xref: /freebsd/sys/kern/vfs_vnops.c (revision b51f459a2098622c31ed54f5c1bf0e03efce403b)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/prng.h>
74 #include <sys/sx.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
78 #include <sys/conf.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
81 #include <sys/user.h>
82 
83 #include <security/audit/audit.h>
84 #include <security/mac/mac_framework.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_extern.h>
88 #include <vm/pmap.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pager.h>
93 
94 #ifdef HWPMC_HOOKS
95 #include <sys/pmckern.h>
96 #endif
97 
98 static fo_rdwr_t	vn_read;
99 static fo_rdwr_t	vn_write;
100 static fo_rdwr_t	vn_io_fault;
101 static fo_truncate_t	vn_truncate;
102 static fo_ioctl_t	vn_ioctl;
103 static fo_poll_t	vn_poll;
104 static fo_kqfilter_t	vn_kqfilter;
105 static fo_close_t	vn_closefile;
106 static fo_mmap_t	vn_mmap;
107 static fo_fallocate_t	vn_fallocate;
108 
109 struct 	fileops vnops = {
110 	.fo_read = vn_io_fault,
111 	.fo_write = vn_io_fault,
112 	.fo_truncate = vn_truncate,
113 	.fo_ioctl = vn_ioctl,
114 	.fo_poll = vn_poll,
115 	.fo_kqfilter = vn_kqfilter,
116 	.fo_stat = vn_statfile,
117 	.fo_close = vn_closefile,
118 	.fo_chmod = vn_chmod,
119 	.fo_chown = vn_chown,
120 	.fo_sendfile = vn_sendfile,
121 	.fo_seek = vn_seek,
122 	.fo_fill_kinfo = vn_fill_kinfo,
123 	.fo_mmap = vn_mmap,
124 	.fo_fallocate = vn_fallocate,
125 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
126 };
127 
128 const u_int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
131     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
132 static int vn_io_fault_prefault = 0;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
134     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static int vn_io_pgcache_read_enable = 1;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
137     &vn_io_pgcache_read_enable, 0,
138     "Enable copying from page cache for reads, avoiding fs");
139 static u_long vn_io_faults_cnt;
140 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
141     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
142 
143 static int vfs_allow_read_dir = 0;
144 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
145     &vfs_allow_read_dir, 0,
146     "Enable read(2) of directory by root for filesystems that support it");
147 
148 /*
149  * Returns true if vn_io_fault mode of handling the i/o request should
150  * be used.
151  */
152 static bool
153 do_vn_io_fault(struct vnode *vp, struct uio *uio)
154 {
155 	struct mount *mp;
156 
157 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
158 	    (mp = vp->v_mount) != NULL &&
159 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
160 }
161 
162 /*
163  * Structure used to pass arguments to vn_io_fault1(), to do either
164  * file- or vnode-based I/O calls.
165  */
166 struct vn_io_fault_args {
167 	enum {
168 		VN_IO_FAULT_FOP,
169 		VN_IO_FAULT_VOP
170 	} kind;
171 	struct ucred *cred;
172 	int flags;
173 	union {
174 		struct fop_args_tag {
175 			struct file *fp;
176 			fo_rdwr_t *doio;
177 		} fop_args;
178 		struct vop_args_tag {
179 			struct vnode *vp;
180 		} vop_args;
181 	} args;
182 };
183 
184 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
185     struct vn_io_fault_args *args, struct thread *td);
186 
187 int
188 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
189 {
190 	struct thread *td = ndp->ni_cnd.cn_thread;
191 
192 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
193 }
194 
195 static uint64_t
196 open2nameif(int fmode, u_int vn_open_flags)
197 {
198 	uint64_t res;
199 
200 	res = ISOPEN | LOCKLEAF;
201 	if ((fmode & O_RESOLVE_BENEATH) != 0)
202 		res |= RBENEATH;
203 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
204 		res |= AUDITVNODE1;
205 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
206 		res |= NOCAPCHECK;
207 	return (res);
208 }
209 
210 /*
211  * Common code for vnode open operations via a name lookup.
212  * Lookup the vnode and invoke VOP_CREATE if needed.
213  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
214  *
215  * Note that this does NOT free nameidata for the successful case,
216  * due to the NDINIT being done elsewhere.
217  */
218 int
219 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
220     struct ucred *cred, struct file *fp)
221 {
222 	struct vnode *vp;
223 	struct mount *mp;
224 	struct thread *td = ndp->ni_cnd.cn_thread;
225 	struct vattr vat;
226 	struct vattr *vap = &vat;
227 	int fmode, error;
228 	bool first_open;
229 
230 restart:
231 	first_open = false;
232 	fmode = *flagp;
233 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
234 	    O_EXCL | O_DIRECTORY))
235 		return (EINVAL);
236 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
237 		ndp->ni_cnd.cn_nameiop = CREATE;
238 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
239 		/*
240 		 * Set NOCACHE to avoid flushing the cache when
241 		 * rolling in many files at once.
242 		 *
243 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
244 		 * exist despite NOCACHE.
245 		 */
246 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
247 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
248 			ndp->ni_cnd.cn_flags |= FOLLOW;
249 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
250 			bwillwrite();
251 		if ((error = namei(ndp)) != 0)
252 			return (error);
253 		if (ndp->ni_vp == NULL) {
254 			VATTR_NULL(vap);
255 			vap->va_type = VREG;
256 			vap->va_mode = cmode;
257 			if (fmode & O_EXCL)
258 				vap->va_vaflags |= VA_EXCLUSIVE;
259 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
260 				NDFREE(ndp, NDF_ONLY_PNBUF);
261 				vput(ndp->ni_dvp);
262 				if ((error = vn_start_write(NULL, &mp,
263 				    V_XSLEEP | PCATCH)) != 0)
264 					return (error);
265 				NDREINIT(ndp);
266 				goto restart;
267 			}
268 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
269 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
270 #ifdef MAC
271 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
272 			    &ndp->ni_cnd, vap);
273 			if (error == 0)
274 #endif
275 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
276 				    &ndp->ni_cnd, vap);
277 			vp = ndp->ni_vp;
278 			if (error == 0 && (fmode & O_EXCL) != 0 &&
279 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
280 				VI_LOCK(vp);
281 				vp->v_iflag |= VI_FOPENING;
282 				VI_UNLOCK(vp);
283 				first_open = true;
284 			}
285 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
286 			    false);
287 			vn_finished_write(mp);
288 			if (error) {
289 				NDFREE(ndp, NDF_ONLY_PNBUF);
290 				if (error == ERELOOKUP) {
291 					NDREINIT(ndp);
292 					goto restart;
293 				}
294 				return (error);
295 			}
296 			fmode &= ~O_TRUNC;
297 		} else {
298 			if (ndp->ni_dvp == ndp->ni_vp)
299 				vrele(ndp->ni_dvp);
300 			else
301 				vput(ndp->ni_dvp);
302 			ndp->ni_dvp = NULL;
303 			vp = ndp->ni_vp;
304 			if (fmode & O_EXCL) {
305 				error = EEXIST;
306 				goto bad;
307 			}
308 			if (vp->v_type == VDIR) {
309 				error = EISDIR;
310 				goto bad;
311 			}
312 			fmode &= ~O_CREAT;
313 		}
314 	} else {
315 		ndp->ni_cnd.cn_nameiop = LOOKUP;
316 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
317 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
318 		    FOLLOW;
319 		if ((fmode & FWRITE) == 0)
320 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
321 		if ((error = namei(ndp)) != 0)
322 			return (error);
323 		vp = ndp->ni_vp;
324 	}
325 	error = vn_open_vnode(vp, fmode, cred, td, fp);
326 	if (first_open) {
327 		VI_LOCK(vp);
328 		vp->v_iflag &= ~VI_FOPENING;
329 		wakeup(vp);
330 		VI_UNLOCK(vp);
331 	}
332 	if (error)
333 		goto bad;
334 	*flagp = fmode;
335 	return (0);
336 bad:
337 	NDFREE(ndp, NDF_ONLY_PNBUF);
338 	vput(vp);
339 	*flagp = fmode;
340 	ndp->ni_vp = NULL;
341 	return (error);
342 }
343 
344 static int
345 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
346 {
347 	struct flock lf;
348 	int error, lock_flags, type;
349 
350 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
351 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
352 		return (0);
353 	KASSERT(fp != NULL, ("open with flock requires fp"));
354 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
355 		return (EOPNOTSUPP);
356 
357 	lock_flags = VOP_ISLOCKED(vp);
358 	VOP_UNLOCK(vp);
359 
360 	lf.l_whence = SEEK_SET;
361 	lf.l_start = 0;
362 	lf.l_len = 0;
363 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
364 	type = F_FLOCK;
365 	if ((fmode & FNONBLOCK) == 0)
366 		type |= F_WAIT;
367 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
368 		type |= F_FIRSTOPEN;
369 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
370 	if (error == 0)
371 		fp->f_flag |= FHASLOCK;
372 
373 	vn_lock(vp, lock_flags | LK_RETRY);
374 	return (error);
375 }
376 
377 /*
378  * Common code for vnode open operations once a vnode is located.
379  * Check permissions, and call the VOP_OPEN routine.
380  */
381 int
382 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
383     struct thread *td, struct file *fp)
384 {
385 	accmode_t accmode;
386 	int error;
387 
388 	if (vp->v_type == VLNK) {
389 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
390 			return (EMLINK);
391 	}
392 	if (vp->v_type == VSOCK)
393 		return (EOPNOTSUPP);
394 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
395 		return (ENOTDIR);
396 
397 	accmode = 0;
398 	if ((fmode & O_PATH) == 0) {
399 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
400 			if (vp->v_type == VDIR)
401 				return (EISDIR);
402 			accmode |= VWRITE;
403 		}
404 		if ((fmode & FREAD) != 0)
405 			accmode |= VREAD;
406 		if ((fmode & O_APPEND) && (fmode & FWRITE))
407 			accmode |= VAPPEND;
408 #ifdef MAC
409 		if ((fmode & O_CREAT) != 0)
410 			accmode |= VCREAT;
411 #endif
412 	}
413 	if ((fmode & FEXEC) != 0)
414 		accmode |= VEXEC;
415 #ifdef MAC
416 	if ((fmode & O_VERIFY) != 0)
417 		accmode |= VVERIFY;
418 	error = mac_vnode_check_open(cred, vp, accmode);
419 	if (error != 0)
420 		return (error);
421 
422 	accmode &= ~(VCREAT | VVERIFY);
423 #endif
424 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
425 		error = VOP_ACCESS(vp, accmode, cred, td);
426 		if (error != 0)
427 			return (error);
428 	}
429 	if ((fmode & O_PATH) != 0) {
430 		error = VOP_ACCESS(vp, VREAD, cred, td);
431 		if (error == 0)
432 			fp->f_flag |= FKQALLOWED;
433 		return (0);
434 	}
435 
436 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
437 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
438 	error = VOP_OPEN(vp, fmode, cred, td, fp);
439 	if (error != 0)
440 		return (error);
441 
442 	error = vn_open_vnode_advlock(vp, fmode, fp);
443 	if (error == 0 && (fmode & FWRITE) != 0) {
444 		error = VOP_ADD_WRITECOUNT(vp, 1);
445 		if (error == 0) {
446 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
447 			     __func__, vp, vp->v_writecount);
448 		}
449 	}
450 
451 	/*
452 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
453 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
454 	 * Arrange for that by having fdrop() to use vn_closefile().
455 	 */
456 	if (error != 0) {
457 		fp->f_flag |= FOPENFAILED;
458 		fp->f_vnode = vp;
459 		if (fp->f_ops == &badfileops) {
460 			fp->f_type = DTYPE_VNODE;
461 			fp->f_ops = &vnops;
462 		}
463 		vref(vp);
464 	}
465 
466 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
467 	return (error);
468 
469 }
470 
471 /*
472  * Check for write permissions on the specified vnode.
473  * Prototype text segments cannot be written.
474  * It is racy.
475  */
476 int
477 vn_writechk(struct vnode *vp)
478 {
479 
480 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
481 	/*
482 	 * If there's shared text associated with
483 	 * the vnode, try to free it up once.  If
484 	 * we fail, we can't allow writing.
