xref: /freebsd/sys/kern/vfs_vnops.c (revision 105fd928b0b5b35ab529e5f6914788dc49582901)
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 #include <sys/ktrace.h>
83 
84 #include <security/audit/audit.h>
85 #include <security/mac/mac_framework.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_extern.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pager.h>
94 
95 #ifdef HWPMC_HOOKS
96 #include <sys/pmckern.h>
97 #endif
98 
99 static fo_rdwr_t	vn_read;
100 static fo_rdwr_t	vn_write;
101 static fo_rdwr_t	vn_io_fault;
102 static fo_truncate_t	vn_truncate;
103 static fo_ioctl_t	vn_ioctl;
104 static fo_poll_t	vn_poll;
105 static fo_kqfilter_t	vn_kqfilter;
106 static fo_close_t	vn_closefile;
107 static fo_mmap_t	vn_mmap;
108 static fo_fallocate_t	vn_fallocate;
109 static fo_fspacectl_t	vn_fspacectl;
110 
111 struct 	fileops vnops = {
112 	.fo_read = vn_io_fault,
113 	.fo_write = vn_io_fault,
114 	.fo_truncate = vn_truncate,
115 	.fo_ioctl = vn_ioctl,
116 	.fo_poll = vn_poll,
117 	.fo_kqfilter = vn_kqfilter,
118 	.fo_stat = vn_statfile,
119 	.fo_close = vn_closefile,
120 	.fo_chmod = vn_chmod,
121 	.fo_chown = vn_chown,
122 	.fo_sendfile = vn_sendfile,
123 	.fo_seek = vn_seek,
124 	.fo_fill_kinfo = vn_fill_kinfo,
125 	.fo_mmap = vn_mmap,
126 	.fo_fallocate = vn_fallocate,
127 	.fo_fspacectl = vn_fspacectl,
128 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
129 };
130 
131 const u_int io_hold_cnt = 16;
132 static int vn_io_fault_enable = 1;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
134     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
135 static int vn_io_fault_prefault = 0;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
137     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
138 static int vn_io_pgcache_read_enable = 1;
139 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
140     &vn_io_pgcache_read_enable, 0,
141     "Enable copying from page cache for reads, avoiding fs");
142 static u_long vn_io_faults_cnt;
143 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
144     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
145 
146 static int vfs_allow_read_dir = 0;
147 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
148     &vfs_allow_read_dir, 0,
149     "Enable read(2) of directory by root for filesystems that support it");
150 
151 /*
152  * Returns true if vn_io_fault mode of handling the i/o request should
153  * be used.
154  */
155 static bool
156 do_vn_io_fault(struct vnode *vp, struct uio *uio)
157 {
158 	struct mount *mp;
159 
160 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
161 	    (mp = vp->v_mount) != NULL &&
162 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
163 }
164 
165 /*
166  * Structure used to pass arguments to vn_io_fault1(), to do either
167  * file- or vnode-based I/O calls.
168  */
169 struct vn_io_fault_args {
170 	enum {
171 		VN_IO_FAULT_FOP,
172 		VN_IO_FAULT_VOP
173 	} kind;
174 	struct ucred *cred;
175 	int flags;
176 	union {
177 		struct fop_args_tag {
178 			struct file *fp;
179 			fo_rdwr_t *doio;
180 		} fop_args;
181 		struct vop_args_tag {
182 			struct vnode *vp;
183 		} vop_args;
184 	} args;
185 };
186 
187 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
188     struct vn_io_fault_args *args, struct thread *td);
189 
190 int
191 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
192 {
193 	struct thread *td = ndp->ni_cnd.cn_thread;
194 
195 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
196 }
197 
198 static uint64_t
199 open2nameif(int fmode, u_int vn_open_flags)
200 {
201 	uint64_t res;
202 
203 	res = ISOPEN | LOCKLEAF;
204 	if ((fmode & O_RESOLVE_BENEATH) != 0)
205 		res |= RBENEATH;
206 	if ((fmode & O_EMPTY_PATH) != 0)
207 		res |= EMPTYPATH;
208 	if ((fmode & FREAD) != 0)
209 		res |= OPENREAD;
210 	if ((fmode & FWRITE) != 0)
211 		res |= OPENWRITE;
212 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
213 		res |= AUDITVNODE1;
214 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
215 		res |= NOCAPCHECK;
216 	return (res);
217 }
218 
219 /*
220  * Common code for vnode open operations via a name lookup.
221  * Lookup the vnode and invoke VOP_CREATE if needed.
222  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
223  *
224  * Note that this does NOT free nameidata for the successful case,
225  * due to the NDINIT being done elsewhere.
226  */
227 int
228 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
229     struct ucred *cred, struct file *fp)
230 {
231 	struct vnode *vp;
232 	struct mount *mp;
233 	struct thread *td = ndp->ni_cnd.cn_thread;
234 	struct vattr vat;
235 	struct vattr *vap = &vat;
236 	int fmode, error;
237 	bool first_open;
238 
239 restart:
240 	first_open = false;
241 	fmode = *flagp;
242 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
243 	    O_EXCL | O_DIRECTORY) ||
244 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
245 		return (EINVAL);
246 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
247 		ndp->ni_cnd.cn_nameiop = CREATE;
248 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
249 		/*
250 		 * Set NOCACHE to avoid flushing the cache when
251 		 * rolling in many files at once.
252 		 *
253 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
254 		 * exist despite NOCACHE.
255 		 */
256 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
257 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
258 			ndp->ni_cnd.cn_flags |= FOLLOW;
259 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
260 			bwillwrite();
261 		if ((error = namei(ndp)) != 0)
262 			return (error);
263 		if (ndp->ni_vp == NULL) {
264 			VATTR_NULL(vap);
265 			vap->va_type = VREG;
266 			vap->va_mode = cmode;
267 			if (fmode & O_EXCL)
268 				vap->va_vaflags |= VA_EXCLUSIVE;
269 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
270 				NDFREE(ndp, NDF_ONLY_PNBUF);
271 				vput(ndp->ni_dvp);
272 				if ((error = vn_start_write(NULL, &mp,
273 				    V_XSLEEP | PCATCH)) != 0)
274 					return (error);
275 				NDREINIT(ndp);
276 				goto restart;
277 			}
278 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
279 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
280 #ifdef MAC
281 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
282 			    &ndp->ni_cnd, vap);
283 			if (error == 0)
284 #endif
285 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
286 				    &ndp->ni_cnd, vap);
287 			vp = ndp->ni_vp;
288 			if (error == 0 && (fmode & O_EXCL) != 0 &&
289 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
290 				VI_LOCK(vp);
291 				vp->v_iflag |= VI_FOPENING;
292 				VI_UNLOCK(vp);
293 				first_open = true;
294 			}
295 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
296 			    false);
297 			vn_finished_write(mp);
298 			if (error) {
299 				NDFREE(ndp, NDF_ONLY_PNBUF);
300 				if (error == ERELOOKUP) {
301 					NDREINIT(ndp);
302 					goto restart;
303 				}
304 				return (error);
305 			}
306 			fmode &= ~O_TRUNC;
307 		} else {
308 			if (ndp->ni_dvp == ndp->ni_vp)
309 				vrele(ndp->ni_dvp);
310 			else
311 				vput(ndp->ni_dvp);
312 			ndp->ni_dvp = NULL;
313 			vp = ndp->ni_vp;
314 			if (fmode & O_EXCL) {
315 				error = EEXIST;
316 				goto bad;
317 			}
318 			if (vp->v_type == VDIR) {
319 				error = EISDIR;
320 				goto bad;
321 			}
322 			fmode &= ~O_CREAT;
323 		}
324 	} else {
325 		ndp->ni_cnd.cn_nameiop = LOOKUP;
326 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
327 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
328 		    FOLLOW;
329 		if ((fmode & FWRITE) == 0)
330 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
331 		if ((error = namei(ndp)) != 0)
332 			return (error);
333 		vp = ndp->ni_vp;
334 	}
335 	error = vn_open_vnode(vp, fmode, cred, td, fp);
336 	if (first_open) {
337 		VI_LOCK(vp);
338 		vp->v_iflag &= ~VI_FOPENING;
339 		wakeup(vp);
340 		VI_UNLOCK(vp);
341 	}
342 	if (error)
343 		goto bad;
344 	*flagp = fmode;
345 	return (0);
346 bad:
347 	NDFREE(ndp, NDF_ONLY_PNBUF);
348 	vput(vp);
349 	*flagp = fmode;
350 	ndp->ni_vp = NULL;
351 	return (error);
352 }
353 
354 static int
355 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
356 {
357 	struct flock lf;
358 	int error, lock_flags, type;
359 
360 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
361 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
362 		return (0);
363 	KASSERT(fp != NULL, ("open with flock requires fp"));
364 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
365 		return (EOPNOTSUPP);
366 
367 	lock_flags = VOP_ISLOCKED(vp);
368 	VOP_UNLOCK(vp);
369 
370 	lf.l_whence = SEEK_SET;
371 	lf.l_start = 0;
372 	lf.l_len = 0;
373 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
374 	type = F_FLOCK;
375 	if ((fmode & FNONBLOCK) == 0)
376 		type |= F_WAIT;
377 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
378 		type |= F_FIRSTOPEN;
379 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
380 	if (error == 0)
381 		fp->f_flag |= FHASLOCK;
382 
383 	vn_lock(vp, lock_flags | LK_RETRY);
384 	return (error);
385 }
386 
387 /*
388  * Common code for vnode open operations once a vnode is located.
389  * Check permissions, and call the VOP_OPEN routine.
390  */
391 int
392 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
393     struct thread *td, struct file *fp)
394 {
395 	accmode_t accmode;
396 	int error;
397 
398 	if (vp->v_type == VLNK) {
399 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
400 			return (EMLINK);
401 	}
402 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
403 		return (ENOTDIR);
404 
405 	accmode = 0;
406 	if ((fmode & O_PATH) == 0) {
407 		if (vp->v_type == VSOCK)
408 			return (EOPNOTSUPP);
409 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
410 			if (vp->v_type == VDIR)
411 				return (EISDIR);
412 			accmode |= VWRITE;
413 		}
414 		if ((fmode & FREAD) != 0)
415 			accmode |= VREAD;
416 		if ((fmode & O_APPEND) && (fmode & FWRITE))
417 			accmode |= VAPPEND;
418 #ifdef MAC
419 		if ((fmode & O_CREAT) != 0)
420 			accmode |= VCREAT;
421 #endif
422 	}
423 	if ((fmode & FEXEC) != 0)
424 		accmode |= VEXEC;
425 #ifdef MAC
426 	if ((fmode & O_VERIFY) != 0)
427 		accmode |= VVERIFY;
428 	error = mac_vnode_check_open(cred, vp, accmode);
429 	if (error != 0)
430 		return (error);
431 
432 	accmode &= ~(VCREAT | VVERIFY);
433 #endif
434 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
435 		error = VOP_ACCESS(vp, accmode, cred, td);
436 		if (error != 0)
437 			return (error);
438 	}
439 	if ((fmode & O_PATH) != 0) {
440 		if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
441 		    VOP_ACCESS(vp, VREAD, cred, td) == 0)
442 			fp->f_flag |= FKQALLOWED;
443 		return (0);
444 	}
445 
446 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
447 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
448 	error = VOP_OPEN(vp, fmode, cred, td, fp);
449 	if (error != 0)
450 		return (error);
451 
452 	error = vn_open_vnode_advlock(vp, fmode, fp);
453 	if (error == 0 && (fmode & FWRITE) != 0) {
454 		error = VOP_ADD_WRITECOUNT(vp, 1);
455 		if (error == 0) {
456 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
457 			     __func__, vp, vp->v_writecount);
458 		}
459 	}
460 
461 	/*
462 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
463 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
464 	 */
465 	if (error != 0) {
466 		if (fp != NULL) {
467 			/*
468 			 * Arrange the call by having fdrop() to use
469 			 * vn_closefile().  This is to satisfy
470 			 * filesystems like devfs or tmpfs, which
471 			 * override fo_close().
472 			 */
473 			fp->f_flag |= FOPENFAILED;
474 			fp->f_vnode = vp;
475 			if (fp->f_ops == &badfileops) {
476 				fp->f_type = DTYPE_VNODE;
477 				fp->f_ops = &vnops;
478 			}
479 			vref(vp);
480 		} else {
481 			/*
482 			 * If there is no fp, due to kernel-mode open,
483 			 * we can call VOP_CLOSE() now.
484 			 */
485 			if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
486 			    !MNT_EXTENDED_SHARED(vp->v_mount) &&
487 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
488 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
489 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
490 			    cred, td);
491 		}
492 	}
493 
494 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
495 	return (error);
496 
497 }
498 
499 /*
500  * Check for write permissions on the specified vnode.
501  * Prototype text segments cannot be written.
502  * It is racy.
503  */
504 int
505 vn_writechk(struct vnode *vp)
506 {
507 
508 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
509 	/*
510 	 * If there's shared text associated with
511 	 * the vnode, try to free it up once.  If
512 	 * we fail, we can't allow writing.
513 	 */
514 	if (VOP_IS_TEXT(vp))
515 		return (ETXTBSY);
516 
517 	return (0);
518 }
519 
520 /*
521  * Vnode close call
522  */
523 static int
524 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
525     struct thread *td, bool keep_ref)
526 {
527 	struct mount *mp;
528 	int error, lock_flags;
529 
530 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
531 	    MNT_EXTENDED_SHARED(vp->v_mount))
532 		lock_flags = LK_SHARED;
533 	else
534 		lock_flags = LK_EXCLUSIVE;
535 
536 	vn_start_write(vp, &mp, V_WAIT);
537 	vn_lock(vp, lock_flags | LK_RETRY);
538 	AUDIT_ARG_VNODE1(vp);
539 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
540 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
541 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
542 		    __func__, vp, vp->v_writecount);
543 	}
544 	error = VOP_CLOSE(vp, flags, file_cred, td);
545 	if (keep_ref)
546 		VOP_UNLOCK(vp);
547 	else
548 		vput(vp);
549 	vn_finished_write(mp);
550 	return (error);
551 }
552 
553 int
554 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
555     struct thread *td)
556 {
557 
558 	return (vn_close1(vp, flags, file_cred, td, false));
559 }
560 
561 /*
562  * Heuristic to detect sequential operation.
