xref: /freebsd/sys/kern/vfs_vnops.c (revision 7a8d62cd046942ab34fcf282ecbbb0747a08a4ee)
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 = curthread;
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 	if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
217 		res |= WANTIOCTLCAPS;
218 	return (res);
219 }
220 
221 /*
222  * Common code for vnode open operations via a name lookup.
223  * Lookup the vnode and invoke VOP_CREATE if needed.
224  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
225  *
226  * Note that this does NOT free nameidata for the successful case,
227  * due to the NDINIT being done elsewhere.
228  */
229 int
230 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
231     struct ucred *cred, struct file *fp)
232 {
233 	struct vnode *vp;
234 	struct mount *mp;
235 	struct vattr vat;
236 	struct vattr *vap = &vat;
237 	int fmode, error;
238 	bool first_open;
239 
240 restart:
241 	first_open = false;
242 	fmode = *flagp;
243 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
244 	    O_EXCL | O_DIRECTORY) ||
245 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
246 		return (EINVAL);
247 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
248 		ndp->ni_cnd.cn_nameiop = CREATE;
249 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
250 		/*
251 		 * Set NOCACHE to avoid flushing the cache when
252 		 * rolling in many files at once.
253 		 *
254 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
255 		 * exist despite NOCACHE.
256 		 */
257 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
258 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
259 			ndp->ni_cnd.cn_flags |= FOLLOW;
260 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
261 			bwillwrite();
262 		if ((error = namei(ndp)) != 0)
263 			return (error);
264 		if (ndp->ni_vp == NULL) {
265 			VATTR_NULL(vap);
266 			vap->va_type = VREG;
267 			vap->va_mode = cmode;
268 			if (fmode & O_EXCL)
269 				vap->va_vaflags |= VA_EXCLUSIVE;
270 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
271 				NDFREE_PNBUF(ndp);
272 				vput(ndp->ni_dvp);
273 				if ((error = vn_start_write(NULL, &mp,
274 				    V_XSLEEP | V_PCATCH)) != 0)
275 					return (error);
276 				NDREINIT(ndp);
277 				goto restart;
278 			}
279 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
280 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
281 #ifdef MAC
282 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
283 			    &ndp->ni_cnd, vap);
284 			if (error == 0)
285 #endif
286 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
287 				    &ndp->ni_cnd, vap);
288 			vp = ndp->ni_vp;
289 			if (error == 0 && (fmode & O_EXCL) != 0 &&
290 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
291 				VI_LOCK(vp);
292 				vp->v_iflag |= VI_FOPENING;
293 				VI_UNLOCK(vp);
294 				first_open = true;
295 			}
296 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
297 			    false);
298 			vn_finished_write(mp);
299 			if (error) {
300 				NDFREE_PNBUF(ndp);
301 				if (error == ERELOOKUP) {
302 					NDREINIT(ndp);
303 					goto restart;
304 				}
305 				return (error);
306 			}
307 			fmode &= ~O_TRUNC;
308 		} else {
309 			if (ndp->ni_dvp == ndp->ni_vp)
310 				vrele(ndp->ni_dvp);
311 			else
312 				vput(ndp->ni_dvp);
313 			ndp->ni_dvp = NULL;
314 			vp = ndp->ni_vp;
315 			if (fmode & O_EXCL) {
316 				error = EEXIST;
317 				goto bad;
318 			}
319 			if (vp->v_type == VDIR) {
320 				error = EISDIR;
321 				goto bad;
322 			}
323 			fmode &= ~O_CREAT;
324 		}
325 	} else {
326 		ndp->ni_cnd.cn_nameiop = LOOKUP;
327 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
328 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
329 		    FOLLOW;
330 		if ((fmode & FWRITE) == 0)
331 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
332 		if ((error = namei(ndp)) != 0)
333 			return (error);
334 		vp = ndp->ni_vp;
335 	}
336 	error = vn_open_vnode(vp, fmode, cred, curthread, fp);
337 	if (first_open) {
338 		VI_LOCK(vp);
339 		vp->v_iflag &= ~VI_FOPENING;
340 		wakeup(vp);
341 		VI_UNLOCK(vp);
342 	}
343 	if (error)
344 		goto bad;
345 	*flagp = fmode;
346 	return (0);
347 bad:
348 	NDFREE_PNBUF(ndp);
349 	vput(vp);
350 	*flagp = fmode;
351 	ndp->ni_vp = NULL;
352 	return (error);
353 }
354 
355 static int
356 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
357 {
358 	struct flock lf;
359 	int error, lock_flags, type;
360 
361 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
362 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
363 		return (0);
364 	KASSERT(fp != NULL, ("open with flock requires fp"));
365 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
366 		return (EOPNOTSUPP);
367 
368 	lock_flags = VOP_ISLOCKED(vp);
369 	VOP_UNLOCK(vp);
370 
371 	lf.l_whence = SEEK_SET;
372 	lf.l_start = 0;
373 	lf.l_len = 0;
374 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
375 	type = F_FLOCK;
376 	if ((fmode & FNONBLOCK) == 0)
377 		type |= F_WAIT;
378 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
379 		type |= F_FIRSTOPEN;
380 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
381 	if (error == 0)
382 		fp->f_flag |= FHASLOCK;
383 
384 	vn_lock(vp, lock_flags | LK_RETRY);
385 	return (error);
386 }
387 
388 /*
389  * Common code for vnode open operations once a vnode is located.
390  * Check permissions, and call the VOP_OPEN routine.
391  */
392 int
393 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
394     struct thread *td, struct file *fp)
395 {
396 	accmode_t accmode;
397 	int error;
398 
399 	if (vp->v_type == VLNK) {
400 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
401 			return (EMLINK);
402 	}
403 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
404 		return (ENOTDIR);
405 
406 	accmode = 0;
407 	if ((fmode & O_PATH) == 0) {
408 		if (vp->v_type == VSOCK)
409 			return (EOPNOTSUPP);
410 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
411 			if (vp->v_type == VDIR)
412 				return (EISDIR);
413 			accmode |= VWRITE;
414 		}
415 		if ((fmode & FREAD) != 0)
416 			accmode |= VREAD;
417 		if ((fmode & O_APPEND) && (fmode & FWRITE))
418 			accmode |= VAPPEND;
419 #ifdef MAC
420 		if ((fmode & O_CREAT) != 0)
421 			accmode |= VCREAT;
422 #endif
423 	}
424 	if ((fmode & FEXEC) != 0)
425 		accmode |= VEXEC;
426 #ifdef MAC
427 	if ((fmode & O_VERIFY) != 0)
428 		accmode |= VVERIFY;
429 	error = mac_vnode_check_open(cred, vp, accmode);
430 	if (error != 0)
431 		return (error);
432 
433 	accmode &= ~(VCREAT | VVERIFY);
434 #endif
435 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
436 		error = VOP_ACCESS(vp, accmode, cred, td);
437 		if (error != 0)
438 			return (error);
439 	}
440 	if ((fmode & O_PATH) != 0) {
441 		if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
442 		    VOP_ACCESS(vp, VREAD, cred, td) == 0)
443 			fp->f_flag |= FKQALLOWED;
444 		return (0);
445 	}
446 
447 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
448 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
449 	error = VOP_OPEN(vp, fmode, cred, td, fp);
450 	if (error != 0)
451 		return (error);
452 
453 	error = vn_open_vnode_advlock(vp, fmode, fp);
454 	if (error == 0 && (fmode & FWRITE) != 0) {
455 		error = VOP_ADD_WRITECOUNT(vp, 1);
456 		if (error == 0) {
457 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
458 			     __func__, vp, vp->v_writecount);
459 		}
460 	}
461 
462 	/*
463 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
464 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
465 	 */
466 	if (error != 0) {
467 		if (fp != NULL) {
468 			/*
469 			 * Arrange the call by having fdrop() to use
470 			 * vn_closefile().  This is to satisfy
471 			 * filesystems like devfs or tmpfs, which
472 			 * override fo_close().
473 			 */
474 			fp->f_flag |= FOPENFAILED;
475 			fp->f_vnode = vp;
476 			if (fp->f_ops == &badfileops) {
477 				fp->f_type = DTYPE_VNODE;
478 				fp->f_ops = &vnops;
479 			}
480 			vref(vp);
481 		} else {
482 			/*
483 			 * If there is no fp, due to kernel-mode open,
484 			 * we can call VOP_CLOSE() now.
485 			 */
486 			if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
487 			    !MNT_EXTENDED_SHARED(vp->v_mount) &&
488 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
489 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
490 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
491 			    cred, td);
492 		}
493 	}
494 
495 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
496 	return (error);
497 
498 }
499 
500 /*
501  * Check for write permissions on the specified vnode.
502  * Prototype text segments cannot be written.
503  * It is racy.
504  */
505 int
506 vn_writechk(struct vnode *vp)
507 {
508 
509 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
510 	/*
511 	 * If there's shared text associated with
512 	 * the vnode, try to free it up once.  If
513 	 * we fail, we can't allow writing.
514 	 */
515 	if (VOP_IS_TEXT(vp))
516 		return (ETXTBSY);
517 
518 	return (0);
519 }
520 
521 /*
522  * Vnode close call
523  */
524 static int
525 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
526     struct thread *td, bool keep_ref)
527 {
528 	struct mount *mp;
529 	int error, lock_flags;
530 
531 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
532 	    MNT_EXTENDED_SHARED(vp->v_mount))
533 		lock_flags = LK_SHARED;
534 	else
535 		lock_flags = LK_EXCLUSIVE;
536 
537 	vn_start_write(vp, &mp, V_WAIT);
538 	vn_lock(vp, lock_flags | LK_RETRY);
539 	AUDIT_ARG_VNODE1(vp);
540 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
541 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
542 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
543 		    __func__, vp, vp->v_writecount);
544 	}
545 	error = VOP_CLOSE(vp, flags, file_cred, td);
546 	if (keep_ref)
547 		VOP_UNLOCK(vp);
548 	else
549 		vput(vp);
550 	vn_finished_write(mp);
551 	return (error);
552 }
553 
554 int
555 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
556     struct thread *td)
557 {
558 
559 	return (vn_close1(vp, flags, file_cred, td, false));
560 }
561 
562 /*
563  * Heuristic to detect sequential operation.
564  */
565 static int
566 sequential_heuristic(struct uio *uio, struct file *fp)
567 {
568 	enum uio_rw rw;
569 
570 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
571 
572 	rw = uio->uio_rw;
573 	if (fp->f_flag & FRDAHEAD)
574 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
575 
576 	/*
577 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
578 	 * that the first I/O is normally considered to be slightly
579 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
580 	 * unless previous seeks have reduced f_seqcount to 0, in which
581 	 * case offset 0 is not special.
582 	 */
583 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
584 	    uio->uio_offset == fp->f_nextoff[rw]) {
585 		/*
586 		 * f_seqcount is in units of fixed-size blocks so that it
587 		 * depends mainly on the amount of sequential I/O and not
588 		 * much on the number of sequential I/O's.  The fixed size
589 		 * of 16384 is hard-coded here since it is (not quite) just
590 		 * a magic size that works well here.  This size is more
591 		 * closely related to the best I/O size for real disks than
592 		 * to any block size used by software.
593 		 */
594 		if (uio->uio_resid >= IO_SEQMAX * 16384)
595 			fp->f_seqcount[rw] = IO_SEQMAX;
596 		else {
597 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
598 			if (fp->f_seqcount[rw] > IO_SEQMAX)
599 				fp->f_seqcount[rw] = IO_SEQMAX;
600 		}
601 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
602 	}
603 
604 	/* Not sequential.  Quickly draw-down sequentiality. */
605 	if (fp->f_seqcount[rw] > 1)
606 		fp->f_seqcount[rw] = 1;
607 	else
608 		fp->f_seqcount[rw] = 0;
609 	return (0);
610 }
611 
612 /*
613  * Package up an I/O request on a vnode into a uio and do it.