485 	 */
486 	if (VOP_IS_TEXT(vp))
487 		return (ETXTBSY);
488 
489 	return (0);
490 }
491 
492 /*
493  * Vnode close call
494  */
495 static int
496 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
497     struct thread *td, bool keep_ref)
498 {
499 	struct mount *mp;
500 	int error, lock_flags;
501 
502 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
503 	    MNT_EXTENDED_SHARED(vp->v_mount))
504 		lock_flags = LK_SHARED;
505 	else
506 		lock_flags = LK_EXCLUSIVE;
507 
508 	vn_start_write(vp, &mp, V_WAIT);
509 	vn_lock(vp, lock_flags | LK_RETRY);
510 	AUDIT_ARG_VNODE1(vp);
511 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
512 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
513 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
514 		    __func__, vp, vp->v_writecount);
515 	}
516 	error = VOP_CLOSE(vp, flags, file_cred, td);
517 	if (keep_ref)
518 		VOP_UNLOCK(vp);
519 	else
520 		vput(vp);
521 	vn_finished_write(mp);
522 	return (error);
523 }
524 
525 int
526 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
527     struct thread *td)
528 {
529 
530 	return (vn_close1(vp, flags, file_cred, td, false));
531 }
532 
533 /*
534  * Heuristic to detect sequential operation.
535  */
536 static int
537 sequential_heuristic(struct uio *uio, struct file *fp)
538 {
539 	enum uio_rw rw;
540 
541 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
542 
543 	rw = uio->uio_rw;
544 	if (fp->f_flag & FRDAHEAD)
545 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
546 
547 	/*
548 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
549 	 * that the first I/O is normally considered to be slightly
550 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
551 	 * unless previous seeks have reduced f_seqcount to 0, in which
552 	 * case offset 0 is not special.
553 	 */
554 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
555 	    uio->uio_offset == fp->f_nextoff[rw]) {
556 		/*
557 		 * f_seqcount is in units of fixed-size blocks so that it
558 		 * depends mainly on the amount of sequential I/O and not
559 		 * much on the number of sequential I/O's.  The fixed size
560 		 * of 16384 is hard-coded here since it is (not quite) just
561 		 * a magic size that works well here.  This size is more
562 		 * closely related to the best I/O size for real disks than
563 		 * to any block size used by software.
564 		 */
565 		if (uio->uio_resid >= IO_SEQMAX * 16384)
566 			fp->f_seqcount[rw] = IO_SEQMAX;
567 		else {
568 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
569 			if (fp->f_seqcount[rw] > IO_SEQMAX)
570 				fp->f_seqcount[rw] = IO_SEQMAX;
571 		}
572 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
573 	}
574 
575 	/* Not sequential.  Quickly draw-down sequentiality. */
576 	if (fp->f_seqcount[rw] > 1)
577 		fp->f_seqcount[rw] = 1;
578 	else
579 		fp->f_seqcount[rw] = 0;
580 	return (0);
581 }
582 
583 /*
584  * Package up an I/O request on a vnode into a uio and do it.
585  */
586 int
587 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
588     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
589     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
590 {
591 	struct uio auio;
592 	struct iovec aiov;
593 	struct mount *mp;
594 	struct ucred *cred;
595 	void *rl_cookie;
596 	struct vn_io_fault_args args;
597 	int error, lock_flags;
598 
599 	if (offset < 0 && vp->v_type != VCHR)
600 		return (EINVAL);
601 	auio.uio_iov = &aiov;
602 	auio.uio_iovcnt = 1;
603 	aiov.iov_base = base;
604 	aiov.iov_len = len;
605 	auio.uio_resid = len;
606 	auio.uio_offset = offset;
607 	auio.uio_segflg = segflg;
608 	auio.uio_rw = rw;
609 	auio.uio_td = td;
610 	error = 0;
611 
612 	if ((ioflg & IO_NODELOCKED) == 0) {
613 		if ((ioflg & IO_RANGELOCKED) == 0) {
614 			if (rw == UIO_READ) {
615 				rl_cookie = vn_rangelock_rlock(vp, offset,
616 				    offset + len);
617 			} else if ((ioflg & IO_APPEND) != 0) {
618 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
619 			} else {
620 				rl_cookie = vn_rangelock_wlock(vp, offset,
621 				    offset + len);
622 			}
623 		} else
624 			rl_cookie = NULL;
625 		mp = NULL;
626 		if (rw == UIO_WRITE) {
627 			if (vp->v_type != VCHR &&
628 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
629 			    != 0)
630 				goto out;
631 			if (MNT_SHARED_WRITES(mp) ||
632 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
633 				lock_flags = LK_SHARED;
634 			else
635 				lock_flags = LK_EXCLUSIVE;
636 		} else
637 			lock_flags = LK_SHARED;
638 		vn_lock(vp, lock_flags | LK_RETRY);
639 	} else
640 		rl_cookie = NULL;
641 
642 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
643 #ifdef MAC
644 	if ((ioflg & IO_NOMACCHECK) == 0) {
645 		if (rw == UIO_READ)
646 			error = mac_vnode_check_read(active_cred, file_cred,
647 			    vp);
648 		else
649 			error = mac_vnode_check_write(active_cred, file_cred,
650 			    vp);
651 	}
652 #endif
653 	if (error == 0) {
654 		if (file_cred != NULL)
655 			cred = file_cred;
656 		else
657 			cred = active_cred;
658 		if (do_vn_io_fault(vp, &auio)) {
659 			args.kind = VN_IO_FAULT_VOP;
660 			args.cred = cred;
661 			args.flags = ioflg;
662 			args.args.vop_args.vp = vp;
663 			error = vn_io_fault1(vp, &auio, &args, td);
664 		} else if (rw == UIO_READ) {
665 			error = VOP_READ(vp, &auio, ioflg, cred);
666 		} else /* if (rw == UIO_WRITE) */ {
667 			error = VOP_WRITE(vp, &auio, ioflg, cred);
668 		}
669 	}
670 	if (aresid)
671 		*aresid = auio.uio_resid;
672 	else
673 		if (auio.uio_resid && error == 0)
674 			error = EIO;
675 	if ((ioflg & IO_NODELOCKED) == 0) {
676 		VOP_UNLOCK(vp);
677 		if (mp != NULL)
678 			vn_finished_write(mp);
679 	}
680  out:
681 	if (rl_cookie != NULL)
682 		vn_rangelock_unlock(vp, rl_cookie);
683 	return (error);
684 }
685 
686 /*
687  * Package up an I/O request on a vnode into a uio and do it.  The I/O
688  * request is split up into smaller chunks and we try to avoid saturating
689  * the buffer cache while potentially holding a vnode locked, so we
690  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
691  * to give other processes a chance to lock the vnode (either other processes
692  * core'ing the same binary, or unrelated processes scanning the directory).
693  */
694 int
695 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
696     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
697     struct ucred *file_cred, size_t *aresid, struct thread *td)
698 {
699 	int error = 0;
700 	ssize_t iaresid;
701 
702 	do {
703 		int chunk;
704 
705 		/*
706 		 * Force `offset' to a multiple of MAXBSIZE except possibly
707 		 * for the first chunk, so that filesystems only need to
708 		 * write full blocks except possibly for the first and last
709 		 * chunks.
710 		 */
711 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
712 
713 		if (chunk > len)
714 			chunk = len;
715 		if (rw != UIO_READ && vp->v_type == VREG)
716 			bwillwrite();
717 		iaresid = 0;
718 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
719 		    ioflg, active_cred, file_cred, &iaresid, td);
720 		len -= chunk;	/* aresid calc already includes length */
721 		if (error)
722 			break;
723 		offset += chunk;
724 		base = (char *)base + chunk;
725 		kern_yield(PRI_USER);
726 	} while (len);
727 	if (aresid)
728 		*aresid = len + iaresid;
729 	return (error);
730 }
731 
732 #if OFF_MAX <= LONG_MAX
733 off_t
734 foffset_lock(struct file *fp, int flags)
735 {
736 	volatile short *flagsp;
737 	off_t res;
738 	short state;
739 
740 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
741 
742 	if ((flags & FOF_NOLOCK) != 0)
743 		return (atomic_load_long(&fp->f_offset));
744 
745 	/*
746 	 * According to McKusick the vn lock was protecting f_offset here.
747 	 * It is now protected by the FOFFSET_LOCKED flag.
748 	 */
749 	flagsp = &fp->f_vnread_flags;
750 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
751 		return (atomic_load_long(&fp->f_offset));
752 
753 	sleepq_lock(&fp->f_vnread_flags);
754 	state = atomic_load_16(flagsp);
755 	for (;;) {
756 		if ((state & FOFFSET_LOCKED) == 0) {
757 			if (!atomic_fcmpset_acq_16(flagsp, &state,
758 			    FOFFSET_LOCKED))
759 				continue;
760 			break;
761 		}
762 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
763 			if (!atomic_fcmpset_acq_16(flagsp, &state,
764 			    state | FOFFSET_LOCK_WAITING))
765 				continue;
766 		}
767 		DROP_GIANT();
768 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
769 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
770 		PICKUP_GIANT();
771 		sleepq_lock(&fp->f_vnread_flags);
772 		state = atomic_load_16(flagsp);
773 	}
774 	res = atomic_load_long(&fp->f_offset);
775 	sleepq_release(&fp->f_vnread_flags);
776 	return (res);
777 }
778 
779 void
780 foffset_unlock(struct file *fp, off_t val, int flags)
781 {
782 	volatile short *flagsp;
783 	short state;
784 
785 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
786 
787 	if ((flags & FOF_NOUPDATE) == 0)
788 		atomic_store_long(&fp->f_offset, val);
789 	if ((flags & FOF_NEXTOFF_R) != 0)
790 		fp->f_nextoff[UIO_READ] = val;
791 	if ((flags & FOF_NEXTOFF_W) != 0)
792 		fp->f_nextoff[UIO_WRITE] = val;
793 
794 	if ((flags & FOF_NOLOCK) != 0)
795 		return;
796 
797 	flagsp = &fp->f_vnread_flags;
798 	state = atomic_load_16(flagsp);
799 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
800 	    atomic_cmpset_rel_16(flagsp, state, 0))
801 		return;
802 
803 	sleepq_lock(&fp->f_vnread_flags);
804 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
805 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
806 	fp->f_vnread_flags = 0;
807 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
808 	sleepq_release(&fp->f_vnread_flags);
809 }
810 #else
811 off_t
812 foffset_lock(struct file *fp, int flags)
813 {
814 	struct mtx *mtxp;
815 	off_t res;
816 
817 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
818 
819 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
820 	mtx_lock(mtxp);
821 	if ((flags & FOF_NOLOCK) == 0) {
822 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
823 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
824 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
825 			    "vofflock", 0);
826 		}
827 		fp->f_vnread_flags |= FOFFSET_LOCKED;
828 	}
829 	res = fp->f_offset;
830 	mtx_unlock(mtxp);
831 	return (res);
832 }
833 
834 void
835 foffset_unlock(struct file *fp, off_t val, int flags)
836 {
837 	struct mtx *mtxp;
838 
839 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
840 
841 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
842 	mtx_lock(mtxp);
843 	if ((flags & FOF_NOUPDATE) == 0)
844 		fp->f_offset = val;
845 	if ((flags & FOF_NEXTOFF_R) != 0)
846 		fp->f_nextoff[UIO_READ] = val;
847 	if ((flags & FOF_NEXTOFF_W) != 0)
848 		fp->f_nextoff[UIO_WRITE] = val;
849 	if ((flags & FOF_NOLOCK) == 0) {
850 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
851 		    ("Lost FOFFSET_LOCKED"));
852 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
853 			wakeup(&fp->f_vnread_flags);
854 		fp->f_vnread_flags = 0;
855 	}
856 	mtx_unlock(mtxp);
857 }
858 #endif
859 
860 void
861 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
862 {
863 
864 	if ((flags & FOF_OFFSET) == 0)
865 		uio->uio_offset = foffset_lock(fp, flags);
866 }
867 
868 void
869 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
870 {
871 
872 	if ((flags & FOF_OFFSET) == 0)
873 		foffset_unlock(fp, uio->uio_offset, flags);
874 }
875 
876 static int
877 get_advice(struct file *fp, struct uio *uio)
878 {
879 	struct mtx *mtxp;
880 	int ret;
881 
882 	ret = POSIX_FADV_NORMAL;
883 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
884 		return (ret);
885 
886 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
887 	mtx_lock(mtxp);
888 	if (fp->f_advice != NULL &&
889 	    uio->uio_offset >= fp->f_advice->fa_start &&
890 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
891 		ret = fp->f_advice->fa_advice;
892 	mtx_unlock(mtxp);
893 	return (ret);
894 }
895 
896 int
897 vn_read_from_obj(struct vnode *vp, struct uio *uio)
898 {
899 	vm_object_t obj;
900 	vm_page_t ma[io_hold_cnt + 2];
901 	off_t off, vsz;
902 	ssize_t resid;
903 	int error, i, j;
904 
905 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
906 	obj = atomic_load_ptr(&vp->v_object);
907 	if (obj == NULL)
908 		return (EJUSTRETURN);
909 
910 	/*
911 	 * Depends on type stability of vm_objects.