563  */
564 static int
565 sequential_heuristic(struct uio *uio, struct file *fp)
566 {
567 	enum uio_rw rw;
568 
569 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
570 
571 	rw = uio->uio_rw;
572 	if (fp->f_flag & FRDAHEAD)
573 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
574 
575 	/*
576 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
577 	 * that the first I/O is normally considered to be slightly
578 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
579 	 * unless previous seeks have reduced f_seqcount to 0, in which
580 	 * case offset 0 is not special.
581 	 */
582 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
583 	    uio->uio_offset == fp->f_nextoff[rw]) {
584 		/*
585 		 * f_seqcount is in units of fixed-size blocks so that it
586 		 * depends mainly on the amount of sequential I/O and not
587 		 * much on the number of sequential I/O's.  The fixed size
588 		 * of 16384 is hard-coded here since it is (not quite) just
589 		 * a magic size that works well here.  This size is more
590 		 * closely related to the best I/O size for real disks than
591 		 * to any block size used by software.
592 		 */
593 		if (uio->uio_resid >= IO_SEQMAX * 16384)
594 			fp->f_seqcount[rw] = IO_SEQMAX;
595 		else {
596 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
597 			if (fp->f_seqcount[rw] > IO_SEQMAX)
598 				fp->f_seqcount[rw] = IO_SEQMAX;
599 		}
600 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
601 	}
602 
603 	/* Not sequential.  Quickly draw-down sequentiality. */
604 	if (fp->f_seqcount[rw] > 1)
605 		fp->f_seqcount[rw] = 1;
606 	else
607 		fp->f_seqcount[rw] = 0;
608 	return (0);
609 }
610 
611 /*
612  * Package up an I/O request on a vnode into a uio and do it.
613  */
614 int
615 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
616     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
617     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
618 {
619 	struct uio auio;
620 	struct iovec aiov;
621 	struct mount *mp;
622 	struct ucred *cred;
623 	void *rl_cookie;
624 	struct vn_io_fault_args args;
625 	int error, lock_flags;
626 
627 	if (offset < 0 && vp->v_type != VCHR)
628 		return (EINVAL);
629 	auio.uio_iov = &aiov;
630 	auio.uio_iovcnt = 1;
631 	aiov.iov_base = base;
632 	aiov.iov_len = len;
633 	auio.uio_resid = len;
634 	auio.uio_offset = offset;
635 	auio.uio_segflg = segflg;
636 	auio.uio_rw = rw;
637 	auio.uio_td = td;
638 	error = 0;
639 
640 	if ((ioflg & IO_NODELOCKED) == 0) {
641 		if ((ioflg & IO_RANGELOCKED) == 0) {
642 			if (rw == UIO_READ) {
643 				rl_cookie = vn_rangelock_rlock(vp, offset,
644 				    offset + len);
645 			} else if ((ioflg & IO_APPEND) != 0) {
646 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
647 			} else {
648 				rl_cookie = vn_rangelock_wlock(vp, offset,
649 				    offset + len);
650 			}
651 		} else
652 			rl_cookie = NULL;
653 		mp = NULL;
654 		if (rw == UIO_WRITE) {
655 			if (vp->v_type != VCHR &&
656 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
657 			    != 0)
658 				goto out;
659 			lock_flags = vn_lktype_write(mp, vp);
660 		} else
661 			lock_flags = LK_SHARED;
662 		vn_lock(vp, lock_flags | LK_RETRY);
663 	} else
664 		rl_cookie = NULL;
665 
666 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
667 #ifdef MAC
668 	if ((ioflg & IO_NOMACCHECK) == 0) {
669 		if (rw == UIO_READ)
670 			error = mac_vnode_check_read(active_cred, file_cred,
671 			    vp);
672 		else
673 			error = mac_vnode_check_write(active_cred, file_cred,
674 			    vp);
675 	}
676 #endif
677 	if (error == 0) {
678 		if (file_cred != NULL)
679 			cred = file_cred;
680 		else
681 			cred = active_cred;
682 		if (do_vn_io_fault(vp, &auio)) {
683 			args.kind = VN_IO_FAULT_VOP;
684 			args.cred = cred;
685 			args.flags = ioflg;
686 			args.args.vop_args.vp = vp;
687 			error = vn_io_fault1(vp, &auio, &args, td);
688 		} else if (rw == UIO_READ) {
689 			error = VOP_READ(vp, &auio, ioflg, cred);
690 		} else /* if (rw == UIO_WRITE) */ {
691 			error = VOP_WRITE(vp, &auio, ioflg, cred);
692 		}
693 	}
694 	if (aresid)
695 		*aresid = auio.uio_resid;
696 	else
697 		if (auio.uio_resid && error == 0)
698 			error = EIO;
699 	if ((ioflg & IO_NODELOCKED) == 0) {
700 		VOP_UNLOCK(vp);
701 		if (mp != NULL)
702 			vn_finished_write(mp);
703 	}
704  out:
705 	if (rl_cookie != NULL)
706 		vn_rangelock_unlock(vp, rl_cookie);
707 	return (error);
708 }
709 
710 /*
711  * Package up an I/O request on a vnode into a uio and do it.  The I/O
712  * request is split up into smaller chunks and we try to avoid saturating
713  * the buffer cache while potentially holding a vnode locked, so we
714  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
715  * to give other processes a chance to lock the vnode (either other processes
716  * core'ing the same binary, or unrelated processes scanning the directory).
717  */
718 int
719 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
720     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
721     struct ucred *file_cred, size_t *aresid, struct thread *td)
722 {
723 	int error = 0;
724 	ssize_t iaresid;
725 
726 	do {
727 		int chunk;
728 
729 		/*
730 		 * Force `offset' to a multiple of MAXBSIZE except possibly
731 		 * for the first chunk, so that filesystems only need to
732 		 * write full blocks except possibly for the first and last
733 		 * chunks.
734 		 */
735 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
736 
737 		if (chunk > len)
738 			chunk = len;
739 		if (rw != UIO_READ && vp->v_type == VREG)
740 			bwillwrite();
741 		iaresid = 0;
742 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
743 		    ioflg, active_cred, file_cred, &iaresid, td);
744 		len -= chunk;	/* aresid calc already includes length */
745 		if (error)
746 			break;
747 		offset += chunk;
748 		base = (char *)base + chunk;
749 		kern_yield(PRI_USER);
750 	} while (len);
751 	if (aresid)
752 		*aresid = len + iaresid;
753 	return (error);
754 }
755 
756 #if OFF_MAX <= LONG_MAX
757 off_t
758 foffset_lock(struct file *fp, int flags)
759 {
760 	volatile short *flagsp;
761 	off_t res;
762 	short state;
763 
764 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
765 
766 	if ((flags & FOF_NOLOCK) != 0)
767 		return (atomic_load_long(&fp->f_offset));
768 
769 	/*
770 	 * According to McKusick the vn lock was protecting f_offset here.
771 	 * It is now protected by the FOFFSET_LOCKED flag.
772 	 */
773 	flagsp = &fp->f_vnread_flags;
774 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
775 		return (atomic_load_long(&fp->f_offset));
776 
777 	sleepq_lock(&fp->f_vnread_flags);
778 	state = atomic_load_16(flagsp);
779 	for (;;) {
780 		if ((state & FOFFSET_LOCKED) == 0) {
781 			if (!atomic_fcmpset_acq_16(flagsp, &state,
782 			    FOFFSET_LOCKED))
783 				continue;
784 			break;
785 		}
786 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
787 			if (!atomic_fcmpset_acq_16(flagsp, &state,
788 			    state | FOFFSET_LOCK_WAITING))
789 				continue;
790 		}
791 		DROP_GIANT();
792 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
793 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
794 		PICKUP_GIANT();
795 		sleepq_lock(&fp->f_vnread_flags);
796 		state = atomic_load_16(flagsp);
797 	}
798 	res = atomic_load_long(&fp->f_offset);
799 	sleepq_release(&fp->f_vnread_flags);
800 	return (res);
801 }
802 
803 void
804 foffset_unlock(struct file *fp, off_t val, int flags)
805 {
806 	volatile short *flagsp;
807 	short state;
808 
809 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
810 
811 	if ((flags & FOF_NOUPDATE) == 0)
812 		atomic_store_long(&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 
818 	if ((flags & FOF_NOLOCK) != 0)
819 		return;
820 
821 	flagsp = &fp->f_vnread_flags;
822 	state = atomic_load_16(flagsp);
823 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
824 	    atomic_cmpset_rel_16(flagsp, state, 0))
825 		return;
826 
827 	sleepq_lock(&fp->f_vnread_flags);
828 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
829 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
830 	fp->f_vnread_flags = 0;
831 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
832 	sleepq_release(&fp->f_vnread_flags);
833 }
834 #else
835 off_t
836 foffset_lock(struct file *fp, int flags)
837 {
838 	struct mtx *mtxp;
839 	off_t res;
840 
841 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
842 
843 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
844 	mtx_lock(mtxp);
845 	if ((flags & FOF_NOLOCK) == 0) {
846 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
847 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
848 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
849 			    "vofflock", 0);
850 		}
851 		fp->f_vnread_flags |= FOFFSET_LOCKED;
852 	}
853 	res = fp->f_offset;
854 	mtx_unlock(mtxp);
855 	return (res);
856 }
857 
858 void
859 foffset_unlock(struct file *fp, off_t val, int flags)
860 {
861 	struct mtx *mtxp;
862 
863 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
864 
865 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
866 	mtx_lock(mtxp);
867 	if ((flags & FOF_NOUPDATE) == 0)
868 		fp->f_offset = val;
869 	if ((flags & FOF_NEXTOFF_R) != 0)
870 		fp->f_nextoff[UIO_READ] = val;
871 	if ((flags & FOF_NEXTOFF_W) != 0)
872 		fp->f_nextoff[UIO_WRITE] = val;
873 	if ((flags & FOF_NOLOCK) == 0) {
874 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
875 		    ("Lost FOFFSET_LOCKED"));
876 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
877 			wakeup(&fp->f_vnread_flags);
878 		fp->f_vnread_flags = 0;
879 	}
880 	mtx_unlock(mtxp);
881 }
882 #endif
883 
884 void
885 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
886 {
887 
888 	if ((flags & FOF_OFFSET) == 0)
889 		uio->uio_offset = foffset_lock(fp, flags);
890 }
891 
892 void
893 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
894 {
895 
896 	if ((flags & FOF_OFFSET) == 0)
897 		foffset_unlock(fp, uio->uio_offset, flags);
898 }
899 
900 static int
901 get_advice(struct file *fp, struct uio *uio)
902 {
903 	struct mtx *mtxp;
904 	int ret;
905 
906 	ret = POSIX_FADV_NORMAL;
907 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
908 		return (ret);
909 
910 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
911 	mtx_lock(mtxp);
912 	if (fp->f_advice != NULL &&
913 	    uio->uio_offset >= fp->f_advice->fa_start &&
914 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
915 		ret = fp->f_advice->fa_advice;
916 	mtx_unlock(mtxp);
917 	return (ret);
918 }
919 
920 static int
921 get_write_ioflag(struct file *fp)
922 {
923 	int ioflag;
924 	struct mount *mp;
925 	struct vnode *vp;
926 
927 	ioflag = 0;
928 	vp = fp->f_vnode;
929 	mp = atomic_load_ptr(&vp->v_mount);
930 
931 	if ((fp->f_flag & O_DIRECT) != 0)
932 		ioflag |= IO_DIRECT;
933 
934 	if ((fp->f_flag & O_FSYNC) != 0 ||
935 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
936 		ioflag |= IO_SYNC;
937 
938 	/*
939 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
940 	 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
941 	 * fall back to full O_SYNC behavior.
942 	 */
943 	if ((fp->f_flag & O_DSYNC) != 0)
944 		ioflag |= IO_SYNC | IO_DATASYNC;
945 
946 	return (ioflag);
947 }
948 
949 int
950 vn_read_from_obj(struct vnode *vp, struct uio *uio)
951 {
952 	vm_object_t obj;
953 	vm_page_t ma[io_hold_cnt + 2];
954 	off_t off, vsz;
955 	ssize_t resid;
956 	int error, i, j;
957 
958 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
959 	obj = atomic_load_ptr(&vp->v_object);
960 	if (obj == NULL)
961 		return (EJUSTRETURN);
962 
963 	/*
964 	 * Depends on type stability of vm_objects.
965 	 */
966 	vm_object_pip_add(obj, 1);
967 	if ((obj->flags & OBJ_DEAD) != 0) {
968 		/*
969 		 * Note that object might be already reused from the
970 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
971 		 * we recheck for DOOMED vnode state after all pages
972 		 * are busied, and retract then.
973 		 *
974 		 * But we check for OBJ_DEAD to ensure that we do not
975 		 * busy pages while vm_object_terminate_pages()
976 		 * processes the queue.
977 		 */
978 		error = EJUSTRETURN;
979 		goto out_pip;
980 	}
981 
982 	resid = uio->uio_resid;
983 	off = uio->uio_offset;
984 	for (i = 0; resid > 0; i++) {
985 		MPASS(i < io_hold_cnt + 2);
986 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
987 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
988 		    VM_ALLOC_NOWAIT);
989 		if (ma[i] == NULL)
990 			break;
991 
992 		/*
993 		 * Skip invalid pages.  Valid mask can be partial only
994 		 * at EOF, and we clip later.