614  */
615 int
616 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
617     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
618     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
619 {
620 	struct uio auio;
621 	struct iovec aiov;
622 	struct mount *mp;
623 	struct ucred *cred;
624 	void *rl_cookie;
625 	struct vn_io_fault_args args;
626 	int error, lock_flags;
627 
628 	if (offset < 0 && vp->v_type != VCHR)
629 		return (EINVAL);
630 	auio.uio_iov = &aiov;
631 	auio.uio_iovcnt = 1;
632 	aiov.iov_base = base;
633 	aiov.iov_len = len;
634 	auio.uio_resid = len;
635 	auio.uio_offset = offset;
636 	auio.uio_segflg = segflg;
637 	auio.uio_rw = rw;
638 	auio.uio_td = td;
639 	error = 0;
640 
641 	if ((ioflg & IO_NODELOCKED) == 0) {
642 		if ((ioflg & IO_RANGELOCKED) == 0) {
643 			if (rw == UIO_READ) {
644 				rl_cookie = vn_rangelock_rlock(vp, offset,
645 				    offset + len);
646 			} else if ((ioflg & IO_APPEND) != 0) {
647 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
648 			} else {
649 				rl_cookie = vn_rangelock_wlock(vp, offset,
650 				    offset + len);
651 			}
652 		} else
653 			rl_cookie = NULL;
654 		mp = NULL;
655 		if (rw == UIO_WRITE) {
656 			if (vp->v_type != VCHR &&
657 			    (error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH))
658 			    != 0)
659 				goto out;
660 			lock_flags = vn_lktype_write(mp, vp);
661 		} else
662 			lock_flags = LK_SHARED;
663 		vn_lock(vp, lock_flags | LK_RETRY);
664 	} else
665 		rl_cookie = NULL;
666 
667 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
668 #ifdef MAC
669 	if ((ioflg & IO_NOMACCHECK) == 0) {
670 		if (rw == UIO_READ)
671 			error = mac_vnode_check_read(active_cred, file_cred,
672 			    vp);
673 		else
674 			error = mac_vnode_check_write(active_cred, file_cred,
675 			    vp);
676 	}
677 #endif
678 	if (error == 0) {
679 		if (file_cred != NULL)
680 			cred = file_cred;
681 		else
682 			cred = active_cred;
683 		if (do_vn_io_fault(vp, &auio)) {
684 			args.kind = VN_IO_FAULT_VOP;
685 			args.cred = cred;
686 			args.flags = ioflg;
687 			args.args.vop_args.vp = vp;
688 			error = vn_io_fault1(vp, &auio, &args, td);
689 		} else if (rw == UIO_READ) {
690 			error = VOP_READ(vp, &auio, ioflg, cred);
691 		} else /* if (rw == UIO_WRITE) */ {
692 			error = VOP_WRITE(vp, &auio, ioflg, cred);
693 		}
694 	}
695 	if (aresid)
696 		*aresid = auio.uio_resid;
697 	else
698 		if (auio.uio_resid && error == 0)
699 			error = EIO;
700 	if ((ioflg & IO_NODELOCKED) == 0) {
701 		VOP_UNLOCK(vp);
702 		if (mp != NULL)
703 			vn_finished_write(mp);
704 	}
705  out:
706 	if (rl_cookie != NULL)
707 		vn_rangelock_unlock(vp, rl_cookie);
708 	return (error);
709 }
710 
711 /*
712  * Package up an I/O request on a vnode into a uio and do it.  The I/O
713  * request is split up into smaller chunks and we try to avoid saturating
714  * the buffer cache while potentially holding a vnode locked, so we
715  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
716  * to give other processes a chance to lock the vnode (either other processes
717  * core'ing the same binary, or unrelated processes scanning the directory).
718  */
719 int
720 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
721     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
722     struct ucred *file_cred, size_t *aresid, struct thread *td)
723 {
724 	int error = 0;
725 	ssize_t iaresid;
726 
727 	do {
728 		int chunk;
729 
730 		/*
731 		 * Force `offset' to a multiple of MAXBSIZE except possibly
732 		 * for the first chunk, so that filesystems only need to
733 		 * write full blocks except possibly for the first and last
734 		 * chunks.
735 		 */
736 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
737 
738 		if (chunk > len)
739 			chunk = len;
740 		if (rw != UIO_READ && vp->v_type == VREG)
741 			bwillwrite();
742 		iaresid = 0;
743 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
744 		    ioflg, active_cred, file_cred, &iaresid, td);
745 		len -= chunk;	/* aresid calc already includes length */
746 		if (error)
747 			break;
748 		offset += chunk;
749 		base = (char *)base + chunk;
750 		kern_yield(PRI_USER);
751 	} while (len);
752 	if (aresid)
753 		*aresid = len + iaresid;
754 	return (error);
755 }
756 
757 #if OFF_MAX <= LONG_MAX
758 off_t
759 foffset_lock(struct file *fp, int flags)
760 {
761 	volatile short *flagsp;
762 	off_t res;
763 	short state;
764 
765 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
766 
767 	if ((flags & FOF_NOLOCK) != 0)
768 		return (atomic_load_long(&fp->f_offset));
769 
770 	/*
771 	 * According to McKusick the vn lock was protecting f_offset here.
772 	 * It is now protected by the FOFFSET_LOCKED flag.
773 	 */
774 	flagsp = &fp->f_vnread_flags;
775 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
776 		return (atomic_load_long(&fp->f_offset));
777 
778 	sleepq_lock(&fp->f_vnread_flags);
779 	state = atomic_load_16(flagsp);
780 	for (;;) {
781 		if ((state & FOFFSET_LOCKED) == 0) {
782 			if (!atomic_fcmpset_acq_16(flagsp, &state,
783 			    FOFFSET_LOCKED))
784 				continue;
785 			break;
786 		}
787 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
788 			if (!atomic_fcmpset_acq_16(flagsp, &state,
789 			    state | FOFFSET_LOCK_WAITING))
790 				continue;
791 		}
792 		DROP_GIANT();
793 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
794 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
795 		PICKUP_GIANT();
796 		sleepq_lock(&fp->f_vnread_flags);
797 		state = atomic_load_16(flagsp);
798 	}
799 	res = atomic_load_long(&fp->f_offset);
800 	sleepq_release(&fp->f_vnread_flags);
801 	return (res);
802 }
803 
804 void
805 foffset_unlock(struct file *fp, off_t val, int flags)
806 {
807 	volatile short *flagsp;
808 	short state;
809 
810 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
811 
812 	if ((flags & FOF_NOUPDATE) == 0)
813 		atomic_store_long(&fp->f_offset, val);
814 	if ((flags & FOF_NEXTOFF_R) != 0)
815 		fp->f_nextoff[UIO_READ] = val;
816 	if ((flags & FOF_NEXTOFF_W) != 0)
817 		fp->f_nextoff[UIO_WRITE] = val;
818 
819 	if ((flags & FOF_NOLOCK) != 0)
820 		return;
821 
822 	flagsp = &fp->f_vnread_flags;
823 	state = atomic_load_16(flagsp);
824 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
825 	    atomic_cmpset_rel_16(flagsp, state, 0))
826 		return;
827 
828 	sleepq_lock(&fp->f_vnread_flags);
829 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
830 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
831 	fp->f_vnread_flags = 0;
832 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
833 	sleepq_release(&fp->f_vnread_flags);
834 }
835 #else
836 off_t
837 foffset_lock(struct file *fp, int flags)
838 {
839 	struct mtx *mtxp;
840 	off_t res;
841 
842 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
843 
844 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
845 	mtx_lock(mtxp);
846 	if ((flags & FOF_NOLOCK) == 0) {
847 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
848 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
849 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
850 			    "vofflock", 0);
851 		}
852 		fp->f_vnread_flags |= FOFFSET_LOCKED;
853 	}
854 	res = fp->f_offset;
855 	mtx_unlock(mtxp);
856 	return (res);
857 }
858 
859 void
860 foffset_unlock(struct file *fp, off_t val, int flags)
861 {
862 	struct mtx *mtxp;
863 
864 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
865 
866 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
867 	mtx_lock(mtxp);
868 	if ((flags & FOF_NOUPDATE) == 0)
869 		fp->f_offset = val;
870 	if ((flags & FOF_NEXTOFF_R) != 0)
871 		fp->f_nextoff[UIO_READ] = val;
872 	if ((flags & FOF_NEXTOFF_W) != 0)
873 		fp->f_nextoff[UIO_WRITE] = val;
874 	if ((flags & FOF_NOLOCK) == 0) {
875 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
876 		    ("Lost FOFFSET_LOCKED"));
877 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
878 			wakeup(&fp->f_vnread_flags);
879 		fp->f_vnread_flags = 0;
880 	}
881 	mtx_unlock(mtxp);
882 }
883 #endif
884 
885 void
886 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
887 {
888 
889 	if ((flags & FOF_OFFSET) == 0)
890 		uio->uio_offset = foffset_lock(fp, flags);
891 }
892 
893 void
894 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
895 {
896 
897 	if ((flags & FOF_OFFSET) == 0)
898 		foffset_unlock(fp, uio->uio_offset, flags);
899 }
900 
901 static int
902 get_advice(struct file *fp, struct uio *uio)
903 {
904 	struct mtx *mtxp;
905 	int ret;
906 
907 	ret = POSIX_FADV_NORMAL;
908 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
909 		return (ret);
910 
911 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
912 	mtx_lock(mtxp);
913 	if (fp->f_advice != NULL &&
914 	    uio->uio_offset >= fp->f_advice->fa_start &&
915 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
916 		ret = fp->f_advice->fa_advice;
917 	mtx_unlock(mtxp);
918 	return (ret);
919 }
920 
921 static int
922 get_write_ioflag(struct file *fp)
923 {
924 	int ioflag;
925 	struct mount *mp;
926 	struct vnode *vp;
927 
928 	ioflag = 0;
929 	vp = fp->f_vnode;
930 	mp = atomic_load_ptr(&vp->v_mount);
931 
932 	if ((fp->f_flag & O_DIRECT) != 0)
933 		ioflag |= IO_DIRECT;
934 
935 	if ((fp->f_flag & O_FSYNC) != 0 ||
936 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
937 		ioflag |= IO_SYNC;
938 
939 	/*
940 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
941 	 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
942 	 * fall back to full O_SYNC behavior.
943 	 */
944 	if ((fp->f_flag & O_DSYNC) != 0)
945 		ioflag |= IO_SYNC | IO_DATASYNC;
946 
947 	return (ioflag);
948 }
949 
950 int
951 vn_read_from_obj(struct vnode *vp, struct uio *uio)
952 {
953 	vm_object_t obj;
954 	vm_page_t ma[io_hold_cnt + 2];
955 	off_t off, vsz;
956 	ssize_t resid;
957 	int error, i, j;
958 
959 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
960 	obj = atomic_load_ptr(&vp->v_object);
961 	if (obj == NULL)
962 		return (EJUSTRETURN);
963 
964 	/*
965 	 * Depends on type stability of vm_objects.
966 	 */
967 	vm_object_pip_add(obj, 1);
968 	if ((obj->flags & OBJ_DEAD) != 0) {
969 		/*
970 		 * Note that object might be already reused from the
971 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
972 		 * we recheck for DOOMED vnode state after all pages
973 		 * are busied, and retract then.
974 		 *
975 		 * But we check for OBJ_DEAD to ensure that we do not
976 		 * busy pages while vm_object_terminate_pages()
977 		 * processes the queue.
978 		 */
979 		error = EJUSTRETURN;
980 		goto out_pip;
981 	}
982 
983 	resid = uio->uio_resid;
984 	off = uio->uio_offset;
985 	for (i = 0; resid > 0; i++) {
986 		MPASS(i < io_hold_cnt + 2);
987 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
988 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
989 		    VM_ALLOC_NOWAIT);
990 		if (ma[i] == NULL)
991 			break;
992 
993 		/*
994 		 * Skip invalid pages.  Valid mask can be partial only
995 		 * at EOF, and we clip later.
996 		 */
997 		if (vm_page_none_valid(ma[i])) {
998 			vm_page_sunbusy(ma[i]);
999 			break;
1000 		}
1001 
1002 		resid -= PAGE_SIZE;
1003 		off += PAGE_SIZE;
1004 	}
1005 	if (i == 0) {
1006 		error = EJUSTRETURN;
1007 		goto out_pip;
1008 	}
1009 
1010 	/*
1011 	 * Check VIRF_DOOMED after we busied our pages.  Since
1012 	 * vgonel() terminates the vnode' vm_object, it cannot
1013 	 * process past pages busied by us.
1014 	 */
1015 	if (VN_IS_DOOMED(vp)) {
1016 		error = EJUSTRETURN;
1017 		goto out;
1018 	}
1019 
1020 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1021 	if (resid > uio->uio_resid)
1022 		resid = uio->uio_resid;
1023 
1024 	/*
1025 	 * Unlocked read of vnp_size is safe because truncation cannot
1026 	 * pass busied page.  But we load vnp_size into a local
1027 	 * variable so that possible concurrent extension does not
1028 	 * break calculation.