912 	 */
913 	vm_object_pip_add(obj, 1);
914 	if ((obj->flags & OBJ_DEAD) != 0) {
915 		/*
916 		 * Note that object might be already reused from the
917 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
918 		 * we recheck for DOOMED vnode state after all pages
919 		 * are busied, and retract then.
920 		 *
921 		 * But we check for OBJ_DEAD to ensure that we do not
922 		 * busy pages while vm_object_terminate_pages()
923 		 * processes the queue.
924 		 */
925 		error = EJUSTRETURN;
926 		goto out_pip;
927 	}
928 
929 	resid = uio->uio_resid;
930 	off = uio->uio_offset;
931 	for (i = 0; resid > 0; i++) {
932 		MPASS(i < io_hold_cnt + 2);
933 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
934 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
935 		    VM_ALLOC_NOWAIT);
936 		if (ma[i] == NULL)
937 			break;
938 
939 		/*
940 		 * Skip invalid pages.  Valid mask can be partial only
941 		 * at EOF, and we clip later.
942 		 */
943 		if (vm_page_none_valid(ma[i])) {
944 			vm_page_sunbusy(ma[i]);
945 			break;
946 		}
947 
948 		resid -= PAGE_SIZE;
949 		off += PAGE_SIZE;
950 	}
951 	if (i == 0) {
952 		error = EJUSTRETURN;
953 		goto out_pip;
954 	}
955 
956 	/*
957 	 * Check VIRF_DOOMED after we busied our pages.  Since
958 	 * vgonel() terminates the vnode' vm_object, it cannot
959 	 * process past pages busied by us.
960 	 */
961 	if (VN_IS_DOOMED(vp)) {
962 		error = EJUSTRETURN;
963 		goto out;
964 	}
965 
966 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
967 	if (resid > uio->uio_resid)
968 		resid = uio->uio_resid;
969 
970 	/*
971 	 * Unlocked read of vnp_size is safe because truncation cannot
972 	 * pass busied page.  But we load vnp_size into a local
973 	 * variable so that possible concurrent extension does not
974 	 * break calculation.
975 	 */
976 #if defined(__powerpc__) && !defined(__powerpc64__)
977 	vsz = obj->un_pager.vnp.vnp_size;
978 #else
979 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
980 #endif
981 	if (uio->uio_offset >= vsz) {
982 		error = EJUSTRETURN;
983 		goto out;
984 	}
985 	if (uio->uio_offset + resid > vsz)
986 		resid = vsz - uio->uio_offset;
987 
988 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
989 
990 out:
991 	for (j = 0; j < i; j++) {
992 		if (error == 0)
993 			vm_page_reference(ma[j]);
994 		vm_page_sunbusy(ma[j]);
995 	}
996 out_pip:
997 	vm_object_pip_wakeup(obj);
998 	if (error != 0)
999 		return (error);
1000 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1001 }
1002 
1003 /*
1004  * File table vnode read routine.
1005  */
1006 static int
1007 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1008     struct thread *td)
1009 {
1010 	struct vnode *vp;
1011 	off_t orig_offset;
1012 	int error, ioflag;
1013 	int advice;
1014 
1015 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1016 	    uio->uio_td, td));
1017 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1018 	vp = fp->f_vnode;
1019 	ioflag = 0;
1020 	if (fp->f_flag & FNONBLOCK)
1021 		ioflag |= IO_NDELAY;
1022 	if (fp->f_flag & O_DIRECT)
1023 		ioflag |= IO_DIRECT;
1024 
1025 	/*
1026 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1027 	 * allows us to avoid unneeded work outright.
1028 	 */
1029 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1030 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1031 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1032 		if (error == 0) {
1033 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1034 			return (0);
1035 		}
1036 		if (error != EJUSTRETURN)
1037 			return (error);
1038 	}
1039 
1040 	advice = get_advice(fp, uio);
1041 	vn_lock(vp, LK_SHARED | LK_RETRY);
1042 
1043 	switch (advice) {
1044 	case POSIX_FADV_NORMAL:
1045 	case POSIX_FADV_SEQUENTIAL:
1046 	case POSIX_FADV_NOREUSE:
1047 		ioflag |= sequential_heuristic(uio, fp);
1048 		break;
1049 	case POSIX_FADV_RANDOM:
1050 		/* Disable read-ahead for random I/O. */
1051 		break;
1052 	}
1053 	orig_offset = uio->uio_offset;
1054 
1055 #ifdef MAC
1056 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1057 	if (error == 0)
1058 #endif
1059 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1060 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1061 	VOP_UNLOCK(vp);
1062 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1063 	    orig_offset != uio->uio_offset)
1064 		/*
1065 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1066 		 * for the backing file after a POSIX_FADV_NOREUSE
1067 		 * read(2).
1068 		 */
1069 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1070 		    POSIX_FADV_DONTNEED);
1071 	return (error);
1072 }
1073 
1074 /*
1075  * File table vnode write routine.
1076  */
1077 static int
1078 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1079     struct thread *td)
1080 {
1081 	struct vnode *vp;
1082 	struct mount *mp;
1083 	off_t orig_offset;
1084 	int error, ioflag, lock_flags;
1085 	int advice;
1086 
1087 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1088 	    uio->uio_td, td));
1089 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1090 	vp = fp->f_vnode;
1091 	if (vp->v_type == VREG)
1092 		bwillwrite();
1093 	ioflag = IO_UNIT;
1094 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1095 		ioflag |= IO_APPEND;
1096 	if (fp->f_flag & FNONBLOCK)
1097 		ioflag |= IO_NDELAY;
1098 	if (fp->f_flag & O_DIRECT)
1099 		ioflag |= IO_DIRECT;
1100 	if ((fp->f_flag & O_FSYNC) ||
1101 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1102 		ioflag |= IO_SYNC;
1103 	/*
1104 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1105 	 * implementations that don't understand IO_DATASYNC fall back to full
1106 	 * O_SYNC behavior.
1107 	 */
1108 	if (fp->f_flag & O_DSYNC)
1109 		ioflag |= IO_SYNC | IO_DATASYNC;
1110 	mp = NULL;
1111 	if (vp->v_type != VCHR &&
1112 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1113 		goto unlock;
1114 
1115 	advice = get_advice(fp, uio);
1116 
1117 	if (MNT_SHARED_WRITES(mp) ||
1118 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1119 		lock_flags = LK_SHARED;
1120 	} else {
1121 		lock_flags = LK_EXCLUSIVE;
1122 	}
1123 
1124 	vn_lock(vp, lock_flags | LK_RETRY);
1125 	switch (advice) {
1126 	case POSIX_FADV_NORMAL:
1127 	case POSIX_FADV_SEQUENTIAL:
1128 	case POSIX_FADV_NOREUSE:
1129 		ioflag |= sequential_heuristic(uio, fp);
1130 		break;
1131 	case POSIX_FADV_RANDOM:
1132 		/* XXX: Is this correct? */
1133 		break;
1134 	}
1135 	orig_offset = uio->uio_offset;
1136 
1137 #ifdef MAC
1138 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1139 	if (error == 0)
1140 #endif
1141 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1142 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1143 	VOP_UNLOCK(vp);
1144 	if (vp->v_type != VCHR)
1145 		vn_finished_write(mp);
1146 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1147 	    orig_offset != uio->uio_offset)
1148 		/*
1149 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1150 		 * for the backing file after a POSIX_FADV_NOREUSE
1151 		 * write(2).
1152 		 */
1153 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1154 		    POSIX_FADV_DONTNEED);
1155 unlock:
1156 	return (error);
1157 }
1158 
1159 /*
1160  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1161  * prevent the following deadlock:
1162  *
1163  * Assume that the thread A reads from the vnode vp1 into userspace
1164  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1165  * currently not resident, then system ends up with the call chain
1166  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1167  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1168  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1169  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1170  * backed by the pages of vnode vp1, and some page in buf2 is not
1171  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1172  *
1173  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1174  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1175  * Instead, it first tries to do the whole range i/o with pagefaults
1176  * disabled. If all pages in the i/o buffer are resident and mapped,
1177  * VOP will succeed (ignoring the genuine filesystem errors).
1178  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1179  * i/o in chunks, with all pages in the chunk prefaulted and held
1180  * using vm_fault_quick_hold_pages().
1181  *
1182  * Filesystems using this deadlock avoidance scheme should use the
1183  * array of the held pages from uio, saved in the curthread->td_ma,
1184  * instead of doing uiomove().  A helper function
1185  * vn_io_fault_uiomove() converts uiomove request into
1186  * uiomove_fromphys() over td_ma array.
1187  *
1188  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1189  * make the current i/o request atomic with respect to other i/os and
1190  * truncations.
1191  */
1192 
1193 /*
1194  * Decode vn_io_fault_args and perform the corresponding i/o.
1195  */
1196 static int
1197 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1198     struct thread *td)
1199 {
1200 	int error, save;
1201 
1202 	error = 0;
1203 	save = vm_fault_disable_pagefaults();
1204 	switch (args->kind) {
1205 	case VN_IO_FAULT_FOP:
1206 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1207 		    uio, args->cred, args->flags, td);
1208 		break;
1209 	case VN_IO_FAULT_VOP:
1210 		if (uio->uio_rw == UIO_READ) {
1211 			error = VOP_READ(args->args.vop_args.vp, uio,
1212 			    args->flags, args->cred);
1213 		} else if (uio->uio_rw == UIO_WRITE) {
1214 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1215 			    args->flags, args->cred);
1216 		}
1217 		break;
1218 	default:
1219 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1220 		    args->kind, uio->uio_rw);
1221 	}
1222 	vm_fault_enable_pagefaults(save);
1223 	return (error);
1224 }
1225 
1226 static int
1227 vn_io_fault_touch(char *base, const struct uio *uio)
1228 {
1229 	int r;
1230 
1231 	r = fubyte(base);
1232 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1233 		return (EFAULT);
1234 	return (0);
1235 }
1236 
1237 static int
1238 vn_io_fault_prefault_user(const struct uio *uio)
1239 {
1240 	char *base;
1241 	const struct iovec *iov;
1242 	size_t len;
1243 	ssize_t resid;
1244 	int error, i;
1245 
1246 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1247 	    ("vn_io_fault_prefault userspace"));
1248 
1249 	error = i = 0;
1250 	iov = uio->uio_iov;
1251 	resid = uio->uio_resid;
1252 	base = iov->iov_base;
1253 	len = iov->iov_len;
1254 	while (resid > 0) {
1255 		error = vn_io_fault_touch(base, uio);
1256 		if (error != 0)
1257 			break;
1258 		if (len < PAGE_SIZE) {
1259 			if (len != 0) {
1260 				error = vn_io_fault_touch(base + len - 1, uio);
1261 				if (error != 0)
1262 					break;
1263 				resid -= len;
1264 			}
1265 			if (++i >= uio->uio_iovcnt)
1266 				break;
1267 			iov = uio->uio_iov + i;
1268 			base = iov->iov_base;
1269 			len = iov->iov_len;
1270 		} else {
1271 			len -= PAGE_SIZE;
1272 			base += PAGE_SIZE;
1273 			resid -= PAGE_SIZE;
1274 		}
1275 	}
1276 	return (error);
1277 }
1278 
1279 /*
1280  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1281  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1282  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1283  * into args and call vn_io_fault1() to handle faults during the user
1284  * mode buffer accesses.
1285  */
1286 static int
1287 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1288     struct thread *td)
1289 {
1290 	vm_page_t ma[io_hold_cnt + 2];
1291 	struct uio *uio_clone, short_uio;
1292 	struct iovec short_iovec[1];
1293 	vm_page_t *prev_td_ma;
1294 	vm_prot_t prot;
1295 	vm_offset_t addr, end;
1296 	size_t len, resid;
1297 	ssize_t adv;
1298 	int error, cnt, saveheld, prev_td_ma_cnt;
1299 
1300 	if (vn_io_fault_prefault) {
1301 		error = vn_io_fault_prefault_user(uio);
1302 		if (error != 0)
1303 			return (error); /* Or ignore ? */
1304 	}
1305 
1306 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1307 
1308 	/*
1309 	 * The UFS follows IO_UNIT directive and replays back both
1310 	 * uio_offset and uio_resid if an error is encountered during the
1311 	 * operation.  But, since the iovec may be already advanced,
1312 	 * uio is still in an inconsistent state.