995 		 */
996 		if (vm_page_none_valid(ma[i])) {
997 			vm_page_sunbusy(ma[i]);
998 			break;
999 		}
1000 
1001 		resid -= PAGE_SIZE;
1002 		off += PAGE_SIZE;
1003 	}
1004 	if (i == 0) {
1005 		error = EJUSTRETURN;
1006 		goto out_pip;
1007 	}
1008 
1009 	/*
1010 	 * Check VIRF_DOOMED after we busied our pages.  Since
1011 	 * vgonel() terminates the vnode' vm_object, it cannot
1012 	 * process past pages busied by us.
1013 	 */
1014 	if (VN_IS_DOOMED(vp)) {
1015 		error = EJUSTRETURN;
1016 		goto out;
1017 	}
1018 
1019 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1020 	if (resid > uio->uio_resid)
1021 		resid = uio->uio_resid;
1022 
1023 	/*
1024 	 * Unlocked read of vnp_size is safe because truncation cannot
1025 	 * pass busied page.  But we load vnp_size into a local
1026 	 * variable so that possible concurrent extension does not
1027 	 * break calculation.
1028 	 */
1029 #if defined(__powerpc__) && !defined(__powerpc64__)
1030 	vsz = obj->un_pager.vnp.vnp_size;
1031 #else
1032 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1033 #endif
1034 	if (uio->uio_offset >= vsz) {
1035 		error = EJUSTRETURN;
1036 		goto out;
1037 	}
1038 	if (uio->uio_offset + resid > vsz)
1039 		resid = vsz - uio->uio_offset;
1040 
1041 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1042 
1043 out:
1044 	for (j = 0; j < i; j++) {
1045 		if (error == 0)
1046 			vm_page_reference(ma[j]);
1047 		vm_page_sunbusy(ma[j]);
1048 	}
1049 out_pip:
1050 	vm_object_pip_wakeup(obj);
1051 	if (error != 0)
1052 		return (error);
1053 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1054 }
1055 
1056 /*
1057  * File table vnode read routine.
1058  */
1059 static int
1060 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1061     struct thread *td)
1062 {
1063 	struct vnode *vp;
1064 	off_t orig_offset;
1065 	int error, ioflag;
1066 	int advice;
1067 
1068 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1069 	    uio->uio_td, td));
1070 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1071 	vp = fp->f_vnode;
1072 	ioflag = 0;
1073 	if (fp->f_flag & FNONBLOCK)
1074 		ioflag |= IO_NDELAY;
1075 	if (fp->f_flag & O_DIRECT)
1076 		ioflag |= IO_DIRECT;
1077 
1078 	/*
1079 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1080 	 * allows us to avoid unneeded work outright.
1081 	 */
1082 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1083 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1084 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1085 		if (error == 0) {
1086 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1087 			return (0);
1088 		}
1089 		if (error != EJUSTRETURN)
1090 			return (error);
1091 	}
1092 
1093 	advice = get_advice(fp, uio);
1094 	vn_lock(vp, LK_SHARED | LK_RETRY);
1095 
1096 	switch (advice) {
1097 	case POSIX_FADV_NORMAL:
1098 	case POSIX_FADV_SEQUENTIAL:
1099 	case POSIX_FADV_NOREUSE:
1100 		ioflag |= sequential_heuristic(uio, fp);
1101 		break;
1102 	case POSIX_FADV_RANDOM:
1103 		/* Disable read-ahead for random I/O. */
1104 		break;
1105 	}
1106 	orig_offset = uio->uio_offset;
1107 
1108 #ifdef MAC
1109 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1110 	if (error == 0)
1111 #endif
1112 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1113 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1114 	VOP_UNLOCK(vp);
1115 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1116 	    orig_offset != uio->uio_offset)
1117 		/*
1118 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1119 		 * for the backing file after a POSIX_FADV_NOREUSE
1120 		 * read(2).
1121 		 */
1122 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1123 		    POSIX_FADV_DONTNEED);
1124 	return (error);
1125 }
1126 
1127 /*
1128  * File table vnode write routine.
1129  */
1130 static int
1131 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1132     struct thread *td)
1133 {
1134 	struct vnode *vp;
1135 	struct mount *mp;
1136 	off_t orig_offset;
1137 	int error, ioflag;
1138 	int advice;
1139 	bool need_finished_write;
1140 
1141 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1142 	    uio->uio_td, td));
1143 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1144 	vp = fp->f_vnode;
1145 	if (vp->v_type == VREG)
1146 		bwillwrite();
1147 	ioflag = IO_UNIT;
1148 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1149 		ioflag |= IO_APPEND;
1150 	if ((fp->f_flag & FNONBLOCK) != 0)
1151 		ioflag |= IO_NDELAY;
1152 	ioflag |= get_write_ioflag(fp);
1153 
1154 	mp = NULL;
1155 	need_finished_write = false;
1156 	if (vp->v_type != VCHR) {
1157 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1158 		if (error != 0)
1159 			goto unlock;
1160 		need_finished_write = true;
1161 	}
1162 
1163 	advice = get_advice(fp, uio);
1164 
1165 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1166 	switch (advice) {
1167 	case POSIX_FADV_NORMAL:
1168 	case POSIX_FADV_SEQUENTIAL:
1169 	case POSIX_FADV_NOREUSE:
1170 		ioflag |= sequential_heuristic(uio, fp);
1171 		break;
1172 	case POSIX_FADV_RANDOM:
1173 		/* XXX: Is this correct? */
1174 		break;
1175 	}
1176 	orig_offset = uio->uio_offset;
1177 
1178 #ifdef MAC
1179 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1180 	if (error == 0)
1181 #endif
1182 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1183 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1184 	VOP_UNLOCK(vp);
1185 	if (need_finished_write)
1186 		vn_finished_write(mp);
1187 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1188 	    orig_offset != uio->uio_offset)
1189 		/*
1190 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1191 		 * for the backing file after a POSIX_FADV_NOREUSE
1192 		 * write(2).
1193 		 */
1194 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1195 		    POSIX_FADV_DONTNEED);
1196 unlock:
1197 	return (error);
1198 }
1199 
1200 /*
1201  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1202  * prevent the following deadlock:
1203  *
1204  * Assume that the thread A reads from the vnode vp1 into userspace
1205  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1206  * currently not resident, then system ends up with the call chain
1207  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1208  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1209  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1210  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1211  * backed by the pages of vnode vp1, and some page in buf2 is not
1212  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1213  *
1214  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1215  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1216  * Instead, it first tries to do the whole range i/o with pagefaults
1217  * disabled. If all pages in the i/o buffer are resident and mapped,
1218  * VOP will succeed (ignoring the genuine filesystem errors).
1219  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1220  * i/o in chunks, with all pages in the chunk prefaulted and held
1221  * using vm_fault_quick_hold_pages().
1222  *
1223  * Filesystems using this deadlock avoidance scheme should use the
1224  * array of the held pages from uio, saved in the curthread->td_ma,
1225  * instead of doing uiomove().  A helper function
1226  * vn_io_fault_uiomove() converts uiomove request into
1227  * uiomove_fromphys() over td_ma array.
1228  *
1229  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1230  * make the current i/o request atomic with respect to other i/os and
1231  * truncations.
1232  */
1233 
1234 /*
1235  * Decode vn_io_fault_args and perform the corresponding i/o.
1236  */
1237 static int
1238 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1239     struct thread *td)
1240 {
1241 	int error, save;
1242 
1243 	error = 0;
1244 	save = vm_fault_disable_pagefaults();
1245 	switch (args->kind) {
1246 	case VN_IO_FAULT_FOP:
1247 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1248 		    uio, args->cred, args->flags, td);
1249 		break;
1250 	case VN_IO_FAULT_VOP:
1251 		if (uio->uio_rw == UIO_READ) {
1252 			error = VOP_READ(args->args.vop_args.vp, uio,
1253 			    args->flags, args->cred);
1254 		} else if (uio->uio_rw == UIO_WRITE) {
1255 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1256 			    args->flags, args->cred);
1257 		}
1258 		break;
1259 	default:
1260 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1261 		    args->kind, uio->uio_rw);
1262 	}
1263 	vm_fault_enable_pagefaults(save);
1264 	return (error);
1265 }
1266 
1267 static int
1268 vn_io_fault_touch(char *base, const struct uio *uio)
1269 {
1270 	int r;
1271 
1272 	r = fubyte(base);
1273 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1274 		return (EFAULT);
1275 	return (0);
1276 }
1277 
1278 static int
1279 vn_io_fault_prefault_user(const struct uio *uio)
1280 {
1281 	char *base;
1282 	const struct iovec *iov;
1283 	size_t len;
1284 	ssize_t resid;
1285 	int error, i;
1286 
1287 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1288 	    ("vn_io_fault_prefault userspace"));
1289 
1290 	error = i = 0;
1291 	iov = uio->uio_iov;
1292 	resid = uio->uio_resid;
1293 	base = iov->iov_base;
1294 	len = iov->iov_len;
1295 	while (resid > 0) {
1296 		error = vn_io_fault_touch(base, uio);
1297 		if (error != 0)
1298 			break;
1299 		if (len < PAGE_SIZE) {
1300 			if (len != 0) {
1301 				error = vn_io_fault_touch(base + len - 1, uio);
1302 				if (error != 0)
1303 					break;
1304 				resid -= len;
1305 			}
1306 			if (++i >= uio->uio_iovcnt)
1307 				break;
1308 			iov = uio->uio_iov + i;
1309 			base = iov->iov_base;
1310 			len = iov->iov_len;
1311 		} else {
1312 			len -= PAGE_SIZE;
1313 			base += PAGE_SIZE;
1314 			resid -= PAGE_SIZE;
1315 		}
1316 	}
1317 	return (error);
1318 }
1319 
1320 /*
1321  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1322  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1323  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1324  * into args and call vn_io_fault1() to handle faults during the user
1325  * mode buffer accesses.
1326  */
1327 static int
1328 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1329     struct thread *td)
1330 {
1331 	vm_page_t ma[io_hold_cnt + 2];
1332 	struct uio *uio_clone, short_uio;
1333 	struct iovec short_iovec[1];
1334 	vm_page_t *prev_td_ma;
1335 	vm_prot_t prot;
1336 	vm_offset_t addr, end;
1337 	size_t len, resid;
1338 	ssize_t adv;
1339 	int error, cnt, saveheld, prev_td_ma_cnt;
1340 
1341 	if (vn_io_fault_prefault) {
1342 		error = vn_io_fault_prefault_user(uio);
1343 		if (error != 0)
1344 			return (error); /* Or ignore ? */
1345 	}
1346 
1347 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1348 
1349 	/*
1350 	 * The UFS follows IO_UNIT directive and replays back both
1351 	 * uio_offset and uio_resid if an error is encountered during the
1352 	 * operation.  But, since the iovec may be already advanced,
1353 	 * uio is still in an inconsistent state.
1354 	 *
1355 	 * Cache a copy of the original uio, which is advanced to the redo
1356 	 * point using UIO_NOCOPY below.
1357 	 */
1358 	uio_clone = cloneuio(uio);
1359 	resid = uio->uio_resid;
1360 
1361 	short_uio.uio_segflg = UIO_USERSPACE;
1362 	short_uio.uio_rw = uio->uio_rw;
1363 	short_uio.uio_td = uio->uio_td;
1364 
1365 	error = vn_io_fault_doio(args, uio, td);
1366 	if (error != EFAULT)
1367 		goto out;
1368 
1369 	atomic_add_long(&vn_io_faults_cnt, 1);
1370 	uio_clone->uio_segflg = UIO_NOCOPY;
1371 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1372 	uio_clone->uio_segflg = uio->uio_segflg;
1373 
1374 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1375 	prev_td_ma = td->td_ma;
1376 	prev_td_ma_cnt = td->td_ma_cnt;
1377 
1378 	while (uio_clone->uio_resid != 0) {
1379 		len = uio_clone->uio_iov->iov_len;
1380 		if (len == 0) {
1381 			KASSERT(uio_clone->uio_iovcnt >= 1,
1382 			    ("iovcnt underflow"));
1383 			uio_clone->uio_iov++;
1384 			uio_clone->uio_iovcnt--;
1385 			continue;
1386 		}
1387 		if (len > ptoa(io_hold_cnt))
1388 			len = ptoa(io_hold_cnt);
1389 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1390 		end = round_page(addr + len);
1391 		if (end < addr) {
1392 			error = EFAULT;
1393 			break;
1394 		}
1395 		cnt = atop(end - trunc_page(addr));
1396 		/*
1397 		 * A perfectly misaligned address and length could cause
1398 		 * both the start and the end of the chunk to use partial
1399 		 * page.  +2 accounts for such a situation.