1029 	 */
1030 #if defined(__powerpc__) && !defined(__powerpc64__)
1031 	vsz = obj->un_pager.vnp.vnp_size;
1032 #else
1033 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1034 #endif
1035 	if (uio->uio_offset >= vsz) {
1036 		error = EJUSTRETURN;
1037 		goto out;
1038 	}
1039 	if (uio->uio_offset + resid > vsz)
1040 		resid = vsz - uio->uio_offset;
1041 
1042 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1043 
1044 out:
1045 	for (j = 0; j < i; j++) {
1046 		if (error == 0)
1047 			vm_page_reference(ma[j]);
1048 		vm_page_sunbusy(ma[j]);
1049 	}
1050 out_pip:
1051 	vm_object_pip_wakeup(obj);
1052 	if (error != 0)
1053 		return (error);
1054 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1055 }
1056 
1057 /*
1058  * File table vnode read routine.
1059  */
1060 static int
1061 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1062     struct thread *td)
1063 {
1064 	struct vnode *vp;
1065 	off_t orig_offset;
1066 	int error, ioflag;
1067 	int advice;
1068 
1069 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1070 	    uio->uio_td, td));
1071 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1072 	vp = fp->f_vnode;
1073 	ioflag = 0;
1074 	if (fp->f_flag & FNONBLOCK)
1075 		ioflag |= IO_NDELAY;
1076 	if (fp->f_flag & O_DIRECT)
1077 		ioflag |= IO_DIRECT;
1078 
1079 	/*
1080 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1081 	 * allows us to avoid unneeded work outright.
1082 	 */
1083 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1084 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1085 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1086 		if (error == 0) {
1087 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1088 			return (0);
1089 		}
1090 		if (error != EJUSTRETURN)
1091 			return (error);
1092 	}
1093 
1094 	advice = get_advice(fp, uio);
1095 	vn_lock(vp, LK_SHARED | LK_RETRY);
1096 
1097 	switch (advice) {
1098 	case POSIX_FADV_NORMAL:
1099 	case POSIX_FADV_SEQUENTIAL:
1100 	case POSIX_FADV_NOREUSE:
1101 		ioflag |= sequential_heuristic(uio, fp);
1102 		break;
1103 	case POSIX_FADV_RANDOM:
1104 		/* Disable read-ahead for random I/O. */
1105 		break;
1106 	}
1107 	orig_offset = uio->uio_offset;
1108 
1109 #ifdef MAC
1110 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1111 	if (error == 0)
1112 #endif
1113 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1114 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1115 	VOP_UNLOCK(vp);
1116 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1117 	    orig_offset != uio->uio_offset)
1118 		/*
1119 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1120 		 * for the backing file after a POSIX_FADV_NOREUSE
1121 		 * read(2).
1122 		 */
1123 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1124 		    POSIX_FADV_DONTNEED);
1125 	return (error);
1126 }
1127 
1128 /*
1129  * File table vnode write routine.
1130  */
1131 static int
1132 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1133     struct thread *td)
1134 {
1135 	struct vnode *vp;
1136 	struct mount *mp;
1137 	off_t orig_offset;
1138 	int error, ioflag;
1139 	int advice;
1140 	bool need_finished_write;
1141 
1142 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1143 	    uio->uio_td, td));
1144 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1145 	vp = fp->f_vnode;
1146 	if (vp->v_type == VREG)
1147 		bwillwrite();
1148 	ioflag = IO_UNIT;
1149 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1150 		ioflag |= IO_APPEND;
1151 	if ((fp->f_flag & FNONBLOCK) != 0)
1152 		ioflag |= IO_NDELAY;
1153 	ioflag |= get_write_ioflag(fp);
1154 
1155 	mp = NULL;
1156 	need_finished_write = false;
1157 	if (vp->v_type != VCHR) {
1158 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1159 		if (error != 0)
1160 			goto unlock;
1161 		need_finished_write = true;
1162 	}
1163 
1164 	advice = get_advice(fp, uio);
1165 
1166 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1167 	switch (advice) {
1168 	case POSIX_FADV_NORMAL:
1169 	case POSIX_FADV_SEQUENTIAL:
1170 	case POSIX_FADV_NOREUSE:
1171 		ioflag |= sequential_heuristic(uio, fp);
1172 		break;
1173 	case POSIX_FADV_RANDOM:
1174 		/* XXX: Is this correct? */
1175 		break;
1176 	}
1177 	orig_offset = uio->uio_offset;
1178 
1179 #ifdef MAC
1180 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1181 	if (error == 0)
1182 #endif
1183 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1184 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1185 	VOP_UNLOCK(vp);
1186 	if (need_finished_write)
1187 		vn_finished_write(mp);
1188 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1189 	    orig_offset != uio->uio_offset)
1190 		/*
1191 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1192 		 * for the backing file after a POSIX_FADV_NOREUSE
1193 		 * write(2).
1194 		 */
1195 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1196 		    POSIX_FADV_DONTNEED);
1197 unlock:
1198 	return (error);
1199 }
1200 
1201 /*
1202  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1203  * prevent the following deadlock:
1204  *
1205  * Assume that the thread A reads from the vnode vp1 into userspace
1206  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1207  * currently not resident, then system ends up with the call chain
1208  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1209  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1210  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1211  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1212  * backed by the pages of vnode vp1, and some page in buf2 is not
1213  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1214  *
1215  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1216  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1217  * Instead, it first tries to do the whole range i/o with pagefaults
1218  * disabled. If all pages in the i/o buffer are resident and mapped,
1219  * VOP will succeed (ignoring the genuine filesystem errors).
1220  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1221  * i/o in chunks, with all pages in the chunk prefaulted and held
1222  * using vm_fault_quick_hold_pages().
1223  *
1224  * Filesystems using this deadlock avoidance scheme should use the
1225  * array of the held pages from uio, saved in the curthread->td_ma,
1226  * instead of doing uiomove().  A helper function
1227  * vn_io_fault_uiomove() converts uiomove request into
1228  * uiomove_fromphys() over td_ma array.
1229  *
1230  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1231  * make the current i/o request atomic with respect to other i/os and
1232  * truncations.
1233  */
1234 
1235 /*
1236  * Decode vn_io_fault_args and perform the corresponding i/o.
1237  */
1238 static int
1239 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1240     struct thread *td)
1241 {
1242 	int error, save;
1243 
1244 	error = 0;
1245 	save = vm_fault_disable_pagefaults();
1246 	switch (args->kind) {
1247 	case VN_IO_FAULT_FOP:
1248 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1249 		    uio, args->cred, args->flags, td);
1250 		break;
1251 	case VN_IO_FAULT_VOP:
1252 		if (uio->uio_rw == UIO_READ) {
1253 			error = VOP_READ(args->args.vop_args.vp, uio,
1254 			    args->flags, args->cred);
1255 		} else if (uio->uio_rw == UIO_WRITE) {
1256 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1257 			    args->flags, args->cred);
1258 		}
1259 		break;
1260 	default:
1261 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1262 		    args->kind, uio->uio_rw);
1263 	}
1264 	vm_fault_enable_pagefaults(save);
1265 	return (error);
1266 }
1267 
1268 static int
1269 vn_io_fault_touch(char *base, const struct uio *uio)
1270 {
1271 	int r;
1272 
1273 	r = fubyte(base);
1274 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1275 		return (EFAULT);
1276 	return (0);
1277 }
1278 
1279 static int
1280 vn_io_fault_prefault_user(const struct uio *uio)
1281 {
1282 	char *base;
1283 	const struct iovec *iov;
1284 	size_t len;
1285 	ssize_t resid;
1286 	int error, i;
1287 
1288 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1289 	    ("vn_io_fault_prefault userspace"));
1290 
1291 	error = i = 0;
1292 	iov = uio->uio_iov;
1293 	resid = uio->uio_resid;
1294 	base = iov->iov_base;
1295 	len = iov->iov_len;
1296 	while (resid > 0) {
1297 		error = vn_io_fault_touch(base, uio);
1298 		if (error != 0)
1299 			break;
1300 		if (len < PAGE_SIZE) {
1301 			if (len != 0) {
1302 				error = vn_io_fault_touch(base + len - 1, uio);
1303 				if (error != 0)
1304 					break;
1305 				resid -= len;
1306 			}
1307 			if (++i >= uio->uio_iovcnt)
1308 				break;
1309 			iov = uio->uio_iov + i;
1310 			base = iov->iov_base;
1311 			len = iov->iov_len;
1312 		} else {
1313 			len -= PAGE_SIZE;
1314 			base += PAGE_SIZE;
1315 			resid -= PAGE_SIZE;
1316 		}
1317 	}
1318 	return (error);
1319 }
1320 
1321 /*
1322  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1323  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1324  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1325  * into args and call vn_io_fault1() to handle faults during the user
1326  * mode buffer accesses.
1327  */
1328 static int
1329 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1330     struct thread *td)
1331 {
1332 	vm_page_t ma[io_hold_cnt + 2];
1333 	struct uio *uio_clone, short_uio;
1334 	struct iovec short_iovec[1];
1335 	vm_page_t *prev_td_ma;
1336 	vm_prot_t prot;
1337 	vm_offset_t addr, end;
1338 	size_t len, resid;
1339 	ssize_t adv;
1340 	int error, cnt, saveheld, prev_td_ma_cnt;
1341 
1342 	if (vn_io_fault_prefault) {
1343 		error = vn_io_fault_prefault_user(uio);
1344 		if (error != 0)
1345 			return (error); /* Or ignore ? */
1346 	}
1347 
1348 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1349 
1350 	/*
1351 	 * The UFS follows IO_UNIT directive and replays back both
1352 	 * uio_offset and uio_resid if an error is encountered during the
1353 	 * operation.  But, since the iovec may be already advanced,
1354 	 * uio is still in an inconsistent state.
1355 	 *
1356 	 * Cache a copy of the original uio, which is advanced to the redo
1357 	 * point using UIO_NOCOPY below.
1358 	 */
1359 	uio_clone = cloneuio(uio);
1360 	resid = uio->uio_resid;
1361 
1362 	short_uio.uio_segflg = UIO_USERSPACE;
1363 	short_uio.uio_rw = uio->uio_rw;
1364 	short_uio.uio_td = uio->uio_td;
1365 
1366 	error = vn_io_fault_doio(args, uio, td);
1367 	if (error != EFAULT)
1368 		goto out;
1369 
1370 	atomic_add_long(&vn_io_faults_cnt, 1);
1371 	uio_clone->uio_segflg = UIO_NOCOPY;
1372 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1373 	uio_clone->uio_segflg = uio->uio_segflg;
1374 
1375 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1376 	prev_td_ma = td->td_ma;
1377 	prev_td_ma_cnt = td->td_ma_cnt;
1378 
1379 	while (uio_clone->uio_resid != 0) {
1380 		len = uio_clone->uio_iov->iov_len;
1381 		if (len == 0) {
1382 			KASSERT(uio_clone->uio_iovcnt >= 1,
1383 			    ("iovcnt underflow"));
1384 			uio_clone->uio_iov++;
1385 			uio_clone->uio_iovcnt--;
1386 			continue;
1387 		}
1388 		if (len > ptoa(io_hold_cnt))
1389 			len = ptoa(io_hold_cnt);
1390 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1391 		end = round_page(addr + len);
1392 		if (end < addr) {
1393 			error = EFAULT;
1394 			break;
1395 		}
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 | V_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 {
1676 	struct vnode *vp = fp->f_vnode;
1677 	int error;
1678 
1679 	vn_lock(vp, LK_SHARED | LK_RETRY);
1680 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1681 	VOP_UNLOCK(vp);
1682 
1683 	return (error);
1684 }
1685 
1686 /*
1687  * File table vnode ioctl routine.
1688  */
1689 static int
1690 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1691     struct thread *td)
1692 {
1693 	struct vattr vattr;
1694 	struct vnode *vp;
1695 	struct fiobmap2_arg *bmarg;
1696 	int error;
1697 
1698 	vp = fp->f_vnode;
1699 	switch (vp->v_type) {
1700 	case VDIR:
1701 	case VREG:
1702 		switch (com) {
1703 		case FIONREAD:
1704 			vn_lock(vp, LK_SHARED | LK_RETRY);
1705 			error = VOP_GETATTR(vp, &vattr, active_cred);
1706 			VOP_UNLOCK(vp);
1707 			if (error == 0)
1708 				*(int *)data = vattr.va_size - fp->f_offset;
1709 			return (error);
1710 		case FIOBMAP2:
1711 			bmarg = (struct fiobmap2_arg *)data;
1712 			vn_lock(vp, LK_SHARED | LK_RETRY);
1713 #ifdef MAC
1714 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1715 			    vp);
1716 			if (error == 0)
1717 #endif
1718 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1719 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1720 			VOP_UNLOCK(vp);
1721 			return (error);
1722 		case FIONBIO:
1723 		case FIOASYNC:
1724 			return (0);
1725 		default:
1726 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1727 			    active_cred, td));
1728 		}
1729 		break;
1730 	case VCHR:
1731 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1732 		    active_cred, td));
1733 	default:
1734 		return (ENOTTY);
1735 	}
1736 }
1737 
1738 /*
1739  * File table vnode poll routine.