1313 	 *
1314 	 * Cache a copy of the original uio, which is advanced to the redo
1315 	 * point using UIO_NOCOPY below.
1316 	 */
1317 	uio_clone = cloneuio(uio);
1318 	resid = uio->uio_resid;
1319 
1320 	short_uio.uio_segflg = UIO_USERSPACE;
1321 	short_uio.uio_rw = uio->uio_rw;
1322 	short_uio.uio_td = uio->uio_td;
1323 
1324 	error = vn_io_fault_doio(args, uio, td);
1325 	if (error != EFAULT)
1326 		goto out;
1327 
1328 	atomic_add_long(&vn_io_faults_cnt, 1);
1329 	uio_clone->uio_segflg = UIO_NOCOPY;
1330 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1331 	uio_clone->uio_segflg = uio->uio_segflg;
1332 
1333 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1334 	prev_td_ma = td->td_ma;
1335 	prev_td_ma_cnt = td->td_ma_cnt;
1336 
1337 	while (uio_clone->uio_resid != 0) {
1338 		len = uio_clone->uio_iov->iov_len;
1339 		if (len == 0) {
1340 			KASSERT(uio_clone->uio_iovcnt >= 1,
1341 			    ("iovcnt underflow"));
1342 			uio_clone->uio_iov++;
1343 			uio_clone->uio_iovcnt--;
1344 			continue;
1345 		}
1346 		if (len > ptoa(io_hold_cnt))
1347 			len = ptoa(io_hold_cnt);
1348 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1349 		end = round_page(addr + len);
1350 		if (end < addr) {
1351 			error = EFAULT;
1352 			break;
1353 		}
1354 		cnt = atop(end - trunc_page(addr));
1355 		/*
1356 		 * A perfectly misaligned address and length could cause
1357 		 * both the start and the end of the chunk to use partial
1358 		 * page.  +2 accounts for such a situation.
1359 		 */
1360 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1361 		    addr, len, prot, ma, io_hold_cnt + 2);
1362 		if (cnt == -1) {
1363 			error = EFAULT;
1364 			break;
1365 		}
1366 		short_uio.uio_iov = &short_iovec[0];
1367 		short_iovec[0].iov_base = (void *)addr;
1368 		short_uio.uio_iovcnt = 1;
1369 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1370 		short_uio.uio_offset = uio_clone->uio_offset;
1371 		td->td_ma = ma;
1372 		td->td_ma_cnt = cnt;
1373 
1374 		error = vn_io_fault_doio(args, &short_uio, td);
1375 		vm_page_unhold_pages(ma, cnt);
1376 		adv = len - short_uio.uio_resid;
1377 
1378 		uio_clone->uio_iov->iov_base =
1379 		    (char *)uio_clone->uio_iov->iov_base + adv;
1380 		uio_clone->uio_iov->iov_len -= adv;
1381 		uio_clone->uio_resid -= adv;
1382 		uio_clone->uio_offset += adv;
1383 
1384 		uio->uio_resid -= adv;
1385 		uio->uio_offset += adv;
1386 
1387 		if (error != 0 || adv == 0)
1388 			break;
1389 	}
1390 	td->td_ma = prev_td_ma;
1391 	td->td_ma_cnt = prev_td_ma_cnt;
1392 	curthread_pflags_restore(saveheld);
1393 out:
1394 	free(uio_clone, M_IOV);
1395 	return (error);
1396 }
1397 
1398 static int
1399 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1400     int flags, struct thread *td)
1401 {
1402 	fo_rdwr_t *doio;
1403 	struct vnode *vp;
1404 	void *rl_cookie;
1405 	struct vn_io_fault_args args;
1406 	int error;
1407 
1408 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1409 	vp = fp->f_vnode;
1410 
1411 	/*
1412 	 * The ability to read(2) on a directory has historically been
1413 	 * allowed for all users, but this can and has been the source of
1414 	 * at least one security issue in the past.  As such, it is now hidden
1415 	 * away behind a sysctl for those that actually need it to use it, and
1416 	 * restricted to root when it's turned on to make it relatively safe to
1417 	 * leave on for longer sessions of need.
1418 	 */
1419 	if (vp->v_type == VDIR) {
1420 		KASSERT(uio->uio_rw == UIO_READ,
1421 		    ("illegal write attempted on a directory"));
1422 		if (!vfs_allow_read_dir)
1423 			return (EISDIR);
1424 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1425 			return (EISDIR);
1426 	}
1427 
1428 	foffset_lock_uio(fp, uio, flags);
1429 	if (do_vn_io_fault(vp, uio)) {
1430 		args.kind = VN_IO_FAULT_FOP;
1431 		args.args.fop_args.fp = fp;
1432 		args.args.fop_args.doio = doio;
1433 		args.cred = active_cred;
1434 		args.flags = flags | FOF_OFFSET;
1435 		if (uio->uio_rw == UIO_READ) {
1436 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1437 			    uio->uio_offset + uio->uio_resid);
1438 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1439 		    (flags & FOF_OFFSET) == 0) {
1440 			/* For appenders, punt and lock the whole range. */
1441 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1442 		} else {
1443 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1444 			    uio->uio_offset + uio->uio_resid);
1445 		}
1446 		error = vn_io_fault1(vp, uio, &args, td);
1447 		vn_rangelock_unlock(vp, rl_cookie);
1448 	} else {
1449 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1450 	}
1451 	foffset_unlock_uio(fp, uio, flags);
1452 	return (error);
1453 }
1454 
1455 /*
1456  * Helper function to perform the requested uiomove operation using
1457  * the held pages for io->uio_iov[0].iov_base buffer instead of
1458  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1459  * instead of iov_base prevents page faults that could occur due to
1460  * pmap_collect() invalidating the mapping created by
1461  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1462  * object cleanup revoking the write access from page mappings.
1463  *
1464  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1465  * instead of plain uiomove().
1466  */
1467 int
1468 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1469 {
1470 	struct uio transp_uio;
1471 	struct iovec transp_iov[1];
1472 	struct thread *td;
1473 	size_t adv;
1474 	int error, pgadv;
1475 
1476 	td = curthread;
1477 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1478 	    uio->uio_segflg != UIO_USERSPACE)
1479 		return (uiomove(data, xfersize, uio));
1480 
1481 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1482 	transp_iov[0].iov_base = data;
1483 	transp_uio.uio_iov = &transp_iov[0];
1484 	transp_uio.uio_iovcnt = 1;
1485 	if (xfersize > uio->uio_resid)
1486 		xfersize = uio->uio_resid;
1487 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1488 	transp_uio.uio_offset = 0;
1489 	transp_uio.uio_segflg = UIO_SYSSPACE;
1490 	/*
1491 	 * Since transp_iov points to data, and td_ma page array
1492 	 * corresponds to original uio->uio_iov, we need to invert the
1493 	 * direction of the i/o operation as passed to
1494 	 * uiomove_fromphys().
1495 	 */
1496 	switch (uio->uio_rw) {
1497 	case UIO_WRITE:
1498 		transp_uio.uio_rw = UIO_READ;
1499 		break;
1500 	case UIO_READ:
1501 		transp_uio.uio_rw = UIO_WRITE;
1502 		break;
1503 	}
1504 	transp_uio.uio_td = uio->uio_td;
1505 	error = uiomove_fromphys(td->td_ma,
1506 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1507 	    xfersize, &transp_uio);
1508 	adv = xfersize - transp_uio.uio_resid;
1509 	pgadv =
1510 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1511 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1512 	td->td_ma += pgadv;
1513 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1514 	    pgadv));
1515 	td->td_ma_cnt -= pgadv;
1516 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1517 	uio->uio_iov->iov_len -= adv;
1518 	uio->uio_resid -= adv;
1519 	uio->uio_offset += adv;
1520 	return (error);
1521 }
1522 
1523 int
1524 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1525     struct uio *uio)
1526 {
1527 	struct thread *td;
1528 	vm_offset_t iov_base;
1529 	int cnt, pgadv;
1530 
1531 	td = curthread;
1532 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1533 	    uio->uio_segflg != UIO_USERSPACE)
1534 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1535 
1536 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1537 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1538 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1539 	switch (uio->uio_rw) {
1540 	case UIO_WRITE:
1541 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1542 		    offset, cnt);
1543 		break;
1544 	case UIO_READ:
1545 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1546 		    cnt);
1547 		break;
1548 	}
1549 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1550 	td->td_ma += pgadv;
1551 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1552 	    pgadv));
1553 	td->td_ma_cnt -= pgadv;
1554 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1555 	uio->uio_iov->iov_len -= cnt;
1556 	uio->uio_resid -= cnt;
1557 	uio->uio_offset += cnt;
1558 	return (0);
1559 }
1560 
1561 /*
1562  * File table truncate routine.
1563  */
1564 static int
1565 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1566     struct thread *td)
1567 {
1568 	struct mount *mp;
1569 	struct vnode *vp;
1570 	void *rl_cookie;
1571 	int error;
1572 
1573 	vp = fp->f_vnode;
1574 
1575 retry:
1576 	/*
1577 	 * Lock the whole range for truncation.  Otherwise split i/o
1578 	 * might happen partly before and partly after the truncation.
1579 	 */
1580 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1581 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1582 	if (error)
1583 		goto out1;
1584 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1585 	AUDIT_ARG_VNODE1(vp);
1586 	if (vp->v_type == VDIR) {
1587 		error = EISDIR;
1588 		goto out;
1589 	}
1590 #ifdef MAC
1591 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1592 	if (error)
1593 		goto out;
1594 #endif
1595 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1596 	    fp->f_cred);
1597 out:
1598 	VOP_UNLOCK(vp);
1599 	vn_finished_write(mp);
1600 out1:
1601 	vn_rangelock_unlock(vp, rl_cookie);
1602 	if (error == ERELOOKUP)
1603 		goto retry;
1604 	return (error);
1605 }
1606 
1607 /*
1608  * Truncate a file that is already locked.
1609  */
1610 int
1611 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1612     struct ucred *cred)
1613 {
1614 	struct vattr vattr;
1615 	int error;
1616 
1617 	error = VOP_ADD_WRITECOUNT(vp, 1);
1618 	if (error == 0) {
1619 		VATTR_NULL(&vattr);
1620 		vattr.va_size = length;
1621 		if (sync)
1622 			vattr.va_vaflags |= VA_SYNC;
1623 		error = VOP_SETATTR(vp, &vattr, cred);
1624 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1625 	}
1626 	return (error);
1627 }
1628 
1629 /*
1630  * File table vnode stat routine.
1631  */
1632 int
1633 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1634     struct thread *td)
1635 {
1636 	struct vnode *vp = fp->f_vnode;
1637 	int error;
1638 
1639 	vn_lock(vp, LK_SHARED | LK_RETRY);
1640 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1641 	VOP_UNLOCK(vp);
1642 
1643 	return (error);
1644 }
1645 
1646 /*
1647  * File table vnode ioctl routine.
1648  */
1649 static int
1650 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1651     struct thread *td)
1652 {
1653 	struct vattr vattr;
1654 	struct vnode *vp;
1655 	struct fiobmap2_arg *bmarg;
1656 	int error;
1657 
1658 	vp = fp->f_vnode;
1659 	switch (vp->v_type) {
1660 	case VDIR:
1661 	case VREG:
1662 		switch (com) {
1663 		case FIONREAD:
1664 			vn_lock(vp, LK_SHARED | LK_RETRY);
1665 			error = VOP_GETATTR(vp, &vattr, active_cred);
1666 			VOP_UNLOCK(vp);
1667 			if (error == 0)
1668 				*(int *)data = vattr.va_size - fp->f_offset;
1669 			return (error);
1670 		case FIOBMAP2:
1671 			bmarg = (struct fiobmap2_arg *)data;
1672 			vn_lock(vp, LK_SHARED | LK_RETRY);
1673 #ifdef MAC
1674 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1675 			    vp);
1676 			if (error == 0)
1677 #endif
1678 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1679 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1680 			VOP_UNLOCK(vp);
1681 			return (error);
1682 		case FIONBIO:
1683 		case FIOASYNC:
1684 			return (0);
1685 		default:
1686 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1687 			    active_cred, td));
1688 		}
1689 		break;
1690 	case VCHR:
1691 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1692 		    active_cred, td));
1693 	default:
1694 		return (ENOTTY);
1695 	}
1696 }
1697 
1698 /*
1699  * File table vnode poll routine.