1400 		 */
1401 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1402 		    addr, len, prot, ma, io_hold_cnt + 2);
1403 		if (cnt == -1) {
1404 			error = EFAULT;
1405 			break;
1406 		}
1407 		short_uio.uio_iov = &short_iovec[0];
1408 		short_iovec[0].iov_base = (void *)addr;
1409 		short_uio.uio_iovcnt = 1;
1410 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1411 		short_uio.uio_offset = uio_clone->uio_offset;
1412 		td->td_ma = ma;
1413 		td->td_ma_cnt = cnt;
1414 
1415 		error = vn_io_fault_doio(args, &short_uio, td);
1416 		vm_page_unhold_pages(ma, cnt);
1417 		adv = len - short_uio.uio_resid;
1418 
1419 		uio_clone->uio_iov->iov_base =
1420 		    (char *)uio_clone->uio_iov->iov_base + adv;
1421 		uio_clone->uio_iov->iov_len -= adv;
1422 		uio_clone->uio_resid -= adv;
1423 		uio_clone->uio_offset += adv;
1424 
1425 		uio->uio_resid -= adv;
1426 		uio->uio_offset += adv;
1427 
1428 		if (error != 0 || adv == 0)
1429 			break;
1430 	}
1431 	td->td_ma = prev_td_ma;
1432 	td->td_ma_cnt = prev_td_ma_cnt;
1433 	curthread_pflags_restore(saveheld);
1434 out:
1435 	free(uio_clone, M_IOV);
1436 	return (error);
1437 }
1438 
1439 static int
1440 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1441     int flags, struct thread *td)
1442 {
1443 	fo_rdwr_t *doio;
1444 	struct vnode *vp;
1445 	void *rl_cookie;
1446 	struct vn_io_fault_args args;
1447 	int error;
1448 
1449 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1450 	vp = fp->f_vnode;
1451 
1452 	/*
1453 	 * The ability to read(2) on a directory has historically been
1454 	 * allowed for all users, but this can and has been the source of
1455 	 * at least one security issue in the past.  As such, it is now hidden
1456 	 * away behind a sysctl for those that actually need it to use it, and
1457 	 * restricted to root when it's turned on to make it relatively safe to
1458 	 * leave on for longer sessions of need.
1459 	 */
1460 	if (vp->v_type == VDIR) {
1461 		KASSERT(uio->uio_rw == UIO_READ,
1462 		    ("illegal write attempted on a directory"));
1463 		if (!vfs_allow_read_dir)
1464 			return (EISDIR);
1465 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1466 			return (EISDIR);
1467 	}
1468 
1469 	foffset_lock_uio(fp, uio, flags);
1470 	if (do_vn_io_fault(vp, uio)) {
1471 		args.kind = VN_IO_FAULT_FOP;
1472 		args.args.fop_args.fp = fp;
1473 		args.args.fop_args.doio = doio;
1474 		args.cred = active_cred;
1475 		args.flags = flags | FOF_OFFSET;
1476 		if (uio->uio_rw == UIO_READ) {
1477 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1478 			    uio->uio_offset + uio->uio_resid);
1479 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1480 		    (flags & FOF_OFFSET) == 0) {
1481 			/* For appenders, punt and lock the whole range. */
1482 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1483 		} else {
1484 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1485 			    uio->uio_offset + uio->uio_resid);
1486 		}
1487 		error = vn_io_fault1(vp, uio, &args, td);
1488 		vn_rangelock_unlock(vp, rl_cookie);
1489 	} else {
1490 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1491 	}
1492 	foffset_unlock_uio(fp, uio, flags);
1493 	return (error);
1494 }
1495 
1496 /*
1497  * Helper function to perform the requested uiomove operation using
1498  * the held pages for io->uio_iov[0].iov_base buffer instead of
1499  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1500  * instead of iov_base prevents page faults that could occur due to
1501  * pmap_collect() invalidating the mapping created by
1502  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1503  * object cleanup revoking the write access from page mappings.
1504  *
1505  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1506  * instead of plain uiomove().
1507  */
1508 int
1509 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1510 {
1511 	struct uio transp_uio;
1512 	struct iovec transp_iov[1];
1513 	struct thread *td;
1514 	size_t adv;
1515 	int error, pgadv;
1516 
1517 	td = curthread;
1518 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1519 	    uio->uio_segflg != UIO_USERSPACE)
1520 		return (uiomove(data, xfersize, uio));
1521 
1522 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1523 	transp_iov[0].iov_base = data;
1524 	transp_uio.uio_iov = &transp_iov[0];
1525 	transp_uio.uio_iovcnt = 1;
1526 	if (xfersize > uio->uio_resid)
1527 		xfersize = uio->uio_resid;
1528 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1529 	transp_uio.uio_offset = 0;
1530 	transp_uio.uio_segflg = UIO_SYSSPACE;
1531 	/*
1532 	 * Since transp_iov points to data, and td_ma page array
1533 	 * corresponds to original uio->uio_iov, we need to invert the
1534 	 * direction of the i/o operation as passed to
1535 	 * uiomove_fromphys().
1536 	 */
1537 	switch (uio->uio_rw) {
1538 	case UIO_WRITE:
1539 		transp_uio.uio_rw = UIO_READ;
1540 		break;
1541 	case UIO_READ:
1542 		transp_uio.uio_rw = UIO_WRITE;
1543 		break;
1544 	}
1545 	transp_uio.uio_td = uio->uio_td;
1546 	error = uiomove_fromphys(td->td_ma,
1547 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1548 	    xfersize, &transp_uio);
1549 	adv = xfersize - transp_uio.uio_resid;
1550 	pgadv =
1551 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1552 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1553 	td->td_ma += pgadv;
1554 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1555 	    pgadv));
1556 	td->td_ma_cnt -= pgadv;
1557 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1558 	uio->uio_iov->iov_len -= adv;
1559 	uio->uio_resid -= adv;
1560 	uio->uio_offset += adv;
1561 	return (error);
1562 }
1563 
1564 int
1565 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1566     struct uio *uio)
1567 {
1568 	struct thread *td;
1569 	vm_offset_t iov_base;
1570 	int cnt, pgadv;
1571 
1572 	td = curthread;
1573 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1574 	    uio->uio_segflg != UIO_USERSPACE)
1575 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1576 
1577 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1578 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1579 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1580 	switch (uio->uio_rw) {
1581 	case UIO_WRITE:
1582 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1583 		    offset, cnt);
1584 		break;
1585 	case UIO_READ:
1586 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1587 		    cnt);
1588 		break;
1589 	}
1590 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1591 	td->td_ma += pgadv;
1592 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1593 	    pgadv));
1594 	td->td_ma_cnt -= pgadv;
1595 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1596 	uio->uio_iov->iov_len -= cnt;
1597 	uio->uio_resid -= cnt;
1598 	uio->uio_offset += cnt;
1599 	return (0);
1600 }
1601 
1602 /*
1603  * File table truncate routine.
1604  */
1605 static int
1606 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1607     struct thread *td)
1608 {
1609 	struct mount *mp;
1610 	struct vnode *vp;
1611 	void *rl_cookie;
1612 	int error;
1613 
1614 	vp = fp->f_vnode;
1615 
1616 retry:
1617 	/*
1618 	 * Lock the whole range for truncation.  Otherwise split i/o
1619 	 * might happen partly before and partly after the truncation.
1620 	 */
1621 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1622 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1623 	if (error)
1624 		goto out1;
1625 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1626 	AUDIT_ARG_VNODE1(vp);
1627 	if (vp->v_type == VDIR) {
1628 		error = EISDIR;
1629 		goto out;
1630 	}
1631 #ifdef MAC
1632 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1633 	if (error)
1634 		goto out;
1635 #endif
1636 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1637 	    fp->f_cred);
1638 out:
1639 	VOP_UNLOCK(vp);
1640 	vn_finished_write(mp);
1641 out1:
1642 	vn_rangelock_unlock(vp, rl_cookie);
1643 	if (error == ERELOOKUP)
1644 		goto retry;
1645 	return (error);
1646 }
1647 
1648 /*
1649  * Truncate a file that is already locked.
1650  */
1651 int
1652 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1653     struct ucred *cred)
1654 {
1655 	struct vattr vattr;
1656 	int error;
1657 
1658 	error = VOP_ADD_WRITECOUNT(vp, 1);
1659 	if (error == 0) {
1660 		VATTR_NULL(&vattr);
1661 		vattr.va_size = length;
1662 		if (sync)
1663 			vattr.va_vaflags |= VA_SYNC;
1664 		error = VOP_SETATTR(vp, &vattr, cred);
1665 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1666 	}
1667 	return (error);
1668 }
1669 
1670 /*
1671  * File table vnode stat routine.
1672  */
1673 int
1674 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1675     struct thread *td)
1676 {
1677 	struct vnode *vp = fp->f_vnode;
1678 	int error;
1679 
1680 	vn_lock(vp, LK_SHARED | LK_RETRY);
1681 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1682 	VOP_UNLOCK(vp);
1683 
1684 	return (error);
1685 }
1686 
1687 /*
1688  * File table vnode ioctl routine.
1689  */
1690 static int
1691 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1692     struct thread *td)
1693 {
1694 	struct vattr vattr;
1695 	struct vnode *vp;
1696 	struct fiobmap2_arg *bmarg;
1697 	int error;
1698 
1699 	vp = fp->f_vnode;
1700 	switch (vp->v_type) {
1701 	case VDIR:
1702 	case VREG:
1703 		switch (com) {
1704 		case FIONREAD:
1705 			vn_lock(vp, LK_SHARED | LK_RETRY);
1706 			error = VOP_GETATTR(vp, &vattr, active_cred);
1707 			VOP_UNLOCK(vp);
1708 			if (error == 0)
1709 				*(int *)data = vattr.va_size - fp->f_offset;
1710 			return (error);
1711 		case FIOBMAP2:
1712 			bmarg = (struct fiobmap2_arg *)data;
1713 			vn_lock(vp, LK_SHARED | LK_RETRY);
1714 #ifdef MAC
1715 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1716 			    vp);
1717 			if (error == 0)
1718 #endif
1719 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1720 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1721 			VOP_UNLOCK(vp);
1722 			return (error);
1723 		case FIONBIO:
1724 		case FIOASYNC:
1725 			return (0);
1726 		default:
1727 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1728 			    active_cred, td));
1729 		}
1730 		break;
1731 	case VCHR:
1732 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1733 		    active_cred, td));
1734 	default:
1735 		return (ENOTTY);
1736 	}
1737 }
1738 
1739 /*
1740  * File table vnode poll routine.
1741  */
1742 static int
1743 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1744     struct thread *td)
1745 {
1746 	struct vnode *vp;
1747 	int error;
1748 
1749 	vp = fp->f_vnode;
1750 #if defined(MAC) || defined(AUDIT)
1751 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1752 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1753 		AUDIT_ARG_VNODE1(vp);
1754 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1755 		VOP_UNLOCK(vp);
1756 		if (error != 0)
1757 			return (error);
1758 	}
1759 #endif
1760 	error = VOP_POLL(vp, events, fp->f_cred, td);
1761 	return (error);
1762 }
1763 
1764 /*
1765  * Acquire the requested lock and then check for validity.  LK_RETRY
1766  * permits vn_lock to return doomed vnodes.
1767  */
1768 static int __noinline
1769 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1770     int error)
1771 {
1772 
1773 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1774 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1775 
1776 	if (error == 0)
1777 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1778 
1779 	if ((flags & LK_RETRY) == 0) {
1780 		if (error == 0) {
1781 			VOP_UNLOCK(vp);
1782 			error = ENOENT;
1783 		}
1784 		return (error);
1785 	}
1786 
1787 	/*
1788 	 * LK_RETRY case.
1789 	 *
1790 	 * Nothing to do if we got the lock.
1791 	 */
1792 	if (error == 0)
1793 		return (0);
1794 
1795 	/*
1796 	 * Interlock was dropped by the call in _vn_lock.
1797 	 */
1798 	flags &= ~LK_INTERLOCK;
1799 	do {
1800 		error = VOP_LOCK1(vp, flags, file, line);
1801 	} while (error != 0);
1802 	return (0);
1803 }
1804 
1805 int
1806 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1807 {
1808 	int error;
1809 
1810 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1811 	    ("vn_lock: no locktype (%d passed)", flags));
1812 	VNPASS(vp->v_holdcnt > 0, vp);
1813 	error = VOP_LOCK1(vp, flags, file, line);
1814 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1815 		return (_vn_lock_fallback(vp, flags, file, line, error));
1816 	return (0);
1817 }
1818 
1819 /*
1820  * File table vnode close routine.
1821  */
1822 static int
1823 vn_closefile(struct file *fp, struct thread *td)
1824 {
1825 	struct vnode *vp;
1826 	struct flock lf;
1827 	int error;
1828 	bool ref;
1829 
1830 	vp = fp->f_vnode;
1831 	fp->f_ops = &badfileops;
1832 	ref = (fp->f_flag & FHASLOCK) != 0;
1833 
1834 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1835 
1836 	if (__predict_false(ref)) {
1837 		lf.l_whence = SEEK_SET;
1838 		lf.l_start = 0;
1839 		lf.l_len = 0;
1840 		lf.l_type = F_UNLCK;
1841 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1842 		vrele(vp);
1843 	}
1844 	return (error);
1845 }
1846 
1847 /*
1848  * Preparing to start a filesystem write operation. If the operation is
1849  * permitted, then we bump the count of operations in progress and
1850  * proceed. If a suspend request is in progress, we wait until the
1851  * suspension is over, and then proceed.
1852  */
1853 static int
1854 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1855 {
1856 	struct mount_pcpu *mpcpu;
1857 	int error, mflags;
1858 
1859 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1860 	    vfs_op_thread_enter(mp, mpcpu)) {
1861 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1862 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1863 		vfs_op_thread_exit(mp, mpcpu);
1864 		return (0);
1865 	}
1866 
1867 	if (mplocked)
1868 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1869 	else
1870 		MNT_ILOCK(mp);
1871 
1872 	error = 0;
1873 
1874 	/*
1875 	 * Check on status of suspension.
1876 	 */
1877 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1878 	    mp->mnt_susp_owner != curthread) {
1879 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1880 		    (flags & PCATCH) : 0) | (PUSER - 1);
1881 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1882 			if (flags & V_NOWAIT) {
1883 				error = EWOULDBLOCK;
1884 				goto unlock;
1885 			}
1886 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1887 			    "suspfs", 0);
1888 			if (error)
1889 				goto unlock;
1890 		}
1891 	}
1892 	if (flags & V_XSLEEP)
1893 		goto unlock;
1894 	mp->mnt_writeopcount++;
1895 unlock:
1896 	if (error != 0 || (flags & V_XSLEEP) != 0)
1897 		MNT_REL(mp);
1898 	MNT_IUNLOCK(mp);
1899 	return (error);
1900 }
1901 
1902 int
1903 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1904 {
1905 	struct mount *mp;
1906 	int error;
1907 
1908 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1909 	    ("V_MNTREF requires mp"));
1910 
1911 	error = 0;
1912 	/*
1913 	 * If a vnode is provided, get and return the mount point that
1914 	 * to which it will write.