1740  */
1741 static int
1742 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1743     struct thread *td)
1744 {
1745 	struct vnode *vp;
1746 	int error;
1747 
1748 	vp = fp->f_vnode;
1749 #if defined(MAC) || defined(AUDIT)
1750 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1751 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1752 		AUDIT_ARG_VNODE1(vp);
1753 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1754 		VOP_UNLOCK(vp);
1755 		if (error != 0)
1756 			return (error);
1757 	}
1758 #endif
1759 	error = VOP_POLL(vp, events, fp->f_cred, td);
1760 	return (error);
1761 }
1762 
1763 /*
1764  * Acquire the requested lock and then check for validity.  LK_RETRY
1765  * permits vn_lock to return doomed vnodes.
1766  */
1767 static int __noinline
1768 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1769     int error)
1770 {
1771 
1772 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1773 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1774 
1775 	if (error == 0)
1776 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1777 
1778 	if ((flags & LK_RETRY) == 0) {
1779 		if (error == 0) {
1780 			VOP_UNLOCK(vp);
1781 			error = ENOENT;
1782 		}
1783 		return (error);
1784 	}
1785 
1786 	/*
1787 	 * LK_RETRY case.
1788 	 *
1789 	 * Nothing to do if we got the lock.
1790 	 */
1791 	if (error == 0)
1792 		return (0);
1793 
1794 	/*
1795 	 * Interlock was dropped by the call in _vn_lock.
1796 	 */
1797 	flags &= ~LK_INTERLOCK;
1798 	do {
1799 		error = VOP_LOCK1(vp, flags, file, line);
1800 	} while (error != 0);
1801 	return (0);
1802 }
1803 
1804 int
1805 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1806 {
1807 	int error;
1808 
1809 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1810 	    ("vn_lock: no locktype (%d passed)", flags));
1811 	VNPASS(vp->v_holdcnt > 0, vp);
1812 	error = VOP_LOCK1(vp, flags, file, line);
1813 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1814 		return (_vn_lock_fallback(vp, flags, file, line, error));
1815 	return (0);
1816 }
1817 
1818 /*
1819  * File table vnode close routine.
1820  */
1821 static int
1822 vn_closefile(struct file *fp, struct thread *td)
1823 {
1824 	struct vnode *vp;
1825 	struct flock lf;
1826 	int error;
1827 	bool ref;
1828 
1829 	vp = fp->f_vnode;
1830 	fp->f_ops = &badfileops;
1831 	ref = (fp->f_flag & FHASLOCK) != 0;
1832 
1833 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1834 
1835 	if (__predict_false(ref)) {
1836 		lf.l_whence = SEEK_SET;
1837 		lf.l_start = 0;
1838 		lf.l_len = 0;
1839 		lf.l_type = F_UNLCK;
1840 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1841 		vrele(vp);
1842 	}
1843 	return (error);
1844 }
1845 
1846 /*
1847  * Preparing to start a filesystem write operation. If the operation is
1848  * permitted, then we bump the count of operations in progress and
1849  * proceed. If a suspend request is in progress, we wait until the
1850  * suspension is over, and then proceed.
1851  */
1852 static int
1853 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1854 {
1855 	struct mount_pcpu *mpcpu;
1856 	int error, mflags;
1857 
1858 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1859 	    vfs_op_thread_enter(mp, mpcpu)) {
1860 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1861 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1862 		vfs_op_thread_exit(mp, mpcpu);
1863 		return (0);
1864 	}
1865 
1866 	if (mplocked)
1867 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1868 	else
1869 		MNT_ILOCK(mp);
1870 
1871 	error = 0;
1872 
1873 	/*
1874 	 * Check on status of suspension.
1875 	 */
1876 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1877 	    mp->mnt_susp_owner != curthread) {
1878 		mflags = 0;
1879 		if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
1880 			if (flags & V_PCATCH)
1881 				mflags |= PCATCH;
1882 		}
1883 		mflags |= (PUSER - 1);
1884 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1885 			if (flags & V_NOWAIT) {
1886 				error = EWOULDBLOCK;
1887 				goto unlock;
1888 			}
1889 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1890 			    "suspfs", 0);
1891 			if (error)
1892 				goto unlock;
1893 		}
1894 	}
1895 	if (flags & V_XSLEEP)
1896 		goto unlock;
1897 	mp->mnt_writeopcount++;
1898 unlock:
1899 	if (error != 0 || (flags & V_XSLEEP) != 0)
1900 		MNT_REL(mp);
1901 	MNT_IUNLOCK(mp);
1902 	return (error);
1903 }
1904 
1905 int
1906 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1907 {
1908 	struct mount *mp;
1909 	int error;
1910 
1911 	KASSERT((flags & ~V_VALID_FLAGS) == 0,
1912 	    ("%s: invalid flags passed %d\n", __func__, flags));
1913 
1914 	error = 0;
1915 	/*
1916 	 * If a vnode is provided, get and return the mount point that
1917 	 * to which it will write.
1918 	 */
1919 	if (vp != NULL) {
1920 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1921 			*mpp = NULL;
1922 			if (error != EOPNOTSUPP)
1923 				return (error);
1924 			return (0);
1925 		}
1926 	}
1927 	if ((mp = *mpp) == NULL)
1928 		return (0);
1929 
1930 	/*
1931 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1932 	 * a vfs_ref().
1933 	 * As long as a vnode is not provided we need to acquire a
1934 	 * refcount for the provided mountpoint too, in order to
1935 	 * emulate a vfs_ref().
1936 	 */
1937 	if (vp == NULL)
1938 		vfs_ref(mp);
1939 
1940 	return (vn_start_write_refed(mp, flags, false));
1941 }
1942 
1943 /*
1944  * Secondary suspension. Used by operations such as vop_inactive
1945  * routines that are needed by the higher level functions. These
1946  * are allowed to proceed until all the higher level functions have
1947  * completed (indicated by mnt_writeopcount dropping to zero). At that
1948  * time, these operations are halted until the suspension is over.
1949  */
1950 int
1951 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1952 {
1953 	struct mount *mp;
1954 	int error, mflags;
1955 
1956 	KASSERT((flags & ~V_VALID_FLAGS) == 0,
1957 	    ("%s: invalid flags passed %d\n", __func__, flags));
1958 
1959  retry:
1960 	if (vp != NULL) {
1961 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1962 			*mpp = NULL;
1963 			if (error != EOPNOTSUPP)
1964 				return (error);
1965 			return (0);
1966 		}
1967 	}
1968 	/*
1969 	 * If we are not suspended or have not yet reached suspended
1970 	 * mode, then let the operation proceed.
1971 	 */
1972 	if ((mp = *mpp) == NULL)
1973 		return (0);
1974 
1975 	/*
1976 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1977 	 * a vfs_ref().
1978 	 * As long as a vnode is not provided we need to acquire a
1979 	 * refcount for the provided mountpoint too, in order to
1980 	 * emulate a vfs_ref().
1981 	 */
1982 	MNT_ILOCK(mp);
1983 	if (vp == NULL)
1984 		MNT_REF(mp);
1985 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1986 		mp->mnt_secondary_writes++;
1987 		mp->mnt_secondary_accwrites++;
1988 		MNT_IUNLOCK(mp);
1989 		return (0);
1990 	}
1991 	if (flags & V_NOWAIT) {
1992 		MNT_REL(mp);
1993 		MNT_IUNLOCK(mp);
1994 		return (EWOULDBLOCK);
1995 	}
1996 	/*
1997 	 * Wait for the suspension to finish.
1998 	 */
1999 	mflags = 0;
2000 	if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
2001 		if (flags & V_PCATCH)
2002 			mflags |= PCATCH;
2003 	}
2004 	mflags |= (PUSER - 1) | PDROP;
2005 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0);
2006 	vfs_rel(mp);
2007 	if (error == 0)
2008 		goto retry;
2009 	return (error);
2010 }
2011 
2012 /*
2013  * Filesystem write operation has completed. If we are suspending and this
2014  * operation is the last one, notify the suspender that the suspension is
2015  * now in effect.
2016  */
2017 void
2018 vn_finished_write(struct mount *mp)
2019 {
2020 	struct mount_pcpu *mpcpu;
2021 	int c;
2022 
2023 	if (mp == NULL)
2024 		return;
2025 
2026 	if (vfs_op_thread_enter(mp, mpcpu)) {
2027 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2028 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2029 		vfs_op_thread_exit(mp, mpcpu);
2030 		return;
2031 	}
2032 
2033 	MNT_ILOCK(mp);
2034 	vfs_assert_mount_counters(mp);
2035 	MNT_REL(mp);
2036 	c = --mp->mnt_writeopcount;
2037 	if (mp->mnt_vfs_ops == 0) {
2038 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2039 		MNT_IUNLOCK(mp);
2040 		return;
2041 	}
2042 	if (c < 0)
2043 		vfs_dump_mount_counters(mp);
2044 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2045 		wakeup(&mp->mnt_writeopcount);
2046 	MNT_IUNLOCK(mp);
2047 }
2048 
2049 /*
2050  * Filesystem secondary write operation has completed. If we are
2051  * suspending and this operation is the last one, notify the suspender
2052  * that the suspension is now in effect.
2053  */
2054 void
2055 vn_finished_secondary_write(struct mount *mp)
2056 {
2057 	if (mp == NULL)
2058 		return;
2059 	MNT_ILOCK(mp);
2060 	MNT_REL(mp);
2061 	mp->mnt_secondary_writes--;
2062 	if (mp->mnt_secondary_writes < 0)
2063 		panic("vn_finished_secondary_write: neg cnt");
2064 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2065 	    mp->mnt_secondary_writes <= 0)
2066 		wakeup(&mp->mnt_secondary_writes);
2067 	MNT_IUNLOCK(mp);
2068 }
2069 
2070 /*
2071  * Request a filesystem to suspend write operations.
2072  */
2073 int
2074 vfs_write_suspend(struct mount *mp, int flags)
2075 {
2076 	int error;
2077 
2078 	vfs_op_enter(mp);
2079 
2080 	MNT_ILOCK(mp);
2081 	vfs_assert_mount_counters(mp);
2082 	if (mp->mnt_susp_owner == curthread) {
2083 		vfs_op_exit_locked(mp);
2084 		MNT_IUNLOCK(mp);
2085 		return (EALREADY);
2086 	}
2087 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2088 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2089 
2090 	/*
2091 	 * Unmount holds a write reference on the mount point.  If we
2092 	 * own busy reference and drain for writers, we deadlock with
2093 	 * the reference draining in the unmount path.  Callers of
2094 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2095 	 * vfs_busy() reference is owned and caller is not in the
2096 	 * unmount context.
2097 	 */
2098 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2099 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2100 		vfs_op_exit_locked(mp);
2101 		MNT_IUNLOCK(mp);
2102 		return (EBUSY);
2103 	}
2104 
2105 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2106 	mp->mnt_susp_owner = curthread;
2107 	if (mp->mnt_writeopcount > 0)
2108 		(void) msleep(&mp->mnt_writeopcount,
2109 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2110 	else
2111 		MNT_IUNLOCK(mp);
2112 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2113 		vfs_write_resume(mp, 0);
2114 		/* vfs_write_resume does vfs_op_exit() for us */
2115 	}
2116 	return (error);
2117 }
2118 
2119 /*
2120  * Request a filesystem to resume write operations.
2121  */
2122 void
2123 vfs_write_resume(struct mount *mp, int flags)
2124 {
2125 
2126 	MNT_ILOCK(mp);
2127 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2128 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2129 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2130 				       MNTK_SUSPENDED);
2131 		mp->mnt_susp_owner = NULL;
2132 		wakeup(&mp->mnt_writeopcount);
2133 		wakeup(&mp->mnt_flag);
2134 		curthread->td_pflags &= ~TDP_IGNSUSP;
2135 		if ((flags & VR_START_WRITE) != 0) {
2136 			MNT_REF(mp);
2137 			mp->mnt_writeopcount++;
2138 		}
2139 		MNT_IUNLOCK(mp);
2140 		if ((flags & VR_NO_SUSPCLR) == 0)
2141 			VFS_SUSP_CLEAN(mp);
2142 		vfs_op_exit(mp);
2143 	} else if ((flags & VR_START_WRITE) != 0) {
2144 		MNT_REF(mp);
2145 		vn_start_write_refed(mp, 0, true);
2146 	} else {
2147 		MNT_IUNLOCK(mp);
2148 	}
2149 }
2150 
2151 /*
2152  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2153  * methods.