1700  */
1701 static int
1702 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1703     struct thread *td)
1704 {
1705 	struct vnode *vp;
1706 	int error;
1707 
1708 	vp = fp->f_vnode;
1709 #if defined(MAC) || defined(AUDIT)
1710 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1711 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1712 		AUDIT_ARG_VNODE1(vp);
1713 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1714 		VOP_UNLOCK(vp);
1715 		if (error != 0)
1716 			return (error);
1717 	}
1718 #endif
1719 	error = VOP_POLL(vp, events, fp->f_cred, td);
1720 	return (error);
1721 }
1722 
1723 /*
1724  * Acquire the requested lock and then check for validity.  LK_RETRY
1725  * permits vn_lock to return doomed vnodes.
1726  */
1727 static int __noinline
1728 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1729     int error)
1730 {
1731 
1732 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1733 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1734 
1735 	if (error == 0)
1736 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1737 
1738 	if ((flags & LK_RETRY) == 0) {
1739 		if (error == 0) {
1740 			VOP_UNLOCK(vp);
1741 			error = ENOENT;
1742 		}
1743 		return (error);
1744 	}
1745 
1746 	/*
1747 	 * LK_RETRY case.
1748 	 *
1749 	 * Nothing to do if we got the lock.
1750 	 */
1751 	if (error == 0)
1752 		return (0);
1753 
1754 	/*
1755 	 * Interlock was dropped by the call in _vn_lock.
1756 	 */
1757 	flags &= ~LK_INTERLOCK;
1758 	do {
1759 		error = VOP_LOCK1(vp, flags, file, line);
1760 	} while (error != 0);
1761 	return (0);
1762 }
1763 
1764 int
1765 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1766 {
1767 	int error;
1768 
1769 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1770 	    ("vn_lock: no locktype (%d passed)", flags));
1771 	VNPASS(vp->v_holdcnt > 0, vp);
1772 	error = VOP_LOCK1(vp, flags, file, line);
1773 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1774 		return (_vn_lock_fallback(vp, flags, file, line, error));
1775 	return (0);
1776 }
1777 
1778 /*
1779  * File table vnode close routine.
1780  */
1781 static int
1782 vn_closefile(struct file *fp, struct thread *td)
1783 {
1784 	struct vnode *vp;
1785 	struct flock lf;
1786 	int error;
1787 	bool ref;
1788 
1789 	vp = fp->f_vnode;
1790 	fp->f_ops = &badfileops;
1791 	ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1792 
1793 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1794 
1795 	if (__predict_false(ref)) {
1796 		lf.l_whence = SEEK_SET;
1797 		lf.l_start = 0;
1798 		lf.l_len = 0;
1799 		lf.l_type = F_UNLCK;
1800 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1801 		vrele(vp);
1802 	}
1803 	return (error);
1804 }
1805 
1806 /*
1807  * Preparing to start a filesystem write operation. If the operation is
1808  * permitted, then we bump the count of operations in progress and
1809  * proceed. If a suspend request is in progress, we wait until the
1810  * suspension is over, and then proceed.
1811  */
1812 static int
1813 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1814 {
1815 	struct mount_pcpu *mpcpu;
1816 	int error, mflags;
1817 
1818 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1819 	    vfs_op_thread_enter(mp, mpcpu)) {
1820 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1821 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1822 		vfs_op_thread_exit(mp, mpcpu);
1823 		return (0);
1824 	}
1825 
1826 	if (mplocked)
1827 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1828 	else
1829 		MNT_ILOCK(mp);
1830 
1831 	error = 0;
1832 
1833 	/*
1834 	 * Check on status of suspension.
1835 	 */
1836 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1837 	    mp->mnt_susp_owner != curthread) {
1838 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1839 		    (flags & PCATCH) : 0) | (PUSER - 1);
1840 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1841 			if (flags & V_NOWAIT) {
1842 				error = EWOULDBLOCK;
1843 				goto unlock;
1844 			}
1845 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1846 			    "suspfs", 0);
1847 			if (error)
1848 				goto unlock;
1849 		}
1850 	}
1851 	if (flags & V_XSLEEP)
1852 		goto unlock;
1853 	mp->mnt_writeopcount++;
1854 unlock:
1855 	if (error != 0 || (flags & V_XSLEEP) != 0)
1856 		MNT_REL(mp);
1857 	MNT_IUNLOCK(mp);
1858 	return (error);
1859 }
1860 
1861 int
1862 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1863 {
1864 	struct mount *mp;
1865 	int error;
1866 
1867 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1868 	    ("V_MNTREF requires mp"));
1869 
1870 	error = 0;
1871 	/*
1872 	 * If a vnode is provided, get and return the mount point that
1873 	 * to which it will write.
1874 	 */
1875 	if (vp != NULL) {
1876 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1877 			*mpp = NULL;
1878 			if (error != EOPNOTSUPP)
1879 				return (error);
1880 			return (0);
1881 		}
1882 	}
1883 	if ((mp = *mpp) == NULL)
1884 		return (0);
1885 
1886 	/*
1887 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1888 	 * a vfs_ref().
1889 	 * As long as a vnode is not provided we need to acquire a
1890 	 * refcount for the provided mountpoint too, in order to
1891 	 * emulate a vfs_ref().
1892 	 */
1893 	if (vp == NULL && (flags & V_MNTREF) == 0)
1894 		vfs_ref(mp);
1895 
1896 	return (vn_start_write_refed(mp, flags, false));
1897 }
1898 
1899 /*
1900  * Secondary suspension. Used by operations such as vop_inactive
1901  * routines that are needed by the higher level functions. These
1902  * are allowed to proceed until all the higher level functions have
1903  * completed (indicated by mnt_writeopcount dropping to zero). At that
1904  * time, these operations are halted until the suspension is over.
1905  */
1906 int
1907 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1908 {
1909 	struct mount *mp;
1910 	int error;
1911 
1912 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1913 	    ("V_MNTREF requires mp"));
1914 
1915  retry:
1916 	if (vp != NULL) {
1917 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1918 			*mpp = NULL;
1919 			if (error != EOPNOTSUPP)
1920 				return (error);
1921 			return (0);
1922 		}
1923 	}
1924 	/*
1925 	 * If we are not suspended or have not yet reached suspended
1926 	 * mode, then let the operation proceed.
1927 	 */
1928 	if ((mp = *mpp) == NULL)
1929 		return (0);
1930 
1931 	/*
1932 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1933 	 * a vfs_ref().
1934 	 * As long as a vnode is not provided we need to acquire a
1935 	 * refcount for the provided mountpoint too, in order to
1936 	 * emulate a vfs_ref().
1937 	 */
1938 	MNT_ILOCK(mp);
1939 	if (vp == NULL && (flags & V_MNTREF) == 0)
1940 		MNT_REF(mp);
1941 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1942 		mp->mnt_secondary_writes++;
1943 		mp->mnt_secondary_accwrites++;
1944 		MNT_IUNLOCK(mp);
1945 		return (0);
1946 	}
1947 	if (flags & V_NOWAIT) {
1948 		MNT_REL(mp);
1949 		MNT_IUNLOCK(mp);
1950 		return (EWOULDBLOCK);
1951 	}
1952 	/*
1953 	 * Wait for the suspension to finish.
1954 	 */
1955 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1956 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1957 	    "suspfs", 0);
1958 	vfs_rel(mp);
1959 	if (error == 0)
1960 		goto retry;
1961 	return (error);
1962 }
1963 
1964 /*
1965  * Filesystem write operation has completed. If we are suspending and this
1966  * operation is the last one, notify the suspender that the suspension is
1967  * now in effect.
1968  */
1969 void
1970 vn_finished_write(struct mount *mp)
1971 {
1972 	struct mount_pcpu *mpcpu;
1973 	int c;
1974 
1975 	if (mp == NULL)
1976 		return;
1977 
1978 	if (vfs_op_thread_enter(mp, mpcpu)) {
1979 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
1980 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
1981 		vfs_op_thread_exit(mp, mpcpu);
1982 		return;
1983 	}
1984 
1985 	MNT_ILOCK(mp);
1986 	vfs_assert_mount_counters(mp);
1987 	MNT_REL(mp);
1988 	c = --mp->mnt_writeopcount;
1989 	if (mp->mnt_vfs_ops == 0) {
1990 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1991 		MNT_IUNLOCK(mp);
1992 		return;
1993 	}
1994 	if (c < 0)
1995 		vfs_dump_mount_counters(mp);
1996 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1997 		wakeup(&mp->mnt_writeopcount);
1998 	MNT_IUNLOCK(mp);
1999 }
2000 
2001 /*
2002  * Filesystem secondary write operation has completed. If we are
2003  * suspending and this operation is the last one, notify the suspender
2004  * that the suspension is now in effect.
2005  */
2006 void
2007 vn_finished_secondary_write(struct mount *mp)
2008 {
2009 	if (mp == NULL)
2010 		return;
2011 	MNT_ILOCK(mp);
2012 	MNT_REL(mp);
2013 	mp->mnt_secondary_writes--;
2014 	if (mp->mnt_secondary_writes < 0)
2015 		panic("vn_finished_secondary_write: neg cnt");
2016 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2017 	    mp->mnt_secondary_writes <= 0)
2018 		wakeup(&mp->mnt_secondary_writes);
2019 	MNT_IUNLOCK(mp);
2020 }
2021 
2022 /*
2023  * Request a filesystem to suspend write operations.
2024  */
2025 int
2026 vfs_write_suspend(struct mount *mp, int flags)
2027 {
2028 	int error;
2029 
2030 	vfs_op_enter(mp);
2031 
2032 	MNT_ILOCK(mp);
2033 	vfs_assert_mount_counters(mp);
2034 	if (mp->mnt_susp_owner == curthread) {
2035 		vfs_op_exit_locked(mp);
2036 		MNT_IUNLOCK(mp);
2037 		return (EALREADY);
2038 	}
2039 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2040 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2041 
2042 	/*
2043 	 * Unmount holds a write reference on the mount point.  If we
2044 	 * own busy reference and drain for writers, we deadlock with
2045 	 * the reference draining in the unmount path.  Callers of
2046 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2047 	 * vfs_busy() reference is owned and caller is not in the
2048 	 * unmount context.
2049 	 */
2050 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2051 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2052 		vfs_op_exit_locked(mp);
2053 		MNT_IUNLOCK(mp);
2054 		return (EBUSY);
2055 	}
2056 
2057 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2058 	mp->mnt_susp_owner = curthread;
2059 	if (mp->mnt_writeopcount > 0)
2060 		(void) msleep(&mp->mnt_writeopcount,
2061 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2062 	else
2063 		MNT_IUNLOCK(mp);
2064 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2065 		vfs_write_resume(mp, 0);
2066 		/* vfs_write_resume does vfs_op_exit() for us */
2067 	}
2068 	return (error);
2069 }
2070 
2071 /*
2072  * Request a filesystem to resume write operations.
2073  */
2074 void
2075 vfs_write_resume(struct mount *mp, int flags)
2076 {
2077 
2078 	MNT_ILOCK(mp);
2079 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2080 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2081 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2082 				       MNTK_SUSPENDED);
2083 		mp->mnt_susp_owner = NULL;
2084 		wakeup(&mp->mnt_writeopcount);
2085 		wakeup(&mp->mnt_flag);
2086 		curthread->td_pflags &= ~TDP_IGNSUSP;
2087 		if ((flags & VR_START_WRITE) != 0) {
2088 			MNT_REF(mp);
2089 			mp->mnt_writeopcount++;
2090 		}
2091 		MNT_IUNLOCK(mp);
2092 		if ((flags & VR_NO_SUSPCLR) == 0)
2093 			VFS_SUSP_CLEAN(mp);
2094 		vfs_op_exit(mp);
2095 	} else if ((flags & VR_START_WRITE) != 0) {
2096 		MNT_REF(mp);
2097 		vn_start_write_refed(mp, 0, true);
2098 	} else {
2099 		MNT_IUNLOCK(mp);
2100 	}
2101 }
2102 
2103 /*
2104  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2105  * methods.
2106  */
2107 int
2108 vfs_write_suspend_umnt(struct mount *mp)
2109 {
2110 	int error;
2111 
2112 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2113 	    ("vfs_write_suspend_umnt: recursed"));
2114 
2115 	/* dounmount() already called vn_start_write(). */
2116 	for (;;) {
2117 		vn_finished_write(mp);
2118 		error = vfs_write_suspend(mp, 0);
2119 		if (error != 0) {
2120 			vn_start_write(NULL, &mp, V_WAIT);
2121 			return (error);
2122 		}
2123 		MNT_ILOCK(mp);
2124 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2125 			break;
2126 		MNT_IUNLOCK(mp);
2127 		vn_start_write(NULL, &mp, V_WAIT);
2128 	}
2129 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2130 	wakeup(&mp->mnt_flag);
2131 	MNT_IUNLOCK(mp);
2132 	curthread->td_pflags |= TDP_IGNSUSP;
2133 	return (0);
2134 }
2135 
2136 /*
2137  * Implement kqueues for files by translating it to vnode operation.