1915 	 */
1916 	if (vp != NULL) {
1917 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1918 			*mpp = NULL;
1919 			if (error != EOPNOTSUPP)
1920 				return (error);
1921 			return (0);
1922 		}
1923 	}
1924 	if ((mp = *mpp) == NULL)
1925 		return (0);
1926 
1927 	/*
1928 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1929 	 * a vfs_ref().
1930 	 * As long as a vnode is not provided we need to acquire a
1931 	 * refcount for the provided mountpoint too, in order to
1932 	 * emulate a vfs_ref().
1933 	 */
1934 	if (vp == NULL && (flags & V_MNTREF) == 0)
1935 		vfs_ref(mp);
1936 
1937 	return (vn_start_write_refed(mp, flags, false));
1938 }
1939 
1940 /*
1941  * Secondary suspension. Used by operations such as vop_inactive
1942  * routines that are needed by the higher level functions. These
1943  * are allowed to proceed until all the higher level functions have
1944  * completed (indicated by mnt_writeopcount dropping to zero). At that
1945  * time, these operations are halted until the suspension is over.
1946  */
1947 int
1948 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1949 {
1950 	struct mount *mp;
1951 	int error;
1952 
1953 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1954 	    ("V_MNTREF requires mp"));
1955 
1956  retry:
1957 	if (vp != NULL) {
1958 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1959 			*mpp = NULL;
1960 			if (error != EOPNOTSUPP)
1961 				return (error);
1962 			return (0);
1963 		}
1964 	}
1965 	/*
1966 	 * If we are not suspended or have not yet reached suspended
1967 	 * mode, then let the operation proceed.
1968 	 */
1969 	if ((mp = *mpp) == NULL)
1970 		return (0);
1971 
1972 	/*
1973 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1974 	 * a vfs_ref().
1975 	 * As long as a vnode is not provided we need to acquire a
1976 	 * refcount for the provided mountpoint too, in order to
1977 	 * emulate a vfs_ref().
1978 	 */
1979 	MNT_ILOCK(mp);
1980 	if (vp == NULL && (flags & V_MNTREF) == 0)
1981 		MNT_REF(mp);
1982 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1983 		mp->mnt_secondary_writes++;
1984 		mp->mnt_secondary_accwrites++;
1985 		MNT_IUNLOCK(mp);
1986 		return (0);
1987 	}
1988 	if (flags & V_NOWAIT) {
1989 		MNT_REL(mp);
1990 		MNT_IUNLOCK(mp);
1991 		return (EWOULDBLOCK);
1992 	}
1993 	/*
1994 	 * Wait for the suspension to finish.
1995 	 */
1996 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1997 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1998 	    "suspfs", 0);
1999 	vfs_rel(mp);
2000 	if (error == 0)
2001 		goto retry;
2002 	return (error);
2003 }
2004 
2005 /*
2006  * Filesystem write operation has completed. If we are suspending and this
2007  * operation is the last one, notify the suspender that the suspension is
2008  * now in effect.
2009  */
2010 void
2011 vn_finished_write(struct mount *mp)
2012 {
2013 	struct mount_pcpu *mpcpu;
2014 	int c;
2015 
2016 	if (mp == NULL)
2017 		return;
2018 
2019 	if (vfs_op_thread_enter(mp, mpcpu)) {
2020 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2021 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2022 		vfs_op_thread_exit(mp, mpcpu);
2023 		return;
2024 	}
2025 
2026 	MNT_ILOCK(mp);
2027 	vfs_assert_mount_counters(mp);
2028 	MNT_REL(mp);
2029 	c = --mp->mnt_writeopcount;
2030 	if (mp->mnt_vfs_ops == 0) {
2031 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2032 		MNT_IUNLOCK(mp);
2033 		return;
2034 	}
2035 	if (c < 0)
2036 		vfs_dump_mount_counters(mp);
2037 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2038 		wakeup(&mp->mnt_writeopcount);
2039 	MNT_IUNLOCK(mp);
2040 }
2041 
2042 /*
2043  * Filesystem secondary write operation has completed. If we are
2044  * suspending and this operation is the last one, notify the suspender
2045  * that the suspension is now in effect.
2046  */
2047 void
2048 vn_finished_secondary_write(struct mount *mp)
2049 {
2050 	if (mp == NULL)
2051 		return;
2052 	MNT_ILOCK(mp);
2053 	MNT_REL(mp);
2054 	mp->mnt_secondary_writes--;
2055 	if (mp->mnt_secondary_writes < 0)
2056 		panic("vn_finished_secondary_write: neg cnt");
2057 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2058 	    mp->mnt_secondary_writes <= 0)
2059 		wakeup(&mp->mnt_secondary_writes);
2060 	MNT_IUNLOCK(mp);
2061 }
2062 
2063 /*
2064  * Request a filesystem to suspend write operations.
2065  */
2066 int
2067 vfs_write_suspend(struct mount *mp, int flags)
2068 {
2069 	int error;
2070 
2071 	vfs_op_enter(mp);
2072 
2073 	MNT_ILOCK(mp);
2074 	vfs_assert_mount_counters(mp);
2075 	if (mp->mnt_susp_owner == curthread) {
2076 		vfs_op_exit_locked(mp);
2077 		MNT_IUNLOCK(mp);
2078 		return (EALREADY);
2079 	}
2080 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2081 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2082 
2083 	/*
2084 	 * Unmount holds a write reference on the mount point.  If we
2085 	 * own busy reference and drain for writers, we deadlock with
2086 	 * the reference draining in the unmount path.  Callers of
2087 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2088 	 * vfs_busy() reference is owned and caller is not in the
2089 	 * unmount context.
2090 	 */
2091 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2092 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2093 		vfs_op_exit_locked(mp);
2094 		MNT_IUNLOCK(mp);
2095 		return (EBUSY);
2096 	}
2097 
2098 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2099 	mp->mnt_susp_owner = curthread;
2100 	if (mp->mnt_writeopcount > 0)
2101 		(void) msleep(&mp->mnt_writeopcount,
2102 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2103 	else
2104 		MNT_IUNLOCK(mp);
2105 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2106 		vfs_write_resume(mp, 0);
2107 		/* vfs_write_resume does vfs_op_exit() for us */
2108 	}
2109 	return (error);
2110 }
2111 
2112 /*
2113  * Request a filesystem to resume write operations.
2114  */
2115 void
2116 vfs_write_resume(struct mount *mp, int flags)
2117 {
2118 
2119 	MNT_ILOCK(mp);
2120 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2121 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2122 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2123 				       MNTK_SUSPENDED);
2124 		mp->mnt_susp_owner = NULL;
2125 		wakeup(&mp->mnt_writeopcount);
2126 		wakeup(&mp->mnt_flag);
2127 		curthread->td_pflags &= ~TDP_IGNSUSP;
2128 		if ((flags & VR_START_WRITE) != 0) {
2129 			MNT_REF(mp);
2130 			mp->mnt_writeopcount++;
2131 		}
2132 		MNT_IUNLOCK(mp);
2133 		if ((flags & VR_NO_SUSPCLR) == 0)
2134 			VFS_SUSP_CLEAN(mp);
2135 		vfs_op_exit(mp);
2136 	} else if ((flags & VR_START_WRITE) != 0) {
2137 		MNT_REF(mp);
2138 		vn_start_write_refed(mp, 0, true);
2139 	} else {
2140 		MNT_IUNLOCK(mp);
2141 	}
2142 }
2143 
2144 /*
2145  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2146  * methods.
2147  */
2148 int
2149 vfs_write_suspend_umnt(struct mount *mp)
2150 {
2151 	int error;
2152 
2153 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2154 	    ("vfs_write_suspend_umnt: recursed"));
2155 
2156 	/* dounmount() already called vn_start_write(). */
2157 	for (;;) {
2158 		vn_finished_write(mp);
2159 		error = vfs_write_suspend(mp, 0);
2160 		if (error != 0) {
2161 			vn_start_write(NULL, &mp, V_WAIT);
2162 			return (error);
2163 		}
2164 		MNT_ILOCK(mp);
2165 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2166 			break;
2167 		MNT_IUNLOCK(mp);
2168 		vn_start_write(NULL, &mp, V_WAIT);
2169 	}
2170 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2171 	wakeup(&mp->mnt_flag);
2172 	MNT_IUNLOCK(mp);
2173 	curthread->td_pflags |= TDP_IGNSUSP;
2174 	return (0);
2175 }
2176 
2177 /*
2178  * Implement kqueues for files by translating it to vnode operation.
2179  */
2180 static int
2181 vn_kqfilter(struct file *fp, struct knote *kn)
2182 {
2183 
2184 	return (VOP_KQFILTER(fp->f_vnode, kn));
2185 }
2186 
2187 int
2188 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2189 {
2190 	if ((fp->f_flag & FKQALLOWED) == 0)
2191 		return (EBADF);
2192 	return (vn_kqfilter(fp, kn));
2193 }
2194 
2195 /*
2196  * Simplified in-kernel wrapper calls for extended attribute access.
2197  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2198  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2199  */
2200 int
2201 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2202     const char *attrname, int *buflen, char *buf, struct thread *td)
2203 {
2204 	struct uio	auio;
2205 	struct iovec	iov;
2206 	int	error;
2207 
2208 	iov.iov_len = *buflen;
2209 	iov.iov_base = buf;
2210 
2211 	auio.uio_iov = &iov;
2212 	auio.uio_iovcnt = 1;
2213 	auio.uio_rw = UIO_READ;
2214 	auio.uio_segflg = UIO_SYSSPACE;
2215 	auio.uio_td = td;
2216 	auio.uio_offset = 0;
2217 	auio.uio_resid = *buflen;
2218 
2219 	if ((ioflg & IO_NODELOCKED) == 0)
2220 		vn_lock(vp, LK_SHARED | LK_RETRY);
2221 
2222 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2223 
2224 	/* authorize attribute retrieval as kernel */
2225 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2226 	    td);
2227 
2228 	if ((ioflg & IO_NODELOCKED) == 0)
2229 		VOP_UNLOCK(vp);
2230 
2231 	if (error == 0) {
2232 		*buflen = *buflen - auio.uio_resid;
2233 	}
2234 
2235 	return (error);
2236 }
2237 
2238 /*
2239  * XXX failure mode if partially written?
2240  */
2241 int
2242 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2243     const char *attrname, int buflen, char *buf, struct thread *td)
2244 {
2245 	struct uio	auio;
2246 	struct iovec	iov;
2247 	struct mount	*mp;
2248 	int	error;
2249 
2250 	iov.iov_len = buflen;
2251 	iov.iov_base = buf;
2252 
2253 	auio.uio_iov = &iov;
2254 	auio.uio_iovcnt = 1;
2255 	auio.uio_rw = UIO_WRITE;
2256 	auio.uio_segflg = UIO_SYSSPACE;
2257 	auio.uio_td = td;
2258 	auio.uio_offset = 0;
2259 	auio.uio_resid = buflen;
2260 
2261 	if ((ioflg & IO_NODELOCKED) == 0) {
2262 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2263 			return (error);
2264 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2265 	}
2266 
2267 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2268 
2269 	/* authorize attribute setting as kernel */
2270 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2271 
2272 	if ((ioflg & IO_NODELOCKED) == 0) {
2273 		vn_finished_write(mp);
2274 		VOP_UNLOCK(vp);
2275 	}
2276 
2277 	return (error);
2278 }
2279 
2280 int
2281 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2282     const char *attrname, struct thread *td)
2283 {
2284 	struct mount	*mp;
2285 	int	error;
2286 
2287 	if ((ioflg & IO_NODELOCKED) == 0) {
2288 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2289 			return (error);
2290 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2291 	}
2292 
2293 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2294 
2295 	/* authorize attribute removal as kernel */
2296 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2297 	if (error == EOPNOTSUPP)
2298 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2299 		    NULL, td);
2300 
2301 	if ((ioflg & IO_NODELOCKED) == 0) {
2302 		vn_finished_write(mp);
2303 		VOP_UNLOCK(vp);
2304 	}
2305 
2306 	return (error);
2307 }
2308 
2309 static int
2310 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2311     struct vnode **rvp)
2312 {
2313 
2314 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2315 }
2316 
2317 int
2318 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2319 {
2320 
2321 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2322 	    lkflags, rvp));
2323 }
2324 
2325 int
2326 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2327     int lkflags, struct vnode **rvp)
2328 {
2329 	struct mount *mp;
2330 	int ltype, error;
2331 
2332 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2333 	mp = vp->v_mount;
2334 	ltype = VOP_ISLOCKED(vp);
2335 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2336 	    ("vn_vget_ino: vp not locked"));
2337 	error = vfs_busy(mp, MBF_NOWAIT);
2338 	if (error != 0) {
2339 		vfs_ref(mp);
2340 		VOP_UNLOCK(vp);
2341 		error = vfs_busy(mp, 0);
2342 		vn_lock(vp, ltype | LK_RETRY);
2343 		vfs_rel(mp);
2344 		if (error != 0)
2345 			return (ENOENT);
2346 		if (VN_IS_DOOMED(vp)) {
2347 			vfs_unbusy(mp);
2348 			return (ENOENT);
2349 		}
2350 	}
2351 	VOP_UNLOCK(vp);
2352 	error = alloc(mp, alloc_arg, lkflags, rvp);
2353 	vfs_unbusy(mp);
2354 	if (error != 0 || *rvp != vp)
2355 		vn_lock(vp, ltype | LK_RETRY);
2356 	if (VN_IS_DOOMED(vp)) {
2357 		if (error == 0) {
2358 			if (*rvp == vp)
2359 				vunref(vp);
2360 			else
2361 				vput(*rvp);
2362 		}
2363 		error = ENOENT;
2364 	}
2365 	return (error);
2366 }
2367 
2368 int
2369 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2370     struct thread *td)
2371 {
2372 	off_t lim;
2373 	bool ktr_write;
2374 
2375 	if (td == NULL)
2376 		return (0);
2377 
2378 	/*
2379 	 * There are conditions where the limit is to be ignored.