2154  */
2155 int
2156 vfs_write_suspend_umnt(struct mount *mp)
2157 {
2158 	int error;
2159 
2160 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2161 	    ("vfs_write_suspend_umnt: recursed"));
2162 
2163 	/* dounmount() already called vn_start_write(). */
2164 	for (;;) {
2165 		vn_finished_write(mp);
2166 		error = vfs_write_suspend(mp, 0);
2167 		if (error != 0) {
2168 			vn_start_write(NULL, &mp, V_WAIT);
2169 			return (error);
2170 		}
2171 		MNT_ILOCK(mp);
2172 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2173 			break;
2174 		MNT_IUNLOCK(mp);
2175 		vn_start_write(NULL, &mp, V_WAIT);
2176 	}
2177 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2178 	wakeup(&mp->mnt_flag);
2179 	MNT_IUNLOCK(mp);
2180 	curthread->td_pflags |= TDP_IGNSUSP;
2181 	return (0);
2182 }
2183 
2184 /*
2185  * Implement kqueues for files by translating it to vnode operation.
2186  */
2187 static int
2188 vn_kqfilter(struct file *fp, struct knote *kn)
2189 {
2190 
2191 	return (VOP_KQFILTER(fp->f_vnode, kn));
2192 }
2193 
2194 int
2195 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2196 {
2197 	if ((fp->f_flag & FKQALLOWED) == 0)
2198 		return (EBADF);
2199 	return (vn_kqfilter(fp, kn));
2200 }
2201 
2202 /*
2203  * Simplified in-kernel wrapper calls for extended attribute access.
2204  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2205  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2206  */
2207 int
2208 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2209     const char *attrname, int *buflen, char *buf, struct thread *td)
2210 {
2211 	struct uio	auio;
2212 	struct iovec	iov;
2213 	int	error;
2214 
2215 	iov.iov_len = *buflen;
2216 	iov.iov_base = buf;
2217 
2218 	auio.uio_iov = &iov;
2219 	auio.uio_iovcnt = 1;
2220 	auio.uio_rw = UIO_READ;
2221 	auio.uio_segflg = UIO_SYSSPACE;
2222 	auio.uio_td = td;
2223 	auio.uio_offset = 0;
2224 	auio.uio_resid = *buflen;
2225 
2226 	if ((ioflg & IO_NODELOCKED) == 0)
2227 		vn_lock(vp, LK_SHARED | LK_RETRY);
2228 
2229 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2230 
2231 	/* authorize attribute retrieval as kernel */
2232 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2233 	    td);
2234 
2235 	if ((ioflg & IO_NODELOCKED) == 0)
2236 		VOP_UNLOCK(vp);
2237 
2238 	if (error == 0) {
2239 		*buflen = *buflen - auio.uio_resid;
2240 	}
2241 
2242 	return (error);
2243 }
2244 
2245 /*
2246  * XXX failure mode if partially written?
2247  */
2248 int
2249 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2250     const char *attrname, int buflen, char *buf, struct thread *td)
2251 {
2252 	struct uio	auio;
2253 	struct iovec	iov;
2254 	struct mount	*mp;
2255 	int	error;
2256 
2257 	iov.iov_len = buflen;
2258 	iov.iov_base = buf;
2259 
2260 	auio.uio_iov = &iov;
2261 	auio.uio_iovcnt = 1;
2262 	auio.uio_rw = UIO_WRITE;
2263 	auio.uio_segflg = UIO_SYSSPACE;
2264 	auio.uio_td = td;
2265 	auio.uio_offset = 0;
2266 	auio.uio_resid = buflen;
2267 
2268 	if ((ioflg & IO_NODELOCKED) == 0) {
2269 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2270 			return (error);
2271 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2272 	}
2273 
2274 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2275 
2276 	/* authorize attribute setting as kernel */
2277 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2278 
2279 	if ((ioflg & IO_NODELOCKED) == 0) {
2280 		vn_finished_write(mp);
2281 		VOP_UNLOCK(vp);
2282 	}
2283 
2284 	return (error);
2285 }
2286 
2287 int
2288 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2289     const char *attrname, struct thread *td)
2290 {
2291 	struct mount	*mp;
2292 	int	error;
2293 
2294 	if ((ioflg & IO_NODELOCKED) == 0) {
2295 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2296 			return (error);
2297 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2298 	}
2299 
2300 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2301 
2302 	/* authorize attribute removal as kernel */
2303 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2304 	if (error == EOPNOTSUPP)
2305 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2306 		    NULL, td);
2307 
2308 	if ((ioflg & IO_NODELOCKED) == 0) {
2309 		vn_finished_write(mp);
2310 		VOP_UNLOCK(vp);
2311 	}
2312 
2313 	return (error);
2314 }
2315 
2316 static int
2317 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2318     struct vnode **rvp)
2319 {
2320 
2321 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2322 }
2323 
2324 int
2325 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2326 {
2327 
2328 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2329 	    lkflags, rvp));
2330 }
2331 
2332 int
2333 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2334     int lkflags, struct vnode **rvp)
2335 {
2336 	struct mount *mp;
2337 	int ltype, error;
2338 
2339 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2340 	mp = vp->v_mount;
2341 	ltype = VOP_ISLOCKED(vp);
2342 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2343 	    ("vn_vget_ino: vp not locked"));
2344 	error = vfs_busy(mp, MBF_NOWAIT);
2345 	if (error != 0) {
2346 		vfs_ref(mp);
2347 		VOP_UNLOCK(vp);
2348 		error = vfs_busy(mp, 0);
2349 		vn_lock(vp, ltype | LK_RETRY);
2350 		vfs_rel(mp);
2351 		if (error != 0)
2352 			return (ENOENT);
2353 		if (VN_IS_DOOMED(vp)) {
2354 			vfs_unbusy(mp);
2355 			return (ENOENT);
2356 		}
2357 	}
2358 	VOP_UNLOCK(vp);
2359 	error = alloc(mp, alloc_arg, lkflags, rvp);
2360 	vfs_unbusy(mp);
2361 	if (error != 0 || *rvp != vp)
2362 		vn_lock(vp, ltype | LK_RETRY);
2363 	if (VN_IS_DOOMED(vp)) {
2364 		if (error == 0) {
2365 			if (*rvp == vp)
2366 				vunref(vp);
2367 			else
2368 				vput(*rvp);
2369 		}
2370 		error = ENOENT;
2371 	}
2372 	return (error);
2373 }
2374 
2375 static void
2376 vn_send_sigxfsz(struct proc *p)
2377 {
2378 	PROC_LOCK(p);
2379 	kern_psignal(p, SIGXFSZ);
2380 	PROC_UNLOCK(p);
2381 }
2382 
2383 int
2384 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2385 {
2386 	if (size <= lim_cur(td, RLIMIT_FSIZE))
2387 		return (0);
2388 	vn_send_sigxfsz(td->td_proc);
2389 	return (EFBIG);
2390 }
2391 
2392 static int
2393 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2394     bool adj, struct thread *td)
2395 {
2396 	off_t lim;
2397 	bool ktr_write;
2398 
2399 	if (vp->v_type != VREG)
2400 		return (0);
2401 
2402 	/*
2403 	 * Handle file system maximum file size.
2404 	 */
2405 	if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2406 		if (!adj || uio->uio_offset >= maxfsz)
2407 			return (EFBIG);
2408 		uio->uio_resid = maxfsz - uio->uio_offset;
2409 	}
2410 
2411 	/*
2412 	 * This is kernel write (e.g. vnode_pager) or accounting
2413 	 * write, ignore limit.
2414 	 */
2415 	if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2416 		return (0);
2417 
2418 	/*
2419 	 * Calculate file size limit.
2420 	 */
2421 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2422 	lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2423 	    lim_cur(td, RLIMIT_FSIZE);
2424 
2425 	/*
2426 	 * Is the limit reached?
2427 	 */
2428 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2429 		return (0);
2430 
2431 	/*
2432 	 * Prepared filesystems can handle writes truncated to the
2433 	 * file size limit.
2434 	 */
2435 	if (adj && (uoff_t)uio->uio_offset < lim) {
2436 		uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2437 		return (0);
2438 	}
2439 
2440 	if (!ktr_write || ktr_filesize_limit_signal)
2441 		vn_send_sigxfsz(td->td_proc);
2442 	return (EFBIG);
2443 }
2444 
2445 /*
2446  * Helper for VOP_WRITE() implementations, the common code to
2447  * handle maximum supported file size on the filesystem, and
2448  * RLIMIT_FSIZE, except for special writes from accounting subsystem
2449  * and ktrace.
2450  *
2451  * For maximum file size (maxfsz argument):
2452  * - return EFBIG if uio_offset is beyond it
2453  * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2454  *
2455  * For RLIMIT_FSIZE:
2456  * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2457  * - otherwise, clamp uio_resid if write would extend file beyond limit.
2458  *
2459  * If clamping occured, the adjustment for uio_resid is stored in
2460  * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2461  * from the VOP.
2462  */
2463 int
2464 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2465     ssize_t *resid_adj, struct thread *td)
2466 {
2467 	ssize_t resid_orig;
2468 	int error;
2469 	bool adj;
2470 
2471 	resid_orig = uio->uio_resid;
2472 	adj = resid_adj != NULL;
2473 	error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2474 	if (adj)
2475 		*resid_adj = resid_orig - uio->uio_resid;
2476 	return (error);
2477 }
2478 
2479 void
2480 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2481 {
2482 	uio->uio_resid += resid_adj;
2483 }
2484 
2485 int
2486 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2487     struct thread *td)
2488 {
2489 	return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2490 	    td));
2491 }
2492 
2493 int
2494 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2495     struct thread *td)
2496 {
2497 	struct vnode *vp;
2498 
2499 	vp = fp->f_vnode;
2500 #ifdef AUDIT
2501 	vn_lock(vp, LK_SHARED | LK_RETRY);
2502 	AUDIT_ARG_VNODE1(vp);
2503 	VOP_UNLOCK(vp);
2504 #endif
2505 	return (setfmode(td, active_cred, vp, mode));
2506 }
2507 
2508 int
2509 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2510     struct thread *td)
2511 {
2512 	struct vnode *vp;
2513 
2514 	vp = fp->f_vnode;
2515 #ifdef AUDIT
2516 	vn_lock(vp, LK_SHARED | LK_RETRY);
2517 	AUDIT_ARG_VNODE1(vp);
2518 	VOP_UNLOCK(vp);
2519 #endif
2520 	return (setfown(td, active_cred, vp, uid, gid));
2521 }
2522 
2523 /*
2524  * Remove pages in the range ["start", "end") from the vnode's VM object.  If
2525  * "end" is 0, then the range extends to the end of the object.
2526  */
2527 void
2528 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2529 {
2530 	vm_object_t object;
2531 
2532 	if ((object = vp->v_object) == NULL)
2533 		return;
2534 	VM_OBJECT_WLOCK(object);
2535 	vm_object_page_remove(object, start, end, 0);
2536 	VM_OBJECT_WUNLOCK(object);
2537 }
2538 
2539 /*
2540  * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2541  * mapped into any process' address space.  Filesystems may use this in
2542  * preference to vn_pages_remove() to avoid blocking on pages busied in
2543  * preparation for a VOP_GETPAGES.
2544  */
2545 void
2546 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2547 {
2548 	vm_object_t object;
2549 
2550 	if ((object = vp->v_object) == NULL)
2551 		return;
2552 	VM_OBJECT_WLOCK(object);
2553 	vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2554 	VM_OBJECT_WUNLOCK(object);
2555 }
2556 
2557 int
2558 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2559     struct ucred *cred)
2560 {
2561 	struct vattr va;
2562 	daddr_t bn, bnp;
2563 	uint64_t bsize;
2564 	off_t noff;
2565 	int error;
2566 
2567 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2568 	    ("%s: Wrong command %lu", __func__, cmd));
2569 	ASSERT_VOP_LOCKED(vp, "vn_bmap_seekhole_locked");
2570 
2571 	if (vp->v_type != VREG) {
2572 		error = ENOTTY;
2573 		goto out;
2574 	}
2575 	error = VOP_GETATTR(vp, &va, cred);
2576 	if (error != 0)
2577 		goto out;
2578 	noff = *off;
2579 	if (noff < 0 || noff >= va.va_size) {
2580 		error = ENXIO;
2581 		goto out;
2582 	}
2583 	bsize = vp->v_mount->mnt_stat.f_iosize;
2584 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2585 	    noff % bsize) {
2586 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2587 		if (error == EOPNOTSUPP) {
2588 			error = ENOTTY;
2589 			goto out;
2590 		}
2591 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2592 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2593 			noff = bn * bsize;
2594 			if (noff < *off)
2595 				noff = *off;
2596 			goto out;
2597 		}
2598 	}
2599 	if (noff > va.va_size)
2600 		noff = va.va_size;
2601 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2602 	if (cmd == FIOSEEKDATA)
2603 		error = ENXIO;
2604 out:
2605 	if (error == 0)
2606 		*off = noff;
2607 	return (error);
2608 }
2609 
2610 int
2611 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2612 {
2613 	int error;
2614 
2615 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2616 	    ("%s: Wrong command %lu", __func__, cmd));
2617 
2618 	if (vn_lock(vp, LK_SHARED) != 0)
2619 		return (EBADF);
2620 	error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2621 	VOP_UNLOCK(vp);
2622 	return (error);
2623 }
2624 
2625 int
2626 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2627 {
2628 	struct ucred *cred;
2629 	struct vnode *vp;
2630 	struct vattr vattr;
2631 	off_t foffset, size;
2632 	int error, noneg;
2633 
2634 	cred = td->td_ucred;
2635 	vp = fp->f_vnode;
2636 	foffset = foffset_lock(fp, 0);
2637 	noneg = (vp->v_type != VCHR);
2638 	error = 0;
2639 	switch (whence) {
2640 	case L_INCR:
2641 		if (noneg &&
2642 		    (foffset < 0 ||
2643 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2644 			error = EOVERFLOW;
2645 			break;
2646 		}
2647 		offset += foffset;
2648 		break;
2649 	case L_XTND:
2650 		vn_lock(vp, LK_SHARED | LK_RETRY);
2651 		error = VOP_GETATTR(vp, &vattr, cred);
2652 		VOP_UNLOCK(vp);
2653 		if (error)
2654 			break;
2655 
2656 		/*
2657 		 * If the file references a disk device, then fetch
2658 		 * the media size and use that to determine the ending
2659 		 * offset.