2138  */
2139 static int
2140 vn_kqfilter(struct file *fp, struct knote *kn)
2141 {
2142 
2143 	return (VOP_KQFILTER(fp->f_vnode, kn));
2144 }
2145 
2146 int
2147 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2148 {
2149 	if ((fp->f_flag & FKQALLOWED) == 0)
2150 		return (EBADF);
2151 	return (vn_kqfilter(fp, kn));
2152 }
2153 
2154 /*
2155  * Simplified in-kernel wrapper calls for extended attribute access.
2156  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2157  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2158  */
2159 int
2160 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2161     const char *attrname, int *buflen, char *buf, struct thread *td)
2162 {
2163 	struct uio	auio;
2164 	struct iovec	iov;
2165 	int	error;
2166 
2167 	iov.iov_len = *buflen;
2168 	iov.iov_base = buf;
2169 
2170 	auio.uio_iov = &iov;
2171 	auio.uio_iovcnt = 1;
2172 	auio.uio_rw = UIO_READ;
2173 	auio.uio_segflg = UIO_SYSSPACE;
2174 	auio.uio_td = td;
2175 	auio.uio_offset = 0;
2176 	auio.uio_resid = *buflen;
2177 
2178 	if ((ioflg & IO_NODELOCKED) == 0)
2179 		vn_lock(vp, LK_SHARED | LK_RETRY);
2180 
2181 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2182 
2183 	/* authorize attribute retrieval as kernel */
2184 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2185 	    td);
2186 
2187 	if ((ioflg & IO_NODELOCKED) == 0)
2188 		VOP_UNLOCK(vp);
2189 
2190 	if (error == 0) {
2191 		*buflen = *buflen - auio.uio_resid;
2192 	}
2193 
2194 	return (error);
2195 }
2196 
2197 /*
2198  * XXX failure mode if partially written?
2199  */
2200 int
2201 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2202     const char *attrname, int buflen, char *buf, struct thread *td)
2203 {
2204 	struct uio	auio;
2205 	struct iovec	iov;
2206 	struct mount	*mp;
2207 	int	error;
2208 
2209 	iov.iov_len = buflen;
2210 	iov.iov_base = buf;
2211 
2212 	auio.uio_iov = &iov;
2213 	auio.uio_iovcnt = 1;
2214 	auio.uio_rw = UIO_WRITE;
2215 	auio.uio_segflg = UIO_SYSSPACE;
2216 	auio.uio_td = td;
2217 	auio.uio_offset = 0;
2218 	auio.uio_resid = buflen;
2219 
2220 	if ((ioflg & IO_NODELOCKED) == 0) {
2221 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2222 			return (error);
2223 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2224 	}
2225 
2226 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2227 
2228 	/* authorize attribute setting as kernel */
2229 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2230 
2231 	if ((ioflg & IO_NODELOCKED) == 0) {
2232 		vn_finished_write(mp);
2233 		VOP_UNLOCK(vp);
2234 	}
2235 
2236 	return (error);
2237 }
2238 
2239 int
2240 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2241     const char *attrname, struct thread *td)
2242 {
2243 	struct mount	*mp;
2244 	int	error;
2245 
2246 	if ((ioflg & IO_NODELOCKED) == 0) {
2247 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2248 			return (error);
2249 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2250 	}
2251 
2252 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2253 
2254 	/* authorize attribute removal as kernel */
2255 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2256 	if (error == EOPNOTSUPP)
2257 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2258 		    NULL, td);
2259 
2260 	if ((ioflg & IO_NODELOCKED) == 0) {
2261 		vn_finished_write(mp);
2262 		VOP_UNLOCK(vp);
2263 	}
2264 
2265 	return (error);
2266 }
2267 
2268 static int
2269 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2270     struct vnode **rvp)
2271 {
2272 
2273 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2274 }
2275 
2276 int
2277 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2278 {
2279 
2280 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2281 	    lkflags, rvp));
2282 }
2283 
2284 int
2285 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2286     int lkflags, struct vnode **rvp)
2287 {
2288 	struct mount *mp;
2289 	int ltype, error;
2290 
2291 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2292 	mp = vp->v_mount;
2293 	ltype = VOP_ISLOCKED(vp);
2294 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2295 	    ("vn_vget_ino: vp not locked"));
2296 	error = vfs_busy(mp, MBF_NOWAIT);
2297 	if (error != 0) {
2298 		vfs_ref(mp);
2299 		VOP_UNLOCK(vp);
2300 		error = vfs_busy(mp, 0);
2301 		vn_lock(vp, ltype | LK_RETRY);
2302 		vfs_rel(mp);
2303 		if (error != 0)
2304 			return (ENOENT);
2305 		if (VN_IS_DOOMED(vp)) {
2306 			vfs_unbusy(mp);
2307 			return (ENOENT);
2308 		}
2309 	}
2310 	VOP_UNLOCK(vp);
2311 	error = alloc(mp, alloc_arg, lkflags, rvp);
2312 	vfs_unbusy(mp);
2313 	if (error != 0 || *rvp != vp)
2314 		vn_lock(vp, ltype | LK_RETRY);
2315 	if (VN_IS_DOOMED(vp)) {
2316 		if (error == 0) {
2317 			if (*rvp == vp)
2318 				vunref(vp);
2319 			else
2320 				vput(*rvp);
2321 		}
2322 		error = ENOENT;
2323 	}
2324 	return (error);
2325 }
2326 
2327 int
2328 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2329     struct thread *td)
2330 {
2331 
2332 	if (vp->v_type != VREG || td == NULL)
2333 		return (0);
2334 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2335 	    lim_cur(td, RLIMIT_FSIZE)) {
2336 		PROC_LOCK(td->td_proc);
2337 		kern_psignal(td->td_proc, SIGXFSZ);
2338 		PROC_UNLOCK(td->td_proc);
2339 		return (EFBIG);
2340 	}
2341 	return (0);
2342 }
2343 
2344 int
2345 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2346     struct thread *td)
2347 {
2348 	struct vnode *vp;
2349 
2350 	vp = fp->f_vnode;
2351 #ifdef AUDIT
2352 	vn_lock(vp, LK_SHARED | LK_RETRY);
2353 	AUDIT_ARG_VNODE1(vp);
2354 	VOP_UNLOCK(vp);
2355 #endif
2356 	return (setfmode(td, active_cred, vp, mode));
2357 }
2358 
2359 int
2360 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2361     struct thread *td)
2362 {
2363 	struct vnode *vp;
2364 
2365 	vp = fp->f_vnode;
2366 #ifdef AUDIT
2367 	vn_lock(vp, LK_SHARED | LK_RETRY);
2368 	AUDIT_ARG_VNODE1(vp);
2369 	VOP_UNLOCK(vp);
2370 #endif
2371 	return (setfown(td, active_cred, vp, uid, gid));
2372 }
2373 
2374 void
2375 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2376 {
2377 	vm_object_t object;
2378 
2379 	if ((object = vp->v_object) == NULL)
2380 		return;
2381 	VM_OBJECT_WLOCK(object);
2382 	vm_object_page_remove(object, start, end, 0);
2383 	VM_OBJECT_WUNLOCK(object);
2384 }
2385 
2386 int
2387 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2388 {
2389 	struct vattr va;
2390 	daddr_t bn, bnp;
2391 	uint64_t bsize;
2392 	off_t noff;
2393 	int error;
2394 
2395 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2396 	    ("Wrong command %lu", cmd));
2397 
2398 	if (vn_lock(vp, LK_SHARED) != 0)
2399 		return (EBADF);
2400 	if (vp->v_type != VREG) {
2401 		error = ENOTTY;
2402 		goto unlock;
2403 	}
2404 	error = VOP_GETATTR(vp, &va, cred);
2405 	if (error != 0)
2406 		goto unlock;
2407 	noff = *off;
2408 	if (noff >= va.va_size) {
2409 		error = ENXIO;
2410 		goto unlock;
2411 	}
2412 	bsize = vp->v_mount->mnt_stat.f_iosize;
2413 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2414 	    noff % bsize) {
2415 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2416 		if (error == EOPNOTSUPP) {
2417 			error = ENOTTY;
2418 			goto unlock;
2419 		}
2420 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2421 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2422 			noff = bn * bsize;
2423 			if (noff < *off)
2424 				noff = *off;
2425 			goto unlock;
2426 		}
2427 	}
2428 	if (noff > va.va_size)
2429 		noff = va.va_size;
2430 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2431 	if (cmd == FIOSEEKDATA)
2432 		error = ENXIO;
2433 unlock:
2434 	VOP_UNLOCK(vp);
2435 	if (error == 0)
2436 		*off = noff;
2437 	return (error);
2438 }
2439 
2440 int
2441 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2442 {
2443 	struct ucred *cred;
2444 	struct vnode *vp;
2445 	struct vattr vattr;
2446 	off_t foffset, size;
2447 	int error, noneg;
2448 
2449 	cred = td->td_ucred;
2450 	vp = fp->f_vnode;
2451 	foffset = foffset_lock(fp, 0);
2452 	noneg = (vp->v_type != VCHR);
2453 	error = 0;
2454 	switch (whence) {
2455 	case L_INCR:
2456 		if (noneg &&
2457 		    (foffset < 0 ||
2458 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2459 			error = EOVERFLOW;
2460 			break;
2461 		}
2462 		offset += foffset;
2463 		break;
2464 	case L_XTND:
2465 		vn_lock(vp, LK_SHARED | LK_RETRY);
2466 		error = VOP_GETATTR(vp, &vattr, cred);
2467 		VOP_UNLOCK(vp);
2468 		if (error)
2469 			break;
2470 
2471 		/*
2472 		 * If the file references a disk device, then fetch
2473 		 * the media size and use that to determine the ending
2474 		 * offset.
2475 		 */
2476 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2477 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2478 			vattr.va_size = size;
2479 		if (noneg &&
2480 		    (vattr.va_size > OFF_MAX ||
2481 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2482 			error = EOVERFLOW;
2483 			break;
2484 		}
2485 		offset += vattr.va_size;
2486 		break;
2487 	case L_SET:
2488 		break;
2489 	case SEEK_DATA:
2490 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2491 		if (error == ENOTTY)
2492 			error = EINVAL;
2493 		break;
2494 	case SEEK_HOLE:
2495 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2496 		if (error == ENOTTY)
2497 			error = EINVAL;
2498 		break;
2499 	default:
2500 		error = EINVAL;
2501 	}
2502 	if (error == 0 && noneg && offset < 0)
2503 		error = EINVAL;
2504 	if (error != 0)
2505 		goto drop;
2506 	VFS_KNOTE_UNLOCKED(vp, 0);
2507 	td->td_uretoff.tdu_off = offset;
2508 drop:
2509 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2510 	return (error);
2511 }
2512 
2513 int
2514 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2515     struct thread *td)
2516 {
2517 	int error;
2518 
2519 	/*
2520 	 * Grant permission if the caller is the owner of the file, or
2521 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2522 	 * on the file.  If the time pointer is null, then write
2523 	 * permission on the file is also sufficient.
2524 	 *
2525 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2526 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2527 	 * will be allowed to set the times [..] to the current
2528 	 * server time.
2529 	 */
2530 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2531 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2532 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2533 	return (error);
2534 }
2535 
2536 int
2537 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2538 {
2539 	struct vnode *vp;
2540 	int error;
2541 
2542 	if (fp->f_type == DTYPE_FIFO)
2543 		kif->kf_type = KF_TYPE_FIFO;
2544 	else
2545 		kif->kf_type = KF_TYPE_VNODE;
2546 	vp = fp->f_vnode;
2547 	vref(vp);
2548 	FILEDESC_SUNLOCK(fdp);
2549 	error = vn_fill_kinfo_vnode(vp, kif);
2550 	vrele(vp);
2551 	FILEDESC_SLOCK(fdp);
2552 	return (error);
2553 }
2554 
2555 static inline void
2556 vn_fill_junk(struct kinfo_file *kif)
2557 {
2558 	size_t len, olen;
2559 
2560 	/*
2561 	 * Simulate vn_fullpath returning changing values for a given
2562 	 * vp during e.g. coredump.