2380 	 * However, since it is almost never reached, check it first.
2381 	 */
2382 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2383 	lim = lim_cur(td, RLIMIT_FSIZE);
2384 	if (__predict_false(ktr_write))
2385 		lim = td->td_ktr_io_lim;
2386 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2387 		return (0);
2388 
2389 	/*
2390 	 * The limit is reached.
2391 	 */
2392 	if (vp->v_type != VREG ||
2393 	    (td->td_pflags2 & TDP2_ACCT) != 0)
2394 		return (0);
2395 
2396 	if (!ktr_write || ktr_filesize_limit_signal) {
2397 		PROC_LOCK(td->td_proc);
2398 		kern_psignal(td->td_proc, SIGXFSZ);
2399 		PROC_UNLOCK(td->td_proc);
2400 	}
2401 	return (EFBIG);
2402 }
2403 
2404 int
2405 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2406     struct thread *td)
2407 {
2408 	struct vnode *vp;
2409 
2410 	vp = fp->f_vnode;
2411 #ifdef AUDIT
2412 	vn_lock(vp, LK_SHARED | LK_RETRY);
2413 	AUDIT_ARG_VNODE1(vp);
2414 	VOP_UNLOCK(vp);
2415 #endif
2416 	return (setfmode(td, active_cred, vp, mode));
2417 }
2418 
2419 int
2420 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2421     struct thread *td)
2422 {
2423 	struct vnode *vp;
2424 
2425 	vp = fp->f_vnode;
2426 #ifdef AUDIT
2427 	vn_lock(vp, LK_SHARED | LK_RETRY);
2428 	AUDIT_ARG_VNODE1(vp);
2429 	VOP_UNLOCK(vp);
2430 #endif
2431 	return (setfown(td, active_cred, vp, uid, gid));
2432 }
2433 
2434 void
2435 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2436 {
2437 	vm_object_t object;
2438 
2439 	if ((object = vp->v_object) == NULL)
2440 		return;
2441 	VM_OBJECT_WLOCK(object);
2442 	vm_object_page_remove(object, start, end, 0);
2443 	VM_OBJECT_WUNLOCK(object);
2444 }
2445 
2446 int
2447 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2448     struct ucred *cred)
2449 {
2450 	struct vattr va;
2451 	daddr_t bn, bnp;
2452 	uint64_t bsize;
2453 	off_t noff;
2454 	int error;
2455 
2456 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2457 	    ("%s: Wrong command %lu", __func__, cmd));
2458 	ASSERT_VOP_LOCKED(vp, "vn_bmap_seekhole_locked");
2459 
2460 	if (vp->v_type != VREG) {
2461 		error = ENOTTY;
2462 		goto out;
2463 	}
2464 	error = VOP_GETATTR(vp, &va, cred);
2465 	if (error != 0)
2466 		goto out;
2467 	noff = *off;
2468 	if (noff >= va.va_size) {
2469 		error = ENXIO;
2470 		goto out;
2471 	}
2472 	bsize = vp->v_mount->mnt_stat.f_iosize;
2473 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2474 	    noff % bsize) {
2475 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2476 		if (error == EOPNOTSUPP) {
2477 			error = ENOTTY;
2478 			goto out;
2479 		}
2480 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2481 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2482 			noff = bn * bsize;
2483 			if (noff < *off)
2484 				noff = *off;
2485 			goto out;
2486 		}
2487 	}
2488 	if (noff > va.va_size)
2489 		noff = va.va_size;
2490 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2491 	if (cmd == FIOSEEKDATA)
2492 		error = ENXIO;
2493 out:
2494 	if (error == 0)
2495 		*off = noff;
2496 	return (error);
2497 }
2498 
2499 int
2500 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2501 {
2502 	int error;
2503 
2504 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2505 	    ("%s: Wrong command %lu", __func__, cmd));
2506 
2507 	if (vn_lock(vp, LK_SHARED) != 0)
2508 		return (EBADF);
2509 	error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2510 	VOP_UNLOCK(vp);
2511 	return (error);
2512 }
2513 
2514 int
2515 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2516 {
2517 	struct ucred *cred;
2518 	struct vnode *vp;
2519 	struct vattr vattr;
2520 	off_t foffset, size;
2521 	int error, noneg;
2522 
2523 	cred = td->td_ucred;
2524 	vp = fp->f_vnode;
2525 	foffset = foffset_lock(fp, 0);
2526 	noneg = (vp->v_type != VCHR);
2527 	error = 0;
2528 	switch (whence) {
2529 	case L_INCR:
2530 		if (noneg &&
2531 		    (foffset < 0 ||
2532 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2533 			error = EOVERFLOW;
2534 			break;
2535 		}
2536 		offset += foffset;
2537 		break;
2538 	case L_XTND:
2539 		vn_lock(vp, LK_SHARED | LK_RETRY);
2540 		error = VOP_GETATTR(vp, &vattr, cred);
2541 		VOP_UNLOCK(vp);
2542 		if (error)
2543 			break;
2544 
2545 		/*
2546 		 * If the file references a disk device, then fetch
2547 		 * the media size and use that to determine the ending
2548 		 * offset.
2549 		 */
2550 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2551 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2552 			vattr.va_size = size;
2553 		if (noneg &&
2554 		    (vattr.va_size > OFF_MAX ||
2555 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2556 			error = EOVERFLOW;
2557 			break;
2558 		}
2559 		offset += vattr.va_size;
2560 		break;
2561 	case L_SET:
2562 		break;
2563 	case SEEK_DATA:
2564 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2565 		if (error == ENOTTY)
2566 			error = EINVAL;
2567 		break;
2568 	case SEEK_HOLE:
2569 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2570 		if (error == ENOTTY)
2571 			error = EINVAL;
2572 		break;
2573 	default:
2574 		error = EINVAL;
2575 	}
2576 	if (error == 0 && noneg && offset < 0)
2577 		error = EINVAL;
2578 	if (error != 0)
2579 		goto drop;
2580 	VFS_KNOTE_UNLOCKED(vp, 0);
2581 	td->td_uretoff.tdu_off = offset;
2582 drop:
2583 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2584 	return (error);
2585 }
2586 
2587 int
2588 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2589     struct thread *td)
2590 {
2591 	int error;
2592 
2593 	/*
2594 	 * Grant permission if the caller is the owner of the file, or
2595 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2596 	 * on the file.  If the time pointer is null, then write
2597 	 * permission on the file is also sufficient.
2598 	 *
2599 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2600 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2601 	 * will be allowed to set the times [..] to the current
2602 	 * server time.
2603 	 */
2604 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2605 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2606 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2607 	return (error);
2608 }
2609 
2610 int
2611 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2612 {
2613 	struct vnode *vp;
2614 	int error;
2615 
2616 	if (fp->f_type == DTYPE_FIFO)
2617 		kif->kf_type = KF_TYPE_FIFO;
2618 	else
2619 		kif->kf_type = KF_TYPE_VNODE;
2620 	vp = fp->f_vnode;
2621 	vref(vp);
2622 	FILEDESC_SUNLOCK(fdp);
2623 	error = vn_fill_kinfo_vnode(vp, kif);
2624 	vrele(vp);
2625 	FILEDESC_SLOCK(fdp);
2626 	return (error);
2627 }
2628 
2629 static inline void
2630 vn_fill_junk(struct kinfo_file *kif)
2631 {
2632 	size_t len, olen;
2633 
2634 	/*
2635 	 * Simulate vn_fullpath returning changing values for a given
2636 	 * vp during e.g. coredump.
2637 	 */
2638 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2639 	olen = strlen(kif->kf_path);
2640 	if (len < olen)
2641 		strcpy(&kif->kf_path[len - 1], "$");
2642 	else
2643 		for (; olen < len; olen++)
2644 			strcpy(&kif->kf_path[olen], "A");
2645 }
2646 
2647 int
2648 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2649 {
2650 	struct vattr va;
2651 	char *fullpath, *freepath;
2652 	int error;
2653 
2654 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2655 	freepath = NULL;
2656 	fullpath = "-";
2657 	error = vn_fullpath(vp, &fullpath, &freepath);
2658 	if (error == 0) {
2659 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2660 	}
2661 	if (freepath != NULL)
2662 		free(freepath, M_TEMP);
2663 
2664 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2665 		vn_fill_junk(kif);
2666 	);
2667 
2668 	/*
2669 	 * Retrieve vnode attributes.
2670 	 */
2671 	va.va_fsid = VNOVAL;
2672 	va.va_rdev = NODEV;
2673 	vn_lock(vp, LK_SHARED | LK_RETRY);
2674 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2675 	VOP_UNLOCK(vp);
2676 	if (error != 0)
2677 		return (error);
2678 	if (va.va_fsid != VNOVAL)
2679 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2680 	else
2681 		kif->kf_un.kf_file.kf_file_fsid =
2682 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2683 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2684 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2685 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2686 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2687 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2688 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2689 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2690 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2691 	return (0);
2692 }
2693 
2694 int
2695 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2696     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2697     struct thread *td)
2698 {
2699 #ifdef HWPMC_HOOKS
2700 	struct pmckern_map_in pkm;
2701 #endif
2702 	struct mount *mp;
2703 	struct vnode *vp;
2704 	vm_object_t object;
2705 	vm_prot_t maxprot;
2706 	boolean_t writecounted;
2707 	int error;
2708 
2709 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2710     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2711 	/*
2712 	 * POSIX shared-memory objects are defined to have
2713 	 * kernel persistence, and are not defined to support
2714 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2715 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2716 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2717 	 * flag to request this behavior.
2718 	 */
2719 	if ((fp->f_flag & FPOSIXSHM) != 0)
2720 		flags |= MAP_NOSYNC;
2721 #endif
2722 	vp = fp->f_vnode;
2723 
2724 	/*
2725 	 * Ensure that file and memory protections are
2726 	 * compatible.  Note that we only worry about
2727 	 * writability if mapping is shared; in this case,
2728 	 * current and max prot are dictated by the open file.
2729 	 * XXX use the vnode instead?  Problem is: what
2730 	 * credentials do we use for determination? What if
2731 	 * proc does a setuid?
2732 	 */
2733 	mp = vp->v_mount;
2734 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2735 		maxprot = VM_PROT_NONE;
2736 		if ((prot & VM_PROT_EXECUTE) != 0)
2737 			return (EACCES);
2738 	} else
2739 		maxprot = VM_PROT_EXECUTE;
2740 	if ((fp->f_flag & FREAD) != 0)
2741 		maxprot |= VM_PROT_READ;
2742 	else if ((prot & VM_PROT_READ) != 0)
2743 		return (EACCES);
2744 
2745 	/*
2746 	 * If we are sharing potential changes via MAP_SHARED and we
2747 	 * are trying to get write permission although we opened it
2748 	 * without asking for it, bail out.
2749 	 */
2750 	if ((flags & MAP_SHARED) != 0) {
2751 		if ((fp->f_flag & FWRITE) != 0)
2752 			maxprot |= VM_PROT_WRITE;
2753 		else if ((prot & VM_PROT_WRITE) != 0)
2754 			return (EACCES);
2755 	} else {
2756 		maxprot |= VM_PROT_WRITE;
2757 		cap_maxprot |= VM_PROT_WRITE;
2758 	}
2759 	maxprot &= cap_maxprot;
2760 
2761 	/*
2762 	 * For regular files and shared memory, POSIX requires that
2763 	 * the value of foff be a legitimate offset within the data
2764 	 * object.  In particular, negative offsets are invalid.
2765 	 * Blocking negative offsets and overflows here avoids
2766 	 * possible wraparound or user-level access into reserved
2767 	 * ranges of the data object later.  In contrast, POSIX does
2768 	 * not dictate how offsets are used by device drivers, so in
2769 	 * the case of a device mapping a negative offset is passed
2770 	 * on.
2771 	 */
2772 	if (
2773 #ifdef _LP64
2774 	    size > OFF_MAX ||
2775 #endif
2776 	    foff > OFF_MAX - size)
2777 		return (EINVAL);
2778 
2779 	writecounted = FALSE;
2780 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2781 	    &foff, &object, &writecounted);
2782 	if (error != 0)
2783 		return (error);
2784 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2785 	    foff, writecounted, td);
2786 	if (error != 0) {
2787 		/*
2788 		 * If this mapping was accounted for in the vnode's
2789 		 * writecount, then undo that now.
2790 		 */
2791 		if (writecounted)
2792 			vm_pager_release_writecount(object, 0, size);
2793 		vm_object_deallocate(object);
2794 	}
2795 #ifdef HWPMC_HOOKS
2796 	/* Inform hwpmc(4) if an executable is being mapped. */
2797 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2798 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2799 			pkm.pm_file = vp;
2800 			pkm.pm_address = (uintptr_t) *addr;
2801 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2802 		}
2803 	}
2804 #endif
2805 	return (error);
2806 }
2807 
2808 void
2809 vn_fsid(struct vnode *vp, struct vattr *va)
2810 {
2811 	fsid_t *f;
2812 
2813 	f = &vp->v_mount->mnt_stat.f_fsid;
2814 	va->va_fsid = (uint32_t)f->val[1];
2815 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2816 	va->va_fsid += (uint32_t)f->val[0];
2817 }
2818 
2819 int
2820 vn_fsync_buf(struct vnode *vp, int waitfor)
2821 {
2822 	struct buf *bp, *nbp;
2823 	struct bufobj *bo;
2824 	struct mount *mp;
2825 	int error, maxretry;
2826 
2827 	error = 0;
2828 	maxretry = 10000;     /* large, arbitrarily chosen */
2829 	mp = NULL;
2830 	if (vp->v_type == VCHR) {
2831 		VI_LOCK(vp);
2832 		mp = vp->v_rdev->si_mountpt;
2833 		VI_UNLOCK(vp);
2834 	}
2835 	bo = &vp->v_bufobj;
2836 	BO_LOCK(bo);
2837 loop1:
2838 	/*
2839 	 * MARK/SCAN initialization to avoid infinite loops.