2660 		 */
2661 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2662 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2663 			vattr.va_size = size;
2664 		if (noneg &&
2665 		    (vattr.va_size > OFF_MAX ||
2666 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2667 			error = EOVERFLOW;
2668 			break;
2669 		}
2670 		offset += vattr.va_size;
2671 		break;
2672 	case L_SET:
2673 		break;
2674 	case SEEK_DATA:
2675 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2676 		if (error == ENOTTY)
2677 			error = EINVAL;
2678 		break;
2679 	case SEEK_HOLE:
2680 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2681 		if (error == ENOTTY)
2682 			error = EINVAL;
2683 		break;
2684 	default:
2685 		error = EINVAL;
2686 	}
2687 	if (error == 0 && noneg && offset < 0)
2688 		error = EINVAL;
2689 	if (error != 0)
2690 		goto drop;
2691 	VFS_KNOTE_UNLOCKED(vp, 0);
2692 	td->td_uretoff.tdu_off = offset;
2693 drop:
2694 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2695 	return (error);
2696 }
2697 
2698 int
2699 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2700     struct thread *td)
2701 {
2702 	int error;
2703 
2704 	/*
2705 	 * Grant permission if the caller is the owner of the file, or
2706 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2707 	 * on the file.  If the time pointer is null, then write
2708 	 * permission on the file is also sufficient.
2709 	 *
2710 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2711 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2712 	 * will be allowed to set the times [..] to the current
2713 	 * server time.
2714 	 */
2715 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2716 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2717 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2718 	return (error);
2719 }
2720 
2721 int
2722 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2723 {
2724 	struct vnode *vp;
2725 	int error;
2726 
2727 	if (fp->f_type == DTYPE_FIFO)
2728 		kif->kf_type = KF_TYPE_FIFO;
2729 	else
2730 		kif->kf_type = KF_TYPE_VNODE;
2731 	vp = fp->f_vnode;
2732 	vref(vp);
2733 	FILEDESC_SUNLOCK(fdp);
2734 	error = vn_fill_kinfo_vnode(vp, kif);
2735 	vrele(vp);
2736 	FILEDESC_SLOCK(fdp);
2737 	return (error);
2738 }
2739 
2740 static inline void
2741 vn_fill_junk(struct kinfo_file *kif)
2742 {
2743 	size_t len, olen;
2744 
2745 	/*
2746 	 * Simulate vn_fullpath returning changing values for a given
2747 	 * vp during e.g. coredump.
2748 	 */
2749 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2750 	olen = strlen(kif->kf_path);
2751 	if (len < olen)
2752 		strcpy(&kif->kf_path[len - 1], "$");
2753 	else
2754 		for (; olen < len; olen++)
2755 			strcpy(&kif->kf_path[olen], "A");
2756 }
2757 
2758 int
2759 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2760 {
2761 	struct vattr va;
2762 	char *fullpath, *freepath;
2763 	int error;
2764 
2765 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2766 	freepath = NULL;
2767 	fullpath = "-";
2768 	error = vn_fullpath(vp, &fullpath, &freepath);
2769 	if (error == 0) {
2770 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2771 	}
2772 	if (freepath != NULL)
2773 		free(freepath, M_TEMP);
2774 
2775 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2776 		vn_fill_junk(kif);
2777 	);
2778 
2779 	/*
2780 	 * Retrieve vnode attributes.
2781 	 */
2782 	va.va_fsid = VNOVAL;
2783 	va.va_rdev = NODEV;
2784 	vn_lock(vp, LK_SHARED | LK_RETRY);
2785 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2786 	VOP_UNLOCK(vp);
2787 	if (error != 0)
2788 		return (error);
2789 	if (va.va_fsid != VNOVAL)
2790 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2791 	else
2792 		kif->kf_un.kf_file.kf_file_fsid =
2793 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2794 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2795 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2796 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2797 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2798 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2799 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2800 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2801 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2802 	return (0);
2803 }
2804 
2805 int
2806 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2807     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2808     struct thread *td)
2809 {
2810 #ifdef HWPMC_HOOKS
2811 	struct pmckern_map_in pkm;
2812 #endif
2813 	struct mount *mp;
2814 	struct vnode *vp;
2815 	vm_object_t object;
2816 	vm_prot_t maxprot;
2817 	boolean_t writecounted;
2818 	int error;
2819 
2820 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2821     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2822 	/*
2823 	 * POSIX shared-memory objects are defined to have
2824 	 * kernel persistence, and are not defined to support
2825 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2826 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2827 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2828 	 * flag to request this behavior.
2829 	 */
2830 	if ((fp->f_flag & FPOSIXSHM) != 0)
2831 		flags |= MAP_NOSYNC;
2832 #endif
2833 	vp = fp->f_vnode;
2834 
2835 	/*
2836 	 * Ensure that file and memory protections are
2837 	 * compatible.  Note that we only worry about
2838 	 * writability if mapping is shared; in this case,
2839 	 * current and max prot are dictated by the open file.
2840 	 * XXX use the vnode instead?  Problem is: what
2841 	 * credentials do we use for determination? What if
2842 	 * proc does a setuid?
2843 	 */
2844 	mp = vp->v_mount;
2845 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2846 		maxprot = VM_PROT_NONE;
2847 		if ((prot & VM_PROT_EXECUTE) != 0)
2848 			return (EACCES);
2849 	} else
2850 		maxprot = VM_PROT_EXECUTE;
2851 	if ((fp->f_flag & FREAD) != 0)
2852 		maxprot |= VM_PROT_READ;
2853 	else if ((prot & VM_PROT_READ) != 0)
2854 		return (EACCES);
2855 
2856 	/*
2857 	 * If we are sharing potential changes via MAP_SHARED and we
2858 	 * are trying to get write permission although we opened it
2859 	 * without asking for it, bail out.
2860 	 */
2861 	if ((flags & MAP_SHARED) != 0) {
2862 		if ((fp->f_flag & FWRITE) != 0)
2863 			maxprot |= VM_PROT_WRITE;
2864 		else if ((prot & VM_PROT_WRITE) != 0)
2865 			return (EACCES);
2866 	} else {
2867 		maxprot |= VM_PROT_WRITE;
2868 		cap_maxprot |= VM_PROT_WRITE;
2869 	}
2870 	maxprot &= cap_maxprot;
2871 
2872 	/*
2873 	 * For regular files and shared memory, POSIX requires that
2874 	 * the value of foff be a legitimate offset within the data
2875 	 * object.  In particular, negative offsets are invalid.
2876 	 * Blocking negative offsets and overflows here avoids
2877 	 * possible wraparound or user-level access into reserved
2878 	 * ranges of the data object later.  In contrast, POSIX does
2879 	 * not dictate how offsets are used by device drivers, so in
2880 	 * the case of a device mapping a negative offset is passed
2881 	 * on.
2882 	 */
2883 	if (
2884 #ifdef _LP64
2885 	    size > OFF_MAX ||
2886 #endif
2887 	    foff > OFF_MAX - size)
2888 		return (EINVAL);
2889 
2890 	writecounted = FALSE;
2891 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2892 	    &foff, &object, &writecounted);
2893 	if (error != 0)
2894 		return (error);
2895 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2896 	    foff, writecounted, td);
2897 	if (error != 0) {
2898 		/*
2899 		 * If this mapping was accounted for in the vnode's
2900 		 * writecount, then undo that now.
2901 		 */
2902 		if (writecounted)
2903 			vm_pager_release_writecount(object, 0, size);
2904 		vm_object_deallocate(object);
2905 	}
2906 #ifdef HWPMC_HOOKS
2907 	/* Inform hwpmc(4) if an executable is being mapped. */
2908 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2909 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2910 			pkm.pm_file = vp;
2911 			pkm.pm_address = (uintptr_t) *addr;
2912 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2913 		}
2914 	}
2915 #endif
2916 	return (error);
2917 }
2918 
2919 void
2920 vn_fsid(struct vnode *vp, struct vattr *va)
2921 {
2922 	fsid_t *f;
2923 
2924 	f = &vp->v_mount->mnt_stat.f_fsid;
2925 	va->va_fsid = (uint32_t)f->val[1];
2926 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2927 	va->va_fsid += (uint32_t)f->val[0];
2928 }
2929 
2930 int
2931 vn_fsync_buf(struct vnode *vp, int waitfor)
2932 {
2933 	struct buf *bp, *nbp;
2934 	struct bufobj *bo;
2935 	struct mount *mp;
2936 	int error, maxretry;
2937 
2938 	error = 0;
2939 	maxretry = 10000;     /* large, arbitrarily chosen */
2940 	mp = NULL;
2941 	if (vp->v_type == VCHR) {
2942 		VI_LOCK(vp);
2943 		mp = vp->v_rdev->si_mountpt;
2944 		VI_UNLOCK(vp);
2945 	}
2946 	bo = &vp->v_bufobj;
2947 	BO_LOCK(bo);
2948 loop1:
2949 	/*
2950 	 * MARK/SCAN initialization to avoid infinite loops.
2951 	 */
2952         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2953 		bp->b_vflags &= ~BV_SCANNED;
2954 		bp->b_error = 0;
2955 	}
2956 
2957 	/*
2958 	 * Flush all dirty buffers associated with a vnode.
2959 	 */
2960 loop2:
2961 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2962 		if ((bp->b_vflags & BV_SCANNED) != 0)
2963 			continue;
2964 		bp->b_vflags |= BV_SCANNED;
2965 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2966 			if (waitfor != MNT_WAIT)
2967 				continue;
2968 			if (BUF_LOCK(bp,
2969 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2970 			    BO_LOCKPTR(bo)) != 0) {
2971 				BO_LOCK(bo);
2972 				goto loop1;
2973 			}
2974 			BO_LOCK(bo);
2975 		}
2976 		BO_UNLOCK(bo);
2977 		KASSERT(bp->b_bufobj == bo,
2978 		    ("bp %p wrong b_bufobj %p should be %p",
2979 		    bp, bp->b_bufobj, bo));
2980 		if ((bp->b_flags & B_DELWRI) == 0)
2981 			panic("fsync: not dirty");
2982 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2983 			vfs_bio_awrite(bp);
2984 		} else {
2985 			bremfree(bp);
2986 			bawrite(bp);
2987 		}
2988 		if (maxretry < 1000)
2989 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2990 		BO_LOCK(bo);
2991 		goto loop2;
2992 	}
2993 
2994 	/*
2995 	 * If synchronous the caller expects us to completely resolve all
2996 	 * dirty buffers in the system.  Wait for in-progress I/O to
2997 	 * complete (which could include background bitmap writes), then
2998 	 * retry if dirty blocks still exist.
2999 	 */
3000 	if (waitfor == MNT_WAIT) {
3001 		bufobj_wwait(bo, 0, 0);
3002 		if (bo->bo_dirty.bv_cnt > 0) {
3003 			/*
3004 			 * If we are unable to write any of these buffers
3005 			 * then we fail now rather than trying endlessly
3006 			 * to write them out.