2563 	 */
2564 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2565 	olen = strlen(kif->kf_path);
2566 	if (len < olen)
2567 		strcpy(&kif->kf_path[len - 1], "$");
2568 	else
2569 		for (; olen < len; olen++)
2570 			strcpy(&kif->kf_path[olen], "A");
2571 }
2572 
2573 int
2574 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2575 {
2576 	struct vattr va;
2577 	char *fullpath, *freepath;
2578 	int error;
2579 
2580 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2581 	freepath = NULL;
2582 	fullpath = "-";
2583 	error = vn_fullpath(vp, &fullpath, &freepath);
2584 	if (error == 0) {
2585 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2586 	}
2587 	if (freepath != NULL)
2588 		free(freepath, M_TEMP);
2589 
2590 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2591 		vn_fill_junk(kif);
2592 	);
2593 
2594 	/*
2595 	 * Retrieve vnode attributes.
2596 	 */
2597 	va.va_fsid = VNOVAL;
2598 	va.va_rdev = NODEV;
2599 	vn_lock(vp, LK_SHARED | LK_RETRY);
2600 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2601 	VOP_UNLOCK(vp);
2602 	if (error != 0)
2603 		return (error);
2604 	if (va.va_fsid != VNOVAL)
2605 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2606 	else
2607 		kif->kf_un.kf_file.kf_file_fsid =
2608 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2609 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2610 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2611 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2612 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2613 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2614 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2615 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2616 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2617 	return (0);
2618 }
2619 
2620 int
2621 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2622     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2623     struct thread *td)
2624 {
2625 #ifdef HWPMC_HOOKS
2626 	struct pmckern_map_in pkm;
2627 #endif
2628 	struct mount *mp;
2629 	struct vnode *vp;
2630 	vm_object_t object;
2631 	vm_prot_t maxprot;
2632 	boolean_t writecounted;
2633 	int error;
2634 
2635 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2636     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2637 	/*
2638 	 * POSIX shared-memory objects are defined to have
2639 	 * kernel persistence, and are not defined to support
2640 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2641 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2642 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2643 	 * flag to request this behavior.
2644 	 */
2645 	if ((fp->f_flag & FPOSIXSHM) != 0)
2646 		flags |= MAP_NOSYNC;
2647 #endif
2648 	vp = fp->f_vnode;
2649 
2650 	/*
2651 	 * Ensure that file and memory protections are
2652 	 * compatible.  Note that we only worry about
2653 	 * writability if mapping is shared; in this case,
2654 	 * current and max prot are dictated by the open file.
2655 	 * XXX use the vnode instead?  Problem is: what
2656 	 * credentials do we use for determination? What if
2657 	 * proc does a setuid?
2658 	 */
2659 	mp = vp->v_mount;
2660 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2661 		maxprot = VM_PROT_NONE;
2662 		if ((prot & VM_PROT_EXECUTE) != 0)
2663 			return (EACCES);
2664 	} else
2665 		maxprot = VM_PROT_EXECUTE;
2666 	if ((fp->f_flag & FREAD) != 0)
2667 		maxprot |= VM_PROT_READ;
2668 	else if ((prot & VM_PROT_READ) != 0)
2669 		return (EACCES);
2670 
2671 	/*
2672 	 * If we are sharing potential changes via MAP_SHARED and we
2673 	 * are trying to get write permission although we opened it
2674 	 * without asking for it, bail out.
2675 	 */
2676 	if ((flags & MAP_SHARED) != 0) {
2677 		if ((fp->f_flag & FWRITE) != 0)
2678 			maxprot |= VM_PROT_WRITE;
2679 		else if ((prot & VM_PROT_WRITE) != 0)
2680 			return (EACCES);
2681 	} else {
2682 		maxprot |= VM_PROT_WRITE;
2683 		cap_maxprot |= VM_PROT_WRITE;
2684 	}
2685 	maxprot &= cap_maxprot;
2686 
2687 	/*
2688 	 * For regular files and shared memory, POSIX requires that
2689 	 * the value of foff be a legitimate offset within the data
2690 	 * object.  In particular, negative offsets are invalid.
2691 	 * Blocking negative offsets and overflows here avoids
2692 	 * possible wraparound or user-level access into reserved
2693 	 * ranges of the data object later.  In contrast, POSIX does
2694 	 * not dictate how offsets are used by device drivers, so in
2695 	 * the case of a device mapping a negative offset is passed
2696 	 * on.
2697 	 */
2698 	if (
2699 #ifdef _LP64
2700 	    size > OFF_MAX ||
2701 #endif
2702 	    foff > OFF_MAX - size)
2703 		return (EINVAL);
2704 
2705 	writecounted = FALSE;
2706 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2707 	    &foff, &object, &writecounted);
2708 	if (error != 0)
2709 		return (error);
2710 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2711 	    foff, writecounted, td);
2712 	if (error != 0) {
2713 		/*
2714 		 * If this mapping was accounted for in the vnode's
2715 		 * writecount, then undo that now.
2716 		 */
2717 		if (writecounted)
2718 			vm_pager_release_writecount(object, 0, size);
2719 		vm_object_deallocate(object);
2720 	}
2721 #ifdef HWPMC_HOOKS
2722 	/* Inform hwpmc(4) if an executable is being mapped. */
2723 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2724 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2725 			pkm.pm_file = vp;
2726 			pkm.pm_address = (uintptr_t) *addr;
2727 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2728 		}
2729 	}
2730 #endif
2731 	return (error);
2732 }
2733 
2734 void
2735 vn_fsid(struct vnode *vp, struct vattr *va)
2736 {
2737 	fsid_t *f;
2738 
2739 	f = &vp->v_mount->mnt_stat.f_fsid;
2740 	va->va_fsid = (uint32_t)f->val[1];
2741 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2742 	va->va_fsid += (uint32_t)f->val[0];
2743 }
2744 
2745 int
2746 vn_fsync_buf(struct vnode *vp, int waitfor)
2747 {
2748 	struct buf *bp, *nbp;
2749 	struct bufobj *bo;
2750 	struct mount *mp;
2751 	int error, maxretry;
2752 
2753 	error = 0;
2754 	maxretry = 10000;     /* large, arbitrarily chosen */
2755 	mp = NULL;
2756 	if (vp->v_type == VCHR) {
2757 		VI_LOCK(vp);
2758 		mp = vp->v_rdev->si_mountpt;
2759 		VI_UNLOCK(vp);
2760 	}
2761 	bo = &vp->v_bufobj;
2762 	BO_LOCK(bo);
2763 loop1:
2764 	/*
2765 	 * MARK/SCAN initialization to avoid infinite loops.
2766 	 */
2767         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2768 		bp->b_vflags &= ~BV_SCANNED;
2769 		bp->b_error = 0;
2770 	}
2771 
2772 	/*
2773 	 * Flush all dirty buffers associated with a vnode.
2774 	 */
2775 loop2:
2776 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2777 		if ((bp->b_vflags & BV_SCANNED) != 0)
2778 			continue;
2779 		bp->b_vflags |= BV_SCANNED;
2780 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2781 			if (waitfor != MNT_WAIT)
2782 				continue;
2783 			if (BUF_LOCK(bp,
2784 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2785 			    BO_LOCKPTR(bo)) != 0) {
2786 				BO_LOCK(bo);
2787 				goto loop1;
2788 			}
2789 			BO_LOCK(bo);
2790 		}
2791 		BO_UNLOCK(bo);
2792 		KASSERT(bp->b_bufobj == bo,
2793 		    ("bp %p wrong b_bufobj %p should be %p",
2794 		    bp, bp->b_bufobj, bo));
2795 		if ((bp->b_flags & B_DELWRI) == 0)
2796 			panic("fsync: not dirty");
2797 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2798 			vfs_bio_awrite(bp);
2799 		} else {
2800 			bremfree(bp);
2801 			bawrite(bp);
2802 		}
2803 		if (maxretry < 1000)
2804 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2805 		BO_LOCK(bo);
2806 		goto loop2;
2807 	}
2808 
2809 	/*
2810 	 * If synchronous the caller expects us to completely resolve all
2811 	 * dirty buffers in the system.  Wait for in-progress I/O to
2812 	 * complete (which could include background bitmap writes), then
2813 	 * retry if dirty blocks still exist.
2814 	 */
2815 	if (waitfor == MNT_WAIT) {
2816 		bufobj_wwait(bo, 0, 0);
2817 		if (bo->bo_dirty.bv_cnt > 0) {
2818 			/*
2819 			 * If we are unable to write any of these buffers
2820 			 * then we fail now rather than trying endlessly
2821 			 * to write them out.
2822 			 */
2823 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2824 				if ((error = bp->b_error) != 0)
2825 					break;
2826 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2827 			    (error == 0 && --maxretry >= 0))
2828 				goto loop1;
2829 			if (error == 0)
2830 				error = EAGAIN;
2831 		}
2832 	}
2833 	BO_UNLOCK(bo);
2834 	if (error != 0)
2835 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2836 
2837 	return (error);
2838 }
2839 
2840 /*
2841  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2842  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2843  * to do the actual copy.
2844  * vn_generic_copy_file_range() is factored out, so it can be called
2845  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2846  * different file systems.
2847  */
2848 int
2849 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2850     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2851     struct ucred *outcred, struct thread *fsize_td)
2852 {
2853 	int error;
2854 	size_t len;
2855 	uint64_t uval;
2856 
2857 	len = *lenp;
2858 	*lenp = 0;		/* For error returns. */
2859 	error = 0;
2860 
2861 	/* Do some sanity checks on the arguments. */
2862 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2863 		error = EISDIR;
2864 	else if (*inoffp < 0 || *outoffp < 0 ||
2865 	    invp->v_type != VREG || outvp->v_type != VREG)
2866 		error = EINVAL;
2867 	if (error != 0)
2868 		goto out;
2869 
2870 	/* Ensure offset + len does not wrap around. */
2871 	uval = *inoffp;
2872 	uval += len;
2873 	if (uval > INT64_MAX)
2874 		len = INT64_MAX - *inoffp;
2875 	uval = *outoffp;
2876 	uval += len;
2877 	if (uval > INT64_MAX)
2878 		len = INT64_MAX - *outoffp;
2879 	if (len == 0)
2880 		goto out;
2881 
2882 	/*
2883 	 * If the two vnode are for the same file system, call
2884 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2885 	 * which can handle copies across multiple file systems.
2886 	 */
2887 	*lenp = len;
2888 	if (invp->v_mount == outvp->v_mount)
2889 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2890 		    lenp, flags, incred, outcred, fsize_td);
2891 	else
2892 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2893 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2894 out:
2895 	return (error);
2896 }
2897 
2898 /*
2899  * Test len bytes of data starting at dat for all bytes == 0.
2900  * Return true if all bytes are zero, false otherwise.
2901  * Expects dat to be well aligned.
2902  */
2903 static bool
2904 mem_iszero(void *dat, int len)
2905 {
2906 	int i;
2907 	const u_int *p;
2908 	const char *cp;
2909 
2910 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2911 		if (len >= sizeof(*p)) {
2912 			if (*p != 0)
2913 				return (false);
2914 		} else {
2915 			cp = (const char *)p;
2916 			for (i = 0; i < len; i++, cp++)
2917 				if (*cp != '\0')
2918 					return (false);
2919 		}
2920 	}
2921 	return (true);
2922 }
2923 
2924 /*
2925  * Look for a hole in the output file and, if found, adjust *outoffp
2926  * and *xferp to skip past the hole.
2927  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2928  * to be written as 0's upon return.
2929  */
2930 static off_t
2931 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2932     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2933 {
2934 	int error;
2935 	off_t delta;
2936 
2937 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2938 		*dataoffp = *outoffp;
2939 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2940 		    curthread);
2941 		if (error == 0) {
2942 			*holeoffp = *dataoffp;
2943 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2944 			    curthread);
2945 		}
2946 		if (error != 0 || *holeoffp == *dataoffp) {
2947 			/*
2948 			 * Since outvp is unlocked, it may be possible for
2949 			 * another thread to do a truncate(), lseek(), write()
2950 			 * creating a hole at startoff between the above
2951 			 * VOP_IOCTL() calls, if the other thread does not do
2952 			 * rangelocking.
2953 			 * If that happens, *holeoffp == *dataoffp and finding
2954 			 * the hole has failed, so disable vn_skip_hole().
2955 			 */
2956 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2957 			return (xfer2);
2958 		}
2959 		KASSERT(*dataoffp >= *outoffp,
2960 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2961 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2962 		KASSERT(*holeoffp > *dataoffp,
2963 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2964 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2965 	}
2966 
2967 	/*
2968 	 * If there is a hole before the data starts, advance *outoffp and
2969 	 * *xferp past the hole.
