, blksize) + rem;
3116 	}
3117 
3118 	if (blksize <= 1)
3119 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3120 		    outvp->v_mount->mnt_stat.f_iosize);
3121 	if (blksize < 4096)
3122 		blksize = 4096;
3123 	else if (blksize > 1024 * 1024)
3124 		blksize = 1024 * 1024;
3125 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3126 
3127 	/*
3128 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3129 	 * to find holes.  Otherwise, just scan the read block for all 0s
3130 	 * in the inner loop where the data copying is done.
3131 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3132 	 * support holes on the server, but do not support FIOSEEKHOLE.
3133 	 */
3134 	eof = false;
3135 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3136 		endoff = 0;			/* To shut up compilers. */
3137 		cantseek = true;
3138 		startoff = *inoffp;
3139 		copylen = len;
3140 
3141 		/*
3142 		 * Find the next data area.  If there is just a hole to EOF,
3143 		 * FIOSEEKDATA should fail and then we drop down into the
3144 		 * inner loop and create the hole on the outvp file.
3145 		 * (I do not know if any file system will report a hole to
3146 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3147 		 *  will fail for those file systems.)
3148 		 *
3149 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3150 		 * the code just falls through to the inner copy loop.
3151 		 */
3152 		error = EINVAL;
3153 		if (holein > 0)
3154 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3155 			    incred, curthread);
3156 		if (error == 0) {
3157 			endoff = startoff;
3158 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3159 			    incred, curthread);
3160 			/*
3161 			 * Since invp is unlocked, it may be possible for
3162 			 * another thread to do a truncate(), lseek(), write()
3163 			 * creating a hole at startoff between the above
3164 			 * VOP_IOCTL() calls, if the other thread does not do
3165 			 * rangelocking.
3166 			 * If that happens, startoff == endoff and finding
3167 			 * the hole has failed, so set an error.
3168 			 */
3169 			if (error == 0 && startoff == endoff)
3170 				error = EINVAL; /* Any error. Reset to 0. */
3171 		}
3172 		if (error == 0) {
3173 			if (startoff > *inoffp) {
3174 				/* Found hole before data block. */
3175 				xfer = MIN(startoff - *inoffp, len);
3176 				if (*outoffp < va.va_size) {
3177 					/* Must write 0s to punch hole. */
3178 					xfer2 = MIN(va.va_size - *outoffp,
3179 					    xfer);
3180 					memset(dat, 0, MIN(xfer2, blksize));
3181 					error = vn_write_outvp(outvp, dat,
3182 					    *outoffp, xfer2, blksize, false,
3183 					    holeout > 0, outcred);
3184 				}
3185 
3186 				if (error == 0 && *outoffp + xfer >
3187 				    va.va_size && xfer == len)
3188 					/* Grow last block. */
3189 					error = vn_write_outvp(outvp, dat,
3190 					    *outoffp, xfer, blksize, true,
3191 					    false, outcred);
3192 				if (error == 0) {
3193 					*inoffp += xfer;
3194 					*outoffp += xfer;
3195 					len -= xfer;
3196 					if (len < savlen)
3197 						interrupted = sig_intr();
3198 				}
3199 			}
3200 			copylen = MIN(len, endoff - startoff);
3201 			cantseek = false;
3202 		} else {
3203 			cantseek = true;
3204 			startoff = *inoffp;
3205 			copylen = len;
3206 			error = 0;
3207 		}
3208 
3209 		xfer = blksize;
3210 		if (cantseek) {
3211 			/*
3212 			 * Set first xfer to end at a block boundary, so that
3213 			 * holes are more likely detected in the loop below via
3214 			 * the for all bytes 0 method.
3215 			 */
3216 			xfer -= (*inoffp % blksize);
3217 		}
3218 		/* Loop copying the data block. */
3219 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3220 			if (copylen < xfer)
3221 				xfer = copylen;
3222 			error = vn_lock(invp, LK_SHARED);
3223 			if (error != 0)
3224 				goto out;
3225 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3226 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3227 			    curthread->td_ucred, incred, &aresid,
3228 			    curthread);
3229 			VOP_UNLOCK(invp);
3230 			lastblock = false;
3231 			if (error == 0 && aresid > 0) {
3232 				/* Stop the copy at EOF on the input file. */
3233 				xfer -= aresid;
3234 				eof = true;
3235 				lastblock = true;
3236 			}
3237 			if (error == 0) {
3238 				/*
3239 				 * Skip the write for holes past the initial EOF
3240 				 * of the output file, unless this is the last
3241 				 * write of the output file at EOF.