3007 			 */
3008 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
3009 				if ((error = bp->b_error) != 0)
3010 					break;
3011 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
3012 			    (error == 0 && --maxretry >= 0))
3013 				goto loop1;
3014 			if (error == 0)
3015 				error = EAGAIN;
3016 		}
3017 	}
3018 	BO_UNLOCK(bo);
3019 	if (error != 0)
3020 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3021 
3022 	return (error);
3023 }
3024 
3025 /*
3026  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
3027  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3028  * to do the actual copy.
3029  * vn_generic_copy_file_range() is factored out, so it can be called
3030  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3031  * different file systems.
3032  */
3033 int
3034 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3035     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3036     struct ucred *outcred, struct thread *fsize_td)
3037 {
3038 	int error;
3039 	size_t len;
3040 	uint64_t uval;
3041 
3042 	len = *lenp;
3043 	*lenp = 0;		/* For error returns. */
3044 	error = 0;
3045 
3046 	/* Do some sanity checks on the arguments. */
3047 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
3048 		error = EISDIR;
3049 	else if (*inoffp < 0 || *outoffp < 0 ||
3050 	    invp->v_type != VREG || outvp->v_type != VREG)
3051 		error = EINVAL;
3052 	if (error != 0)
3053 		goto out;
3054 
3055 	/* Ensure offset + len does not wrap around. */
3056 	uval = *inoffp;
3057 	uval += len;
3058 	if (uval > INT64_MAX)
3059 		len = INT64_MAX - *inoffp;
3060 	uval = *outoffp;
3061 	uval += len;
3062 	if (uval > INT64_MAX)
3063 		len = INT64_MAX - *outoffp;
3064 	if (len == 0)
3065 		goto out;
3066 
3067 	/*
3068 	 * If the two vnode are for the same file system, call
3069 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3070 	 * which can handle copies across multiple file systems.
3071 	 */
3072 	*lenp = len;
3073 	if (invp->v_mount == outvp->v_mount)
3074 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
3075 		    lenp, flags, incred, outcred, fsize_td);
3076 	else
3077 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
3078 		    outoffp, lenp, flags, incred, outcred, fsize_td);
3079 out:
3080 	return (error);
3081 }
3082 
3083 /*
3084  * Test len bytes of data starting at dat for all bytes == 0.
3085  * Return true if all bytes are zero, false otherwise.
3086  * Expects dat to be well aligned.
3087  */
3088 static bool
3089 mem_iszero(void *dat, int len)
3090 {
3091 	int i;
3092 	const u_int *p;
3093 	const char *cp;
3094 
3095 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
3096 		if (len >= sizeof(*p)) {
3097 			if (*p != 0)
3098 				return (false);
3099 		} else {
3100 			cp = (const char *)p;
3101 			for (i = 0; i < len; i++, cp++)
3102 				if (*cp != '\0')
3103 					return (false);
3104 		}
3105 	}
3106 	return (true);
3107 }
3108 
3109 /*
3110  * Look for a hole in the output file and, if found, adjust *outoffp
3111  * and *xferp to skip past the hole.
3112  * *xferp is the entire hole length to be written and xfer2 is how many bytes
3113  * to be written as 0's upon return.
3114  */
3115 static off_t
3116 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3117     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3118 {
3119 	int error;
3120 	off_t delta;
3121 
3122 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3123 		*dataoffp = *outoffp;
3124 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3125 		    curthread);
3126 		if (error == 0) {
3127 			*holeoffp = *dataoffp;
3128 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3129 			    curthread);
3130 		}
3131 		if (error != 0 || *holeoffp == *dataoffp) {
3132 			/*
3133 			 * Since outvp is unlocked, it may be possible for
3134 			 * another thread to do a truncate(), lseek(), write()
3135 			 * creating a hole at startoff between the above
3136 			 * VOP_IOCTL() calls, if the other thread does not do
3137 			 * rangelocking.
3138 			 * If that happens, *holeoffp == *dataoffp and finding
3139 			 * the hole has failed, so disable vn_skip_hole().
3140 			 */
3141 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3142 			return (xfer2);
3143 		}
3144 		KASSERT(*dataoffp >= *outoffp,
3145 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3146 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3147 		KASSERT(*holeoffp > *dataoffp,
3148 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3149 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3150 	}
3151 
3152 	/*
3153 	 * If there is a hole before the data starts, advance *outoffp and
3154 	 * *xferp past the hole.
3155 	 */
3156 	if (*dataoffp > *outoffp) {
3157 		delta = *dataoffp - *outoffp;
3158 		if (delta >= *xferp) {
3159 			/* Entire *xferp is a hole. */
3160 			*outoffp += *xferp;
3161 			*xferp = 0;
3162 			return (0);
3163 		}
3164 		*xferp -= delta;
3165 		*outoffp += delta;
3166 		xfer2 = MIN(xfer2, *xferp);
3167 	}
3168 
3169 	/*
3170 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3171 	 * that the write ends at the start of the hole.
3172 	 * *holeoffp should always be greater than *outoffp, but for the
3173 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3174 	 * value.
3175 	 */
3176 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3177 		xfer2 = *holeoffp - *outoffp;
3178 	return (xfer2);
3179 }
3180 
3181 /*
3182  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3183  * dat is a maximum of blksize in length and can be written repeatedly in
3184  * the chunk.
3185  * If growfile == true, just grow the file via vn_truncate_locked() instead
3186  * of doing actual writes.
3187  * If checkhole == true, a hole is being punched, so skip over any hole
3188  * already in the output file.
3189  */
3190 static int
3191 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3192     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3193 {
3194 	struct mount *mp;
3195 	off_t dataoff, holeoff, xfer2;
3196 	int error;
3197 
3198 	/*
3199 	 * Loop around doing writes of blksize until write has been completed.
3200 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3201 	 * done for each iteration, since the xfer argument can be very
3202 	 * large if there is a large hole to punch in the output file.
3203 	 */
3204 	error = 0;
3205 	holeoff = 0;
3206 	do {
3207 		xfer2 = MIN(xfer, blksize);
3208 		if (checkhole) {
3209 			/*
3210 			 * Punching a hole.  Skip writing if there is
3211 			 * already a hole in the output file.
3212 			 */
3213 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3214 			    &dataoff, &holeoff, cred);
3215 			if (xfer == 0)
3216 				break;
3217 			if (holeoff < 0)
3218 				checkhole = false;
3219 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3220 			    (intmax_t)xfer2));
3221 		}
3222 		bwillwrite();
3223 		mp = NULL;
3224 		error = vn_start_write(outvp, &mp, V_WAIT);
3225 		if (error != 0)
3226 			break;
3227 		if (growfile) {
3228 			error = vn_lock(outvp, LK_EXCLUSIVE);
3229 			if (error == 0) {
3230 				error = vn_truncate_locked(outvp, outoff + xfer,
3231 				    false, cred);
3232 				VOP_UNLOCK(outvp);
3233 			}
3234 		} else {
3235 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3236 			if (error == 0) {
3237 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3238 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3239 				    curthread->td_ucred, cred, NULL, curthread);
3240 				outoff += xfer2;
3241 				xfer -= xfer2;
3242 				VOP_UNLOCK(outvp);
3243 			}
3244 		}
3245 		if (mp != NULL)
3246 			vn_finished_write(mp);
3247 	} while (!growfile && xfer > 0 && error == 0);
3248 	return (error);
3249 }
3250 
3251 /*
3252  * Copy a byte range of one file to another.  This function can handle the
3253  * case where invp and outvp are on different file systems.
3254  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3255  * is no better file system specific way to do it.
3256  */
3257 int
3258 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3259     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3260     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3261 {
3262 	struct vattr va, inva;
3263 	struct mount *mp;
3264 	off_t startoff, endoff, xfer, xfer2;
3265 	u_long blksize;
3266 	int error, interrupted;
3267 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3268 	ssize_t aresid, r = 0;
3269 	size_t copylen, len, savlen;
3270 	char *dat;
3271 	long holein, holeout;
3272 	struct timespec curts, endts;
3273 
3274 	holein = holeout = 0;
3275 	savlen = len = *lenp;
3276 	error = 0;
3277 	interrupted = 0;
3278 	dat = NULL;
3279 
3280 	error = vn_lock(invp, LK_SHARED);
3281 	if (error != 0)
3282 		goto out;
3283 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3284 		holein = 0;
3285 	if (holein > 0)
3286 		error = VOP_GETATTR(invp, &inva, incred);
3287 	VOP_UNLOCK(invp);
3288 	if (error != 0)
3289 		goto out;
3290 
3291 	mp = NULL;
3292 	error = vn_start_write(outvp, &mp, V_WAIT);
3293 	if (error == 0)
3294 		error = vn_lock(outvp, LK_EXCLUSIVE);
3295 	if (error == 0) {
3296 		/*
3297 		 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3298 		 * now that outvp is locked.
3299 		 */
3300 		if (fsize_td != NULL) {
3301 			struct uio io;
3302 
3303 			io.uio_offset = *outoffp;
3304 			io.uio_resid = len;
3305 			error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3306 			len = savlen = io.uio_resid;
3307 			/*
3308 			 * No need to call vn_rlimit_fsizex_res before return,
3309 			 * since the uio is local.
3310 			 */
3311 		}
3312 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3313 			holeout = 0;
3314 		/*
3315 		 * Holes that are past EOF do not need to be written as a block
3316 		 * of zero bytes.  So, truncate the output file as far as
3317 		 * possible and then use va.va_size to decide if writing 0
3318 		 * bytes is necessary in the loop below.
3319 		 */
3320 		if (error == 0)
3321 			error = VOP_GETATTR(outvp, &va, outcred);
3322 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3323 		    *outoffp + len) {
3324 #ifdef MAC
3325 			error = mac_vnode_check_write(curthread->td_ucred,
3326 			    outcred, outvp);
3327 			if (error == 0)
3328 #endif
3329 				error = vn_truncate_locked(outvp, *outoffp,
3330 				    false, outcred);
3331 			if (error == 0)
3332 				va.va_size = *outoffp;
3333 		}
3334 		VOP_UNLOCK(outvp);
3335 	}
3336 	if (mp != NULL)
3337 		vn_finished_write(mp);
3338 	if (error != 0)
3339 		goto out;
3340 
3341 	if (holein == 0 && holeout > 0) {
3342 		/*
3343 		 * For this special case, the input data will be scanned
3344 		 * for blocks of all 0 bytes.  For these blocks, the
3345 		 * write can be skipped for the output file to create
3346 		 * an unallocated region.
3347 		 * Therefore, use the appropriate size for the output file.
3348 		 */
3349 		blksize = holeout;
3350 		if (blksize <= 512) {
3351 			/*
3352 			 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3353 			 * of 512, although it actually only creates
3354 			 * unallocated regions for blocks >= f_iosize.
3355 			 */
3356 			blksize = outvp->v_mount->mnt_stat.f_iosize;
3357 		}
3358 	} else {
3359 		/*
3360 		 * Use the larger of the two f_iosize values.  If they are
3361 		 * not the same size, one will normally be an exact multiple of
3362 		 * the other, since they are both likely to be a power of 2.
3363 		 */
3364 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3365 		    outvp->v_mount->mnt_stat.f_iosize);
3366 	}
3367 
3368 	/* Clip to sane limits. */
3369 	if (blksize < 4096)
3370 		blksize = 4096;
3371 	else if (blksize > maxphys)
3372 		blksize = maxphys;
3373 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3374 
3375 	/*
3376 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3377 	 * to find holes.  Otherwise, just scan the read block for all 0s
3378 	 * in the inner loop where the data copying is done.
3379 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3380 	 * support holes on the server, but do not support FIOSEEKHOLE.
3381 	 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3382 	 * that this function should return after 1second with a partial
3383 	 * completion.
3384 	 */
3385 	if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3386 		getnanouptime(&endts);
3387 		endts.tv_sec++;
3388 	} else
3389 		timespecclear(&endts);
3390 	holetoeof = eof = false;
3391 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3392 		endoff = 0;			/* To shut up compilers. */
3393 		cantseek = true;
3394 		startoff = *inoffp;
3395 		copylen = len;
3396 
3397 		/*
3398 		 * Find the next data area.  If there is just a hole to EOF,
3399 		 * FIOSEEKDATA should fail with ENXIO.
3400 		 * (I do not know if any file system will report a hole to
3401 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3402 		 *  will fail for those file systems.)
3403 		 *
3404 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3405 		 * the code just falls through to the inner copy loop.
3406 		 */
3407 		error = EINVAL;
3408 		if (holein > 0) {
3409 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3410 			    incred, curthread);
3411 			if (error == ENXIO) {
3412 				startoff = endoff = inva.va_size;
3413 				eof = holetoeof = true;
3414 				error = 0;
3415 			}
3416 		}
3417 		if (error == 0 && !holetoeof) {
3418 			endoff = startoff;
3419 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3420 			    incred, curthread);
3421 			/*
3422 			 * Since invp is unlocked, it may be possible for
3423 			 * another thread to do a truncate(), lseek(), write()
3424 			 * creating a hole at startoff between the above
3425 			 * VOP_IOCTL() calls, if the other thread does not do
3426 			 * rangelocking.