2840 	 */
2841         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2842 		bp->b_vflags &= ~BV_SCANNED;
2843 		bp->b_error = 0;
2844 	}
2845 
2846 	/*
2847 	 * Flush all dirty buffers associated with a vnode.
2848 	 */
2849 loop2:
2850 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2851 		if ((bp->b_vflags & BV_SCANNED) != 0)
2852 			continue;
2853 		bp->b_vflags |= BV_SCANNED;
2854 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2855 			if (waitfor != MNT_WAIT)
2856 				continue;
2857 			if (BUF_LOCK(bp,
2858 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2859 			    BO_LOCKPTR(bo)) != 0) {
2860 				BO_LOCK(bo);
2861 				goto loop1;
2862 			}
2863 			BO_LOCK(bo);
2864 		}
2865 		BO_UNLOCK(bo);
2866 		KASSERT(bp->b_bufobj == bo,
2867 		    ("bp %p wrong b_bufobj %p should be %p",
2868 		    bp, bp->b_bufobj, bo));
2869 		if ((bp->b_flags & B_DELWRI) == 0)
2870 			panic("fsync: not dirty");
2871 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2872 			vfs_bio_awrite(bp);
2873 		} else {
2874 			bremfree(bp);
2875 			bawrite(bp);
2876 		}
2877 		if (maxretry < 1000)
2878 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2879 		BO_LOCK(bo);
2880 		goto loop2;
2881 	}
2882 
2883 	/*
2884 	 * If synchronous the caller expects us to completely resolve all
2885 	 * dirty buffers in the system.  Wait for in-progress I/O to
2886 	 * complete (which could include background bitmap writes), then
2887 	 * retry if dirty blocks still exist.
2888 	 */
2889 	if (waitfor == MNT_WAIT) {
2890 		bufobj_wwait(bo, 0, 0);
2891 		if (bo->bo_dirty.bv_cnt > 0) {
2892 			/*
2893 			 * If we are unable to write any of these buffers
2894 			 * then we fail now rather than trying endlessly
2895 			 * to write them out.
2896 			 */
2897 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2898 				if ((error = bp->b_error) != 0)
2899 					break;
2900 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2901 			    (error == 0 && --maxretry >= 0))
2902 				goto loop1;
2903 			if (error == 0)
2904 				error = EAGAIN;
2905 		}
2906 	}
2907 	BO_UNLOCK(bo);
2908 	if (error != 0)
2909 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2910 
2911 	return (error);
2912 }
2913 
2914 /*
2915  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2916  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2917  * to do the actual copy.
2918  * vn_generic_copy_file_range() is factored out, so it can be called
2919  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2920  * different file systems.
2921  */
2922 int
2923 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2924     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2925     struct ucred *outcred, struct thread *fsize_td)
2926 {
2927 	int error;
2928 	size_t len;
2929 	uint64_t uval;
2930 
2931 	len = *lenp;
2932 	*lenp = 0;		/* For error returns. */
2933 	error = 0;
2934 
2935 	/* Do some sanity checks on the arguments. */
2936 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2937 		error = EISDIR;
2938 	else if (*inoffp < 0 || *outoffp < 0 ||
2939 	    invp->v_type != VREG || outvp->v_type != VREG)
2940 		error = EINVAL;
2941 	if (error != 0)
2942 		goto out;
2943 
2944 	/* Ensure offset + len does not wrap around. */
2945 	uval = *inoffp;
2946 	uval += len;
2947 	if (uval > INT64_MAX)
2948 		len = INT64_MAX - *inoffp;
2949 	uval = *outoffp;
2950 	uval += len;
2951 	if (uval > INT64_MAX)
2952 		len = INT64_MAX - *outoffp;
2953 	if (len == 0)
2954 		goto out;
2955 
2956 	/*
2957 	 * If the two vnode are for the same file system, call
2958 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2959 	 * which can handle copies across multiple file systems.
2960 	 */
2961 	*lenp = len;
2962 	if (invp->v_mount == outvp->v_mount)
2963 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2964 		    lenp, flags, incred, outcred, fsize_td);
2965 	else
2966 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2967 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2968 out:
2969 	return (error);
2970 }
2971 
2972 /*
2973  * Test len bytes of data starting at dat for all bytes == 0.
2974  * Return true if all bytes are zero, false otherwise.
2975  * Expects dat to be well aligned.
2976  */
2977 static bool
2978 mem_iszero(void *dat, int len)
2979 {
2980 	int i;
2981 	const u_int *p;
2982 	const char *cp;
2983 
2984 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2985 		if (len >= sizeof(*p)) {
2986 			if (*p != 0)
2987 				return (false);
2988 		} else {
2989 			cp = (const char *)p;
2990 			for (i = 0; i < len; i++, cp++)
2991 				if (*cp != '\0')
2992 					return (false);
2993 		}
2994 	}
2995 	return (true);
2996 }
2997 
2998 /*
2999  * Look for a hole in the output file and, if found, adjust *outoffp
3000  * and *xferp to skip past the hole.
3001  * *xferp is the entire hole length to be written and xfer2 is how many bytes
3002  * to be written as 0's upon return.
3003  */
3004 static off_t
3005 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3006     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3007 {
3008 	int error;
3009 	off_t delta;
3010 
3011 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3012 		*dataoffp = *outoffp;
3013 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3014 		    curthread);
3015 		if (error == 0) {
3016 			*holeoffp = *dataoffp;
3017 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3018 			    curthread);
3019 		}
3020 		if (error != 0 || *holeoffp == *dataoffp) {
3021 			/*
3022 			 * Since outvp is unlocked, it may be possible for
3023 			 * another thread to do a truncate(), lseek(), write()
3024 			 * creating a hole at startoff between the above
3025 			 * VOP_IOCTL() calls, if the other thread does not do
3026 			 * rangelocking.
3027 			 * If that happens, *holeoffp == *dataoffp and finding
3028 			 * the hole has failed, so disable vn_skip_hole().
3029 			 */
3030 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3031 			return (xfer2);
3032 		}
3033 		KASSERT(*dataoffp >= *outoffp,
3034 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3035 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3036 		KASSERT(*holeoffp > *dataoffp,
3037 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3038 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3039 	}
3040 
3041 	/*
3042 	 * If there is a hole before the data starts, advance *outoffp and
3043 	 * *xferp past the hole.
3044 	 */
3045 	if (*dataoffp > *outoffp) {
3046 		delta = *dataoffp - *outoffp;
3047 		if (delta >= *xferp) {
3048 			/* Entire *xferp is a hole. */
3049 			*outoffp += *xferp;
3050 			*xferp = 0;
3051 			return (0);
3052 		}
3053 		*xferp -= delta;
3054 		*outoffp += delta;
3055 		xfer2 = MIN(xfer2, *xferp);
3056 	}
3057 
3058 	/*
3059 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3060 	 * that the write ends at the start of the hole.
3061 	 * *holeoffp should always be greater than *outoffp, but for the
3062 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3063 	 * value.
3064 	 */
3065 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3066 		xfer2 = *holeoffp - *outoffp;
3067 	return (xfer2);
3068 }
3069 
3070 /*
3071  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3072  * dat is a maximum of blksize in length and can be written repeatedly in
3073  * the chunk.
3074  * If growfile == true, just grow the file via vn_truncate_locked() instead
3075  * of doing actual writes.
3076  * If checkhole == true, a hole is being punched, so skip over any hole
3077  * already in the output file.
3078  */
3079 static int
3080 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3081     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3082 {
3083 	struct mount *mp;
3084 	off_t dataoff, holeoff, xfer2;
3085 	int error;
3086 
3087 	/*
3088 	 * Loop around doing writes of blksize until write has been completed.
3089 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3090 	 * done for each iteration, since the xfer argument can be very
3091 	 * large if there is a large hole to punch in the output file.
3092 	 */
3093 	error = 0;
3094 	holeoff = 0;
3095 	do {
3096 		xfer2 = MIN(xfer, blksize);
3097 		if (checkhole) {
3098 			/*
3099 			 * Punching a hole.  Skip writing if there is
3100 			 * already a hole in the output file.
3101 			 */
3102 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3103 			    &dataoff, &holeoff, cred);
3104 			if (xfer == 0)
3105 				break;
3106 			if (holeoff < 0)
3107 				checkhole = false;
3108 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3109 			    (intmax_t)xfer2));
3110 		}
3111 		bwillwrite();
3112 		mp = NULL;
3113 		error = vn_start_write(outvp, &mp, V_WAIT);
3114 		if (error != 0)
3115 			break;
3116 		if (growfile) {
3117 			error = vn_lock(outvp, LK_EXCLUSIVE);
3118 			if (error == 0) {
3119 				error = vn_truncate_locked(outvp, outoff + xfer,
3120 				    false, cred);
3121 				VOP_UNLOCK(outvp);
3122 			}
3123 		} else {
3124 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3125 			if (error == 0) {
3126 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3127 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3128 				    curthread->td_ucred, cred, NULL, curthread);
3129 				outoff += xfer2;
3130 				xfer -= xfer2;
3131 				VOP_UNLOCK(outvp);
3132 			}
3133 		}
3134 		if (mp != NULL)
3135 			vn_finished_write(mp);
3136 	} while (!growfile && xfer > 0 && error == 0);
3137 	return (error);
3138 }
3139 
3140 /*
3141  * Copy a byte range of one file to another.  This function can handle the
3142  * case where invp and outvp are on different file systems.
3143  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3144  * is no better file system specific way to do it.
3145  */
3146 int
3147 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3148     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3149     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3150 {
3151 	struct vattr va, inva;
3152 	struct mount *mp;
3153 	struct uio io;
3154 	off_t startoff, endoff, xfer, xfer2;
3155 	u_long blksize;
3156 	int error, interrupted;
3157 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3158 	ssize_t aresid;
3159 	size_t copylen, len, rem, savlen;
3160 	char *dat;
3161 	long holein, holeout;
3162 	struct timespec curts, endts;
3163 
3164 	holein = holeout = 0;
3165 	savlen = len = *lenp;
3166 	error = 0;
3167 	interrupted = 0;
3168 	dat = NULL;
3169 
3170 	error = vn_lock(invp, LK_SHARED);
3171 	if (error != 0)
3172 		goto out;
3173 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3174 		holein = 0;
3175 	if (holein > 0)
3176 		error = VOP_GETATTR(invp, &inva, incred);
3177 	VOP_UNLOCK(invp);
3178 	if (error != 0)
3179 		goto out;
3180 
3181 	mp = NULL;
3182 	error = vn_start_write(outvp, &mp, V_WAIT);
3183 	if (error == 0)
3184 		error = vn_lock(outvp, LK_EXCLUSIVE);
3185 	if (error == 0) {
3186 		/*
3187 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3188 		 * now that outvp is locked.
3189 		 */
3190 		if (fsize_td != NULL) {
3191 			io.uio_offset = *outoffp;
3192 			io.uio_resid = len;
3193 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3194 			if (error != 0)
3195 				error = EFBIG;
3196 		}
3197 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3198 			holeout = 0;
3199 		/*
3200 		 * Holes that are past EOF do not need to be written as a block
3201 		 * of zero bytes.  So, truncate the output file as far as
3202 		 * possible and then use va.va_size to decide if writing 0
3203 		 * bytes is necessary in the loop below.
3204 		 */
3205 		if (error == 0)
3206 			error = VOP_GETATTR(outvp, &va, outcred);
3207 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3208 		    *outoffp + len) {
3209 #ifdef MAC
3210 			error = mac_vnode_check_write(curthread->td_ucred,
3211 			    outcred, outvp);
3212 			if (error == 0)
3213 #endif
3214 				error = vn_truncate_locked(outvp, *outoffp,
3215 				    false, outcred);
3216 			if (error == 0)
3217 				va.va_size = *outoffp;
3218 		}
3219 		VOP_UNLOCK(outvp);
3220 	}
3221 	if (mp != NULL)
3222 		vn_finished_write(mp);
3223 	if (error != 0)
3224 		goto out;
3225 
3226 	/*
3227 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3228 	 * If hole sizes aren't available, set the blksize to the larger
3229 	 * f_iosize of invp and outvp.
3230 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3231 	 * This value is clipped at 4Kbytes and 1Mbyte.
3232 	 */
3233 	blksize = MAX(holein, holeout);
3234 
3235 	/* Clip len to end at an exact multiple of hole size. */
3236 	if (blksize > 1) {
3237 		rem = *inoffp % blksize;
3238 		if (rem > 0)
3239 			rem = blksize - rem;
3240 		if (len > rem && len - rem > blksize)
3241 			len = savlen = rounddown(len - rem, blksize) + rem;
3242 	}
3243 
3244 	if (blksize <= 1)
3245 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3246 		    outvp->v_mount->mnt_stat.f_iosize);
3247 	if (blksize < 4096)
3248 		blksize = 4096;
3249 	else if (blksize > 1024 * 1024)
3250 		blksize = 1024 * 1024;
3251 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3252 
3253 	/*
3254 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3255 	 * to find holes.  Otherwise, just scan the read block for all 0s
3256 	 * in the inner loop where the data copying is done.
3257 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3258 	 * support holes on the server, but do not support FIOSEEKHOLE.