3242 				 */
3243 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3244 				    false;
3245 				if (xfer == len)
3246 					lastblock = true;
3247 				if (!cantseek || *outoffp < va.va_size ||
3248 				    lastblock || !readzeros)
3249 					error = vn_write_outvp(outvp, dat,
3250 					    *outoffp, xfer, blksize,
3251 					    readzeros && lastblock &&
3252 					    *outoffp >= va.va_size, false,
3253 					    outcred);
3254 				if (error == 0) {
3255 					*inoffp += xfer;
3256 					startoff += xfer;
3257 					*outoffp += xfer;
3258 					copylen -= xfer;
3259 					len -= xfer;
3260 					if (len < savlen)
3261 						interrupted = sig_intr();
3262 				}
3263 			}
3264 			xfer = blksize;
3265 		}
3266 	}
3267 out:
3268 	*lenp = savlen - len;
3269 	free(dat, M_TEMP);
3270 	return (error);
3271 }
3272 
3273 static int
3274 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3275 {
3276 	struct mount *mp;
3277 	struct vnode *vp;
3278 	off_t olen, ooffset;
3279 	int error;
3280 #ifdef AUDIT
3281 	int audited_vnode1 = 0;
3282 #endif
3283 
3284 	vp = fp->f_vnode;
3285 	if (vp->v_type != VREG)
3286 		return (ENODEV);
3287 
3288 	/* Allocating blocks may take a long time, so iterate. */
3289 	for (;;) {
3290 		olen = len;
3291 		ooffset = offset;
3292 
3293 		bwillwrite();
3294 		mp = NULL;
3295 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3296 		if (error != 0)
3297 			break;
3298 		error = vn_lock(vp, LK_EXCLUSIVE);
3299 		if (error != 0) {
3300 			vn_finished_write(mp);
3301 			break;
3302 		}
3303 #ifdef AUDIT
3304 		if (!audited_vnode1) {
3305 			AUDIT_ARG_VNODE1(vp);
3306 			audited_vnode1 = 1;
3307 		}
3308 #endif
3309 #ifdef MAC
3310 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3311 		if (error == 0)
3312 #endif
3313 			error = VOP_ALLOCATE(vp, &offset, &len);
3314 		VOP_UNLOCK(vp);
3315 		vn_finished_write(mp);
3316 
3317 		if (olen + ooffset != offset + len) {
3318 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3319 			    ooffset, olen, offset, len);
3320 		}
3321 		if (error != 0 || len == 0)
3322 			break;
3323 		KASSERT(olen > len, ("Iteration did not make progress?"));
3324 		maybe_yield();
3325 	}
3326 
3327 	return (error);
3328 }
3329 
3330 static u_long vn_lock_pair_pause_cnt;
3331 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3332     &vn_lock_pair_pause_cnt, 0,
3333     "Count of vn_lock_pair deadlocks");
3334 
3335 static void
3336 vn_lock_pair_pause(const char *wmesg)
3337 {
3338 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3339 	pause(wmesg, prng32_bounded(hz / 10));
3340 }
3341 
3342 /*
3343  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3344  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3345  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3346  * can be NULL.
3347  *
3348  * The function returns with both vnodes exclusively locked, and
3349  * guarantees that it does not create lock order reversal with other
3350  * threads during its execution.  Both vnodes could be unlocked
3351  * temporary (and reclaimed).
3352  */
3353 void
3354 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3355     bool vp2_locked)
3356 {
3357 	int error;
3358 
3359 	if (vp1 == NULL && vp2 == NULL)
3360 		return;
3361 	if (vp1 != NULL) {
3362 		if (vp1_locked)
3363 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3364 		else
3365 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3366 	} else {
3367 		vp1_locked = true;
3368 	}
3369 	if (vp2 != NULL) {
3370 		if (vp2_locked)
3371 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3372 		else
3373 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3374 	} else {
3375 		vp2_locked = true;
3376 	}
3377 	if (!vp1_locked && !vp2_locked) {
3378 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3379 		vp1_locked = true;
3380 	}
3381 
3382 	for (;;) {
3383 		if (vp1_locked && vp2_locked)
3384 			break;
3385 		if (vp1_locked && vp2 != NULL) {
3386 			if (vp1 != NULL) {
3387 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3388 				    __FILE__, __LINE__);
3389 				if (error == 0)
3390 					break;
3391 				VOP_UNLOCK(vp1);
3392 				vp1_locked = false;
3393 				vn_lock_pair_pause("vlp1");
3394 			}
3395 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3396 			vp2_locked = true;
3397 		}
3398 		if (vp2_locked && vp1 != NULL) {
3399 			if (vp2 != NULL) {
3400 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3401 				    __FILE__, __LINE__);
3402 				if (error == 0)
3403 					break;
3404 				VOP_UNLOCK(vp2);
3405 				vp2_locked = false;
3406 				vn_lock_pair_pause("vlp2");
3407 			}
3408 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3409 			vp1_locked = true;
3410 		}
3411 	}
3412 	if (vp1 != NULL)
3413 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3414 	if (vp2 != NULL)
3415 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3416 }
3417