3427 			 * If that happens, startoff == endoff and finding
3428 			 * the hole has failed, so set an error.
3429 			 */
3430 			if (error == 0 && startoff == endoff)
3431 				error = EINVAL; /* Any error. Reset to 0. */
3432 		}
3433 		if (error == 0) {
3434 			if (startoff > *inoffp) {
3435 				/* Found hole before data block. */
3436 				xfer = MIN(startoff - *inoffp, len);
3437 				if (*outoffp < va.va_size) {
3438 					/* Must write 0s to punch hole. */
3439 					xfer2 = MIN(va.va_size - *outoffp,
3440 					    xfer);
3441 					memset(dat, 0, MIN(xfer2, blksize));
3442 					error = vn_write_outvp(outvp, dat,
3443 					    *outoffp, xfer2, blksize, false,
3444 					    holeout > 0, outcred);
3445 				}
3446 
3447 				if (error == 0 && *outoffp + xfer >
3448 				    va.va_size && (xfer == len || holetoeof)) {
3449 					/* Grow output file (hole at end). */
3450 					error = vn_write_outvp(outvp, dat,
3451 					    *outoffp, xfer, blksize, true,
3452 					    false, outcred);
3453 				}
3454 				if (error == 0) {
3455 					*inoffp += xfer;
3456 					*outoffp += xfer;
3457 					len -= xfer;
3458 					if (len < savlen) {
3459 						interrupted = sig_intr();
3460 						if (timespecisset(&endts) &&
3461 						    interrupted == 0) {
3462 							getnanouptime(&curts);
3463 							if (timespeccmp(&curts,
3464 							    &endts, >=))
3465 								interrupted =
3466 								    EINTR;
3467 						}
3468 					}
3469 				}
3470 			}
3471 			copylen = MIN(len, endoff - startoff);
3472 			cantseek = false;
3473 		} else {
3474 			cantseek = true;
3475 			startoff = *inoffp;
3476 			copylen = len;
3477 			error = 0;
3478 		}
3479 
3480 		xfer = blksize;
3481 		if (cantseek) {
3482 			/*
3483 			 * Set first xfer to end at a block boundary, so that
3484 			 * holes are more likely detected in the loop below via
3485 			 * the for all bytes 0 method.
3486 			 */
3487 			xfer -= (*inoffp % blksize);
3488 		}
3489 		/* Loop copying the data block. */
3490 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3491 			if (copylen < xfer)
3492 				xfer = copylen;
3493 			error = vn_lock(invp, LK_SHARED);
3494 			if (error != 0)
3495 				goto out;
3496 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3497 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3498 			    curthread->td_ucred, incred, &aresid,
3499 			    curthread);
3500 			VOP_UNLOCK(invp);
3501 			lastblock = false;
3502 			if (error == 0 && aresid > 0) {
3503 				/* Stop the copy at EOF on the input file. */
3504 				xfer -= aresid;
3505 				eof = true;
3506 				lastblock = true;
3507 			}
3508 			if (error == 0) {
3509 				/*
3510 				 * Skip the write for holes past the initial EOF
3511 				 * of the output file, unless this is the last
3512 				 * write of the output file at EOF.
3513 				 */
3514 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3515 				    false;
3516 				if (xfer == len)
3517 					lastblock = true;
3518 				if (!cantseek || *outoffp < va.va_size ||
3519 				    lastblock || !readzeros)
3520 					error = vn_write_outvp(outvp, dat,
3521 					    *outoffp, xfer, blksize,
3522 					    readzeros && lastblock &&
3523 					    *outoffp >= va.va_size, false,
3524 					    outcred);
3525 				if (error == 0) {
3526 					*inoffp += xfer;
3527 					startoff += xfer;
3528 					*outoffp += xfer;
3529 					copylen -= xfer;
3530 					len -= xfer;
3531 					if (len < savlen) {
3532 						interrupted = sig_intr();
3533 						if (timespecisset(&endts) &&
3534 						    interrupted == 0) {
3535 							getnanouptime(&curts);
3536 							if (timespeccmp(&curts,
3537 							    &endts, >=))
3538 								interrupted =
3539 								    EINTR;
3540 						}
3541 					}
3542 				}
3543 			}
3544 			xfer = blksize;
3545 		}
3546 	}
3547 out:
3548 	*lenp = savlen - len;
3549 	free(dat, M_TEMP);
3550 	return (error);
3551 }
3552 
3553 static int
3554 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3555 {
3556 	struct mount *mp;
3557 	struct vnode *vp;
3558 	off_t olen, ooffset;
3559 	int error;
3560 #ifdef AUDIT
3561 	int audited_vnode1 = 0;
3562 #endif
3563 
3564 	vp = fp->f_vnode;
3565 	if (vp->v_type != VREG)
3566 		return (ENODEV);
3567 
3568 	/* Allocating blocks may take a long time, so iterate. */
3569 	for (;;) {
3570 		olen = len;
3571 		ooffset = offset;
3572 
3573 		bwillwrite();
3574 		mp = NULL;
3575 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
3576 		if (error != 0)
3577 			break;
3578 		error = vn_lock(vp, LK_EXCLUSIVE);
3579 		if (error != 0) {
3580 			vn_finished_write(mp);
3581 			break;
3582 		}
3583 #ifdef AUDIT
3584 		if (!audited_vnode1) {
3585 			AUDIT_ARG_VNODE1(vp);
3586 			audited_vnode1 = 1;
3587 		}
3588 #endif
3589 #ifdef MAC
3590 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3591 		if (error == 0)
3592 #endif
3593 			error = VOP_ALLOCATE(vp, &offset, &len, 0,
3594 			    td->td_ucred);
3595 		VOP_UNLOCK(vp);
3596 		vn_finished_write(mp);
3597 
3598 		if (olen + ooffset != offset + len) {
3599 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3600 			    ooffset, olen, offset, len);
3601 		}
3602 		if (error != 0 || len == 0)
3603 			break;
3604 		KASSERT(olen > len, ("Iteration did not make progress?"));
3605 		maybe_yield();
3606 	}
3607 
3608 	return (error);
3609 }
3610 
3611 static int
3612 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3613     int ioflag, struct ucred *cred, struct ucred *active_cred,
3614     struct ucred *file_cred)
3615 {
3616 	struct mount *mp;
3617 	void *rl_cookie;
3618 	off_t off, len;
3619 	int error;
3620 #ifdef AUDIT
3621 	bool audited_vnode1 = false;
3622 #endif
3623 
3624 	rl_cookie = NULL;
3625 	error = 0;
3626 	mp = NULL;
3627 	off = *offset;
3628 	len = *length;
3629 
3630 	if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3631 		rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3632 	while (len > 0 && error == 0) {
3633 		/*
3634 		 * Try to deallocate the longest range in one pass.
3635 		 * In case a pass takes too long to be executed, it returns
3636 		 * partial result. The residue will be proceeded in the next
3637 		 * pass.
3638 		 */
3639 
3640 		if ((ioflag & IO_NODELOCKED) == 0) {
3641 			bwillwrite();
3642 			if ((error = vn_start_write(vp, &mp,
3643 			    V_WAIT | V_PCATCH)) != 0)
3644 				goto out;
3645 			vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3646 		}
3647 #ifdef AUDIT
3648 		if (!audited_vnode1) {
3649 			AUDIT_ARG_VNODE1(vp);
3650 			audited_vnode1 = true;
3651 		}
3652 #endif
3653 
3654 #ifdef MAC
3655 		if ((ioflag & IO_NOMACCHECK) == 0)
3656 			error = mac_vnode_check_write(active_cred, file_cred,
3657 			    vp);
3658 #endif
3659 		if (error == 0)
3660 			error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3661 			    cred);
3662 
3663 		if ((ioflag & IO_NODELOCKED) == 0) {
3664 			VOP_UNLOCK(vp);
3665 			if (mp != NULL) {
3666 				vn_finished_write(mp);
3667 				mp = NULL;
3668 			}
3669 		}
3670 		if (error == 0 && len != 0)
3671 			maybe_yield();
3672 	}
3673 out:
3674 	if (rl_cookie != NULL)
3675 		vn_rangelock_unlock(vp, rl_cookie);
3676 	*offset = off;
3677 	*length = len;
3678 	return (error);
3679 }
3680 
3681 /*
3682  * This function is supposed to be used in the situations where the deallocation
3683  * is not triggered by a user request.
3684  */
3685 int
3686 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3687     int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3688 {
3689 	struct ucred *cred;
3690 
3691 	if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3692 	    flags != 0)
3693 		return (EINVAL);
3694 	if (vp->v_type != VREG)
3695 		return (ENODEV);
3696 
3697 	cred = file_cred != NOCRED ? file_cred : active_cred;
3698 	return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3699 	    active_cred, file_cred));
3700 }
3701 
3702 static int
3703 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3704     struct ucred *active_cred, struct thread *td)
3705 {
3706 	int error;
3707 	struct vnode *vp;
3708 	int ioflag;
3709 
3710 	KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd"));
3711 	KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
3712 	    ("vn_fspacectl: non-zero flags"));
3713 	KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
3714 	    ("vn_fspacectl: offset/length overflow or underflow"));
3715 	vp = fp->f_vnode;
3716 
3717 	if (vp->v_type != VREG)
3718 		return (ENODEV);
3719 
3720 	ioflag = get_write_ioflag(fp);
3721 
3722 	switch (cmd) {
3723 	case SPACECTL_DEALLOC:
3724 		error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3725 		    active_cred, active_cred, fp->f_cred);
3726 		break;
3727 	default:
3728 		panic("vn_fspacectl: unknown cmd %d", cmd);
3729 	}
3730 
3731 	return (error);
3732 }
3733 
3734 static u_long vn_lock_pair_pause_cnt;
3735 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3736     &vn_lock_pair_pause_cnt, 0,
3737     "Count of vn_lock_pair deadlocks");
3738 
3739 u_int vn_lock_pair_pause_max;
3740 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3741     &vn_lock_pair_pause_max, 0,
3742     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3743 
3744 static void
3745 vn_lock_pair_pause(const char *wmesg)
3746 {
3747 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3748 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3749 }
3750 
3751 /*
3752  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3753  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3754  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3755  * can be NULL.
3756  *
3757  * The function returns with both vnodes exclusively locked, and
3758  * guarantees that it does not create lock order reversal with other
3759  * threads during its execution.  Both vnodes could be unlocked
3760  * temporary (and reclaimed).
3761  */
3762 void
3763 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3764     bool vp2_locked)
3765 {
3766 	int error;
3767 
3768 	if (vp1 == NULL && vp2 == NULL)
3769 		return;
3770 	if (vp1 != NULL) {
3771 		if (vp1_locked)
3772 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3773 		else
3774 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3775 	} else {
3776 		vp1_locked = true;
3777 	}
3778 	if (vp2 != NULL) {
3779 		if (vp2_locked)
3780 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3781 		else
3782 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3783 	} else {
3784 		vp2_locked = true;
3785 	}
3786 	if (!vp1_locked && !vp2_locked) {
3787 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3788 		vp1_locked = true;
3789 	}
3790 
3791 	for (;;) {
3792 		if (vp1_locked && vp2_locked)
3793 			break;
3794 		if (vp1_locked && vp2 != NULL) {
3795 			if (vp1 != NULL) {
3796 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3797 				    __FILE__, __LINE__);
3798 				if (error == 0)
3799 					break;
3800 				VOP_UNLOCK(vp1);
3801 				vp1_locked = false;
3802 				vn_lock_pair_pause("vlp1");
3803 			}
3804 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3805 			vp2_locked = true;
3806 		}
3807 		if (vp2_locked && vp1 != NULL) {
3808 			if (vp2 != NULL) {
3809 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3810 				    __FILE__, __LINE__);
3811 				if (error == 0)
3812 					break;
3813 				VOP_UNLOCK(vp2);
3814 				vp2_locked = false;
3815 				vn_lock_pair_pause("vlp2");
3816 			}
3817 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3818 			vp1_locked = true;
3819 		}
3820 	}
3821 	if (vp1 != NULL)
3822 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3823 	if (vp2 != NULL)
3824 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3825 }
3826 
3827 int
3828 vn_lktype_write(struct mount *mp, struct vnode *vp)
3829 {
3830 	if (MNT_SHARED_WRITES(mp) ||
3831 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3832 		return (LK_SHARED);
3833 	return (LK_EXCLUSIVE);
3834 }
3835