3259 	 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3260 	 * that this function should return after 1second with a partial
3261 	 * completion.
3262 	 */
3263 	if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3264 		getnanouptime(&endts);
3265 		endts.tv_sec++;
3266 	} else
3267 		timespecclear(&endts);
3268 	holetoeof = eof = false;
3269 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3270 		endoff = 0;			/* To shut up compilers. */
3271 		cantseek = true;
3272 		startoff = *inoffp;
3273 		copylen = len;
3274 
3275 		/*
3276 		 * Find the next data area.  If there is just a hole to EOF,
3277 		 * FIOSEEKDATA should fail with ENXIO.
3278 		 * (I do not know if any file system will report a hole to
3279 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3280 		 *  will fail for those file systems.)
3281 		 *
3282 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3283 		 * the code just falls through to the inner copy loop.
3284 		 */
3285 		error = EINVAL;
3286 		if (holein > 0) {
3287 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3288 			    incred, curthread);
3289 			if (error == ENXIO) {
3290 				startoff = endoff = inva.va_size;
3291 				eof = holetoeof = true;
3292 				error = 0;
3293 			}
3294 		}
3295 		if (error == 0 && !holetoeof) {
3296 			endoff = startoff;
3297 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3298 			    incred, curthread);
3299 			/*
3300 			 * Since invp is unlocked, it may be possible for
3301 			 * another thread to do a truncate(), lseek(), write()
3302 			 * creating a hole at startoff between the above
3303 			 * VOP_IOCTL() calls, if the other thread does not do
3304 			 * rangelocking.
3305 			 * If that happens, startoff == endoff and finding
3306 			 * the hole has failed, so set an error.
3307 			 */
3308 			if (error == 0 && startoff == endoff)
3309 				error = EINVAL; /* Any error. Reset to 0. */
3310 		}
3311 		if (error == 0) {
3312 			if (startoff > *inoffp) {
3313 				/* Found hole before data block. */
3314 				xfer = MIN(startoff - *inoffp, len);
3315 				if (*outoffp < va.va_size) {
3316 					/* Must write 0s to punch hole. */
3317 					xfer2 = MIN(va.va_size - *outoffp,
3318 					    xfer);
3319 					memset(dat, 0, MIN(xfer2, blksize));
3320 					error = vn_write_outvp(outvp, dat,
3321 					    *outoffp, xfer2, blksize, false,
3322 					    holeout > 0, outcred);
3323 				}
3324 
3325 				if (error == 0 && *outoffp + xfer >
3326 				    va.va_size && (xfer == len || holetoeof)) {
3327 					/* Grow output file (hole at end). */
3328 					error = vn_write_outvp(outvp, dat,
3329 					    *outoffp, xfer, blksize, true,
3330 					    false, outcred);
3331 				}
3332 				if (error == 0) {
3333 					*inoffp += xfer;
3334 					*outoffp += xfer;
3335 					len -= xfer;
3336 					if (len < savlen) {
3337 						interrupted = sig_intr();
3338 						if (timespecisset(&endts) &&
3339 						    interrupted == 0) {
3340 							getnanouptime(&curts);
3341 							if (timespeccmp(&curts,
3342 							    &endts, >=))
3343 								interrupted =
3344 								    EINTR;
3345 						}
3346 					}
3347 				}
3348 			}
3349 			copylen = MIN(len, endoff - startoff);
3350 			cantseek = false;
3351 		} else {
3352 			cantseek = true;
3353 			startoff = *inoffp;
3354 			copylen = len;
3355 			error = 0;
3356 		}
3357 
3358 		xfer = blksize;
3359 		if (cantseek) {
3360 			/*
3361 			 * Set first xfer to end at a block boundary, so that
3362 			 * holes are more likely detected in the loop below via
3363 			 * the for all bytes 0 method.
3364 			 */
3365 			xfer -= (*inoffp % blksize);
3366 		}
3367 		/* Loop copying the data block. */
3368 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3369 			if (copylen < xfer)
3370 				xfer = copylen;
3371 			error = vn_lock(invp, LK_SHARED);
3372 			if (error != 0)
3373 				goto out;
3374 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3375 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3376 			    curthread->td_ucred, incred, &aresid,
3377 			    curthread);
3378 			VOP_UNLOCK(invp);
3379 			lastblock = false;
3380 			if (error == 0 && aresid > 0) {
3381 				/* Stop the copy at EOF on the input file. */
3382 				xfer -= aresid;
3383 				eof = true;
3384 				lastblock = true;
3385 			}
3386 			if (error == 0) {
3387 				/*
3388 				 * Skip the write for holes past the initial EOF
3389 				 * of the output file, unless this is the last
3390 				 * write of the output file at EOF.
3391 				 */
3392 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3393 				    false;
3394 				if (xfer == len)
3395 					lastblock = true;
3396 				if (!cantseek || *outoffp < va.va_size ||
3397 				    lastblock || !readzeros)
3398 					error = vn_write_outvp(outvp, dat,
3399 					    *outoffp, xfer, blksize,
3400 					    readzeros && lastblock &&
3401 					    *outoffp >= va.va_size, false,
3402 					    outcred);
3403 				if (error == 0) {
3404 					*inoffp += xfer;
3405 					startoff += xfer;
3406 					*outoffp += xfer;
3407 					copylen -= xfer;
3408 					len -= xfer;
3409 					if (len < savlen) {
3410 						interrupted = sig_intr();
3411 						if (timespecisset(&endts) &&
3412 						    interrupted == 0) {
3413 							getnanouptime(&curts);
3414 							if (timespeccmp(&curts,
3415 							    &endts, >=))
3416 								interrupted =
3417 								    EINTR;
3418 						}
3419 					}
3420 				}
3421 			}
3422 			xfer = blksize;
3423 		}
3424 	}
3425 out:
3426 	*lenp = savlen - len;
3427 	free(dat, M_TEMP);
3428 	return (error);
3429 }
3430 
3431 static int
3432 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3433 {
3434 	struct mount *mp;
3435 	struct vnode *vp;
3436 	off_t olen, ooffset;
3437 	int error;
3438 #ifdef AUDIT
3439 	int audited_vnode1 = 0;
3440 #endif
3441 
3442 	vp = fp->f_vnode;
3443 	if (vp->v_type != VREG)
3444 		return (ENODEV);
3445 
3446 	/* Allocating blocks may take a long time, so iterate. */
3447 	for (;;) {
3448 		olen = len;
3449 		ooffset = offset;
3450 
3451 		bwillwrite();
3452 		mp = NULL;
3453 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3454 		if (error != 0)
3455 			break;
3456 		error = vn_lock(vp, LK_EXCLUSIVE);
3457 		if (error != 0) {
3458 			vn_finished_write(mp);
3459 			break;
3460 		}
3461 #ifdef AUDIT
3462 		if (!audited_vnode1) {
3463 			AUDIT_ARG_VNODE1(vp);
3464 			audited_vnode1 = 1;
3465 		}
3466 #endif
3467 #ifdef MAC
3468 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3469 		if (error == 0)
3470 #endif
3471 			error = VOP_ALLOCATE(vp, &offset, &len);
3472 		VOP_UNLOCK(vp);
3473 		vn_finished_write(mp);
3474 
3475 		if (olen + ooffset != offset + len) {
3476 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3477 			    ooffset, olen, offset, len);
3478 		}
3479 		if (error != 0 || len == 0)
3480 			break;
3481 		KASSERT(olen > len, ("Iteration did not make progress?"));
3482 		maybe_yield();
3483 	}
3484 
3485 	return (error);
3486 }
3487 
3488 static int
3489 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3490     int ioflag, struct ucred *cred, struct ucred *active_cred,
3491     struct ucred *file_cred)
3492 {
3493 	struct mount *mp;
3494 	void *rl_cookie;
3495 	off_t off, len;
3496 	int error;
3497 #ifdef AUDIT
3498 	bool audited_vnode1 = false;
3499 #endif
3500 
3501 	rl_cookie = NULL;
3502 	error = 0;
3503 	mp = NULL;
3504 	off = *offset;
3505 	len = *length;
3506 
3507 	if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3508 		rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3509 	while (len > 0 && error == 0) {
3510 		/*
3511 		 * Try to deallocate the longest range in one pass.
3512 		 * In case a pass takes too long to be executed, it returns
3513 		 * partial result. The residue will be proceeded in the next
3514 		 * pass.
3515 		 */
3516 
3517 		if ((ioflag & IO_NODELOCKED) == 0) {
3518 			bwillwrite();
3519 			if ((error = vn_start_write(vp, &mp,
3520 			    V_WAIT | PCATCH)) != 0)
3521 				goto out;
3522 			vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3523 		}
3524 #ifdef AUDIT
3525 		if (!audited_vnode1) {
3526 			AUDIT_ARG_VNODE1(vp);
3527 			audited_vnode1 = true;
3528 		}
3529 #endif
3530 
3531 #ifdef MAC
3532 		if ((ioflag & IO_NOMACCHECK) == 0)
3533 			error = mac_vnode_check_write(active_cred, file_cred,
3534 			    vp);
3535 #endif
3536 		if (error == 0)
3537 			error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3538 			    cred);
3539 
3540 		if ((ioflag & IO_NODELOCKED) == 0) {
3541 			VOP_UNLOCK(vp);
3542 			if (mp != NULL) {
3543 				vn_finished_write(mp);
3544 				mp = NULL;
3545 			}
3546 		}
3547 		if (error == 0 && len != 0)
3548 			maybe_yield();
3549 	}
3550 out:
3551 	if (rl_cookie != NULL)
3552 		vn_rangelock_unlock(vp, rl_cookie);
3553 	*offset = off;
3554 	*length = len;
3555 	return (error);
3556 }
3557 
3558 /*
3559  * This function is supposed to be used in the situations where the deallocation
3560  * is not triggered by a user request.
3561  */
3562 int
3563 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3564     int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3565 {
3566 	struct ucred *cred;
3567 
3568 	if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3569 	    flags != 0)
3570 		return (EINVAL);
3571 	if (vp->v_type != VREG)
3572 		return (ENODEV);
3573 
3574 	cred = file_cred != NOCRED ? file_cred : active_cred;
3575 	return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3576 	    active_cred, file_cred));
3577 }
3578 
3579 static int
3580 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3581     struct ucred *active_cred, struct thread *td)
3582 {
3583 	int error;
3584 	struct vnode *vp;
3585 	int ioflag;
3586 
3587 	vp = fp->f_vnode;
3588 
3589 	if (cmd != SPACECTL_DEALLOC || *offset < 0 || *length <= 0 ||
3590 	    *length > OFF_MAX - *offset || flags != 0)
3591 		return (EINVAL);
3592 	if (vp->v_type != VREG)
3593 		return (ENODEV);
3594 
3595 	ioflag = get_write_ioflag(fp);
3596 
3597 	switch (cmd) {
3598 	case SPACECTL_DEALLOC:
3599 		error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3600 		    active_cred, active_cred, fp->f_cred);
3601 		break;
3602 	default:
3603 		panic("vn_fspacectl: unknown cmd %d", cmd);
3604 	}
3605 
3606 	return (error);
3607 }
3608 
3609 static u_long vn_lock_pair_pause_cnt;
3610 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3611     &vn_lock_pair_pause_cnt, 0,
3612     "Count of vn_lock_pair deadlocks");
3613 
3614 u_int vn_lock_pair_pause_max;
3615 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3616     &vn_lock_pair_pause_max, 0,
3617     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3618 
3619 static void
3620 vn_lock_pair_pause(const char *wmesg)
3621 {
3622 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3623 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3624 }
3625 
3626 /*
3627  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3628  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3629  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3630  * can be NULL.
3631  *
3632  * The function returns with both vnodes exclusively locked, and
3633  * guarantees that it does not create lock order reversal with other
3634  * threads during its execution.  Both vnodes could be unlocked
3635  * temporary (and reclaimed).
3636  */
3637 void
3638 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3639     bool vp2_locked)
3640 {
3641 	int error;
3642 
3643 	if (vp1 == NULL && vp2 == NULL)
3644 		return;
3645 	if (vp1 != NULL) {
3646 		if (vp1_locked)
3647 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3648 		else
3649 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3650 	} else {
3651 		vp1_locked = true;
3652 	}
3653 	if (vp2 != NULL) {
3654 		if (vp2_locked)
3655 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3656 		else
3657 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3658 	} else {
3659 		vp2_locked = true;
3660 	}
3661 	if (!vp1_locked && !vp2_locked) {
3662 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3663 		vp1_locked = true;
3664 	}
3665 
3666 	for (;;) {
3667 		if (vp1_locked && vp2_locked)
3668 			break;
3669 		if (vp1_locked && vp2 != NULL) {
3670 			if (vp1 != NULL) {
3671 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3672 				    __FILE__, __LINE__);
3673 				if (error == 0)
3674 					break;
3675 				VOP_UNLOCK(vp1);
3676 				vp1_locked = false;
3677 				vn_lock_pair_pause("vlp1");
3678 			}
3679 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3680 			vp2_locked = true;
3681 		}
3682 		if (vp2_locked && vp1 != NULL) {
3683 			if (vp2 != NULL) {
3684 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3685 				    __FILE__, __LINE__);
3686 				if (error == 0)
3687 					break;
3688 				VOP_UNLOCK(vp2);
3689 				vp2_locked = false;
3690 				vn_lock_pair_pause("vlp2");
3691 			}
3692 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3693 			vp1_locked = true;
3694 		}
3695 	}
3696 	if (vp1 != NULL)
3697 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3698 	if (vp2 != NULL)
3699 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3700 }
3701 
3702 int
3703 vn_lktype_write(struct mount *mp, struct vnode *vp)
3704 {
3705 	if (MNT_SHARED_WRITES(mp) ||
3706 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3707 		return (LK_SHARED);
3708 	return (LK_EXCLUSIVE);
3709 }
3710