2970 	 */
2971 	if (*dataoffp > *outoffp) {
2972 		delta = *dataoffp - *outoffp;
2973 		if (delta >= *xferp) {
2974 			/* Entire *xferp is a hole. */
2975 			*outoffp += *xferp;
2976 			*xferp = 0;
2977 			return (0);
2978 		}
2979 		*xferp -= delta;
2980 		*outoffp += delta;
2981 		xfer2 = MIN(xfer2, *xferp);
2982 	}
2983 
2984 	/*
2985 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2986 	 * that the write ends at the start of the hole.
2987 	 * *holeoffp should always be greater than *outoffp, but for the
2988 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2989 	 * value.
2990 	 */
2991 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2992 		xfer2 = *holeoffp - *outoffp;
2993 	return (xfer2);
2994 }
2995 
2996 /*
2997  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2998  * dat is a maximum of blksize in length and can be written repeatedly in
2999  * the chunk.
3000  * If growfile == true, just grow the file via vn_truncate_locked() instead
3001  * of doing actual writes.
3002  * If checkhole == true, a hole is being punched, so skip over any hole
3003  * already in the output file.
3004  */
3005 static int
3006 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3007     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3008 {
3009 	struct mount *mp;
3010 	off_t dataoff, holeoff, xfer2;
3011 	int error, lckf;
3012 
3013 	/*
3014 	 * Loop around doing writes of blksize until write has been completed.
3015 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3016 	 * done for each iteration, since the xfer argument can be very
3017 	 * large if there is a large hole to punch in the output file.
3018 	 */
3019 	error = 0;
3020 	holeoff = 0;
3021 	do {
3022 		xfer2 = MIN(xfer, blksize);
3023 		if (checkhole) {
3024 			/*
3025 			 * Punching a hole.  Skip writing if there is
3026 			 * already a hole in the output file.
3027 			 */
3028 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3029 			    &dataoff, &holeoff, cred);
3030 			if (xfer == 0)
3031 				break;
3032 			if (holeoff < 0)
3033 				checkhole = false;
3034 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3035 			    (intmax_t)xfer2));
3036 		}
3037 		bwillwrite();
3038 		mp = NULL;
3039 		error = vn_start_write(outvp, &mp, V_WAIT);
3040 		if (error != 0)
3041 			break;
3042 		if (growfile) {
3043 			error = vn_lock(outvp, LK_EXCLUSIVE);
3044 			if (error == 0) {
3045 				error = vn_truncate_locked(outvp, outoff + xfer,
3046 				    false, cred);
3047 				VOP_UNLOCK(outvp);
3048 			}
3049 		} else {
3050 			if (MNT_SHARED_WRITES(mp))
3051 				lckf = LK_SHARED;
3052 			else
3053 				lckf = LK_EXCLUSIVE;
3054 			error = vn_lock(outvp, lckf);
3055 			if (error == 0) {
3056 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3057 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3058 				    curthread->td_ucred, cred, NULL, curthread);
3059 				outoff += xfer2;
3060 				xfer -= xfer2;
3061 				VOP_UNLOCK(outvp);
3062 			}
3063 		}
3064 		if (mp != NULL)
3065 			vn_finished_write(mp);
3066 	} while (!growfile && xfer > 0 && error == 0);
3067 	return (error);
3068 }
3069 
3070 /*
3071  * Copy a byte range of one file to another.  This function can handle the
3072  * case where invp and outvp are on different file systems.
3073  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3074  * is no better file system specific way to do it.
3075  */
3076 int
3077 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3078     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3079     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3080 {
3081 	struct vattr va;
3082 	struct mount *mp;
3083 	struct uio io;
3084 	off_t startoff, endoff, xfer, xfer2;
3085 	u_long blksize;
3086 	int error, interrupted;
3087 	bool cantseek, readzeros, eof, lastblock;
3088 	ssize_t aresid;
3089 	size_t copylen, len, rem, savlen;
3090 	char *dat;
3091 	long holein, holeout;
3092 
3093 	holein = holeout = 0;
3094 	savlen = len = *lenp;
3095 	error = 0;
3096 	interrupted = 0;
3097 	dat = NULL;
3098 
3099 	error = vn_lock(invp, LK_SHARED);
3100 	if (error != 0)
3101 		goto out;
3102 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3103 		holein = 0;
3104 	VOP_UNLOCK(invp);
3105 
3106 	mp = NULL;
3107 	error = vn_start_write(outvp, &mp, V_WAIT);
3108 	if (error == 0)
3109 		error = vn_lock(outvp, LK_EXCLUSIVE);
3110 	if (error == 0) {
3111 		/*
3112 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3113 		 * now that outvp is locked.
3114 		 */
3115 		if (fsize_td != NULL) {
3116 			io.uio_offset = *outoffp;
3117 			io.uio_resid = len;
3118 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3119 			if (error != 0)
3120 				error = EFBIG;
3121 		}
3122 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3123 			holeout = 0;
3124 		/*
3125 		 * Holes that are past EOF do not need to be written as a block
3126 		 * of zero bytes.  So, truncate the output file as far as
3127 		 * possible and then use va.va_size to decide if writing 0
3128 		 * bytes is necessary in the loop below.
3129 		 */
3130 		if (error == 0)
3131 			error = VOP_GETATTR(outvp, &va, outcred);
3132 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3133 		    *outoffp + len) {
3134 #ifdef MAC
3135 			error = mac_vnode_check_write(curthread->td_ucred,
3136 			    outcred, outvp);
3137 			if (error == 0)
3138 #endif
3139 				error = vn_truncate_locked(outvp, *outoffp,
3140 				    false, outcred);
3141 			if (error == 0)
3142 				va.va_size = *outoffp;
3143 		}
3144 		VOP_UNLOCK(outvp);
3145 	}
3146 	if (mp != NULL)
3147 		vn_finished_write(mp);
3148 	if (error != 0)
3149 		goto out;
3150 
3151 	/*
3152 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3153 	 * If hole sizes aren't available, set the blksize to the larger
3154 	 * f_iosize of invp and outvp.
3155 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3156 	 * This value is clipped at 4Kbytes and 1Mbyte.
3157 	 */
3158 	blksize = MAX(holein, holeout);
3159 
3160 	/* Clip len to end at an exact multiple of hole size. */
3161 	if (blksize > 1) {
3162 		rem = *inoffp % blksize;
3163 		if (rem > 0)
3164 			rem = blksize - rem;
3165 		if (len > rem && len - rem > blksize)
3166 			len = savlen = rounddown(len - rem, blksize) + rem;
3167 	}
3168 
3169 	if (blksize <= 1)
3170 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3171 		    outvp->v_mount->mnt_stat.f_iosize);
3172 	if (blksize < 4096)
3173 		blksize = 4096;
3174 	else if (blksize > 1024 * 1024)
3175 		blksize = 1024 * 1024;
3176 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3177 
3178 	/*
3179 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3180 	 * to find holes.  Otherwise, just scan the read block for all 0s
3181 	 * in the inner loop where the data copying is done.
3182 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3183 	 * support holes on the server, but do not support FIOSEEKHOLE.
3184 	 */
3185 	eof = false;
3186 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3187 		endoff = 0;			/* To shut up compilers. */
3188 		cantseek = true;
3189 		startoff = *inoffp;
3190 		copylen = len;
3191 
3192 		/*
3193 		 * Find the next data area.  If there is just a hole to EOF,
3194 		 * FIOSEEKDATA should fail and then we drop down into the
3195 		 * inner loop and create the hole on the outvp file.
3196 		 * (I do not know if any file system will report a hole to
3197 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3198 		 *  will fail for those file systems.)
3199 		 *
3200 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3201 		 * the code just falls through to the inner copy loop.
3202 		 */
3203 		error = EINVAL;
3204 		if (holein > 0)
3205 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3206 			    incred, curthread);
3207 		if (error == 0) {
3208 			endoff = startoff;
3209 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3210 			    incred, curthread);
3211 			/*
3212 			 * Since invp is unlocked, it may be possible for
3213 			 * another thread to do a truncate(), lseek(), write()
3214 			 * creating a hole at startoff between the above
3215 			 * VOP_IOCTL() calls, if the other thread does not do
3216 			 * rangelocking.
3217 			 * If that happens, startoff == endoff and finding
3218 			 * the hole has failed, so set an error.
3219 			 */
3220 			if (error == 0 && startoff == endoff)
3221 				error = EINVAL; /* Any error. Reset to 0. */
3222 		}
3223 		if (error == 0) {
3224 			if (startoff > *inoffp) {
3225 				/* Found hole before data block. */
3226 				xfer = MIN(startoff - *inoffp, len);
3227 				if (*outoffp < va.va_size) {
3228 					/* Must write 0s to punch hole. */
3229 					xfer2 = MIN(va.va_size - *outoffp,
3230 					    xfer);
3231 					memset(dat, 0, MIN(xfer2, blksize));
3232 					error = vn_write_outvp(outvp, dat,
3233 					    *outoffp, xfer2, blksize, false,
3234 					    holeout > 0, outcred);
3235 				}
3236 
3237 				if (error == 0 && *outoffp + xfer >
3238 				    va.va_size && xfer == len)
3239 					/* Grow last block. */
3240 					error = vn_write_outvp(outvp, dat,
3241 					    *outoffp, xfer, blksize, true,
3242 					    false, outcred);
3243 				if (error == 0) {
3244 					*inoffp += xfer;
3245 					*outoffp += xfer;
3246 					len -= xfer;
3247 					if (len < savlen)
3248 						interrupted = sig_intr();
3249 				}
3250 			}
3251 			copylen = MIN(len, endoff - startoff);
3252 			cantseek = false;
3253 		} else {
3254 			cantseek = true;
3255 			startoff = *inoffp;
3256 			copylen = len;
3257 			error = 0;
3258 		}
3259 
3260 		xfer = blksize;
3261 		if (cantseek) {
3262 			/*
3263 			 * Set first xfer to end at a block boundary, so that
3264 			 * holes are more likely detected in the loop below via
3265 			 * the for all bytes 0 method.
3266 			 */
3267 			xfer -= (*inoffp % blksize);
3268 		}
3269 		/* Loop copying the data block. */
3270 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3271 			if (copylen < xfer)
3272 				xfer = copylen;
3273 			error = vn_lock(invp, LK_SHARED);
3274 			if (error != 0)
3275 				goto out;
3276 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3277 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3278 			    curthread->td_ucred, incred, &aresid,
3279 			    curthread);
3280 			VOP_UNLOCK(invp);
3281 			lastblock = false;
3282 			if (error == 0 && aresid > 0) {
3283 				/* Stop the copy at EOF on the input file. */
3284 				xfer -= aresid;
3285 				eof = true;
3286 				lastblock = true;
3287 			}
3288 			if (error == 0) {
3289 				/*
3290 				 * Skip the write for holes past the initial EOF
3291 				 * of the output file, unless this is the last
3292 				 * write of the output file at EOF.
3293 				 */
3294 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3295 				    false;
3296 				if (xfer == len)
3297 					lastblock = true;
3298 				if (!cantseek || *outoffp < va.va_size ||
3299 				    lastblock || !readzeros)
3300 					error = vn_write_outvp(outvp, dat,
3301 					    *outoffp, xfer, blksize,
3302 					    readzeros && lastblock &&
3303 					    *outoffp >= va.va_size, false,
3304 					    outcred);
3305 				if (error == 0) {
3306 					*inoffp += xfer;
3307 					startoff += xfer;
3308 					*outoffp += xfer;
3309 					copylen -= xfer;
3310 					len -= xfer;
3311 					if (len < savlen)
3312 						interrupted = sig_intr();
3313 				}
3314 			}
3315 			xfer = blksize;
3316 		}
3317 	}
3318 out:
3319 	*lenp = savlen - len;
3320 	free(dat, M_TEMP);
3321 	return (error);
3322 }
3323 
3324 static int
3325 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3326 {
3327 	struct mount *mp;
3328 	struct vnode *vp;
3329 	off_t olen, ooffset;
3330 	int error;
3331 #ifdef AUDIT
3332 	int audited_vnode1 = 0;
3333 #endif
3334 
3335 	vp = fp->f_vnode;
3336 	if (vp->v_type != VREG)
3337 		return (ENODEV);
3338 
3339 	/* Allocating blocks may take a long time, so iterate. */
3340 	for (;;) {
3341 		olen = len;
3342 		ooffset = offset;
3343 
3344 		bwillwrite();
3345 		mp = NULL;
3346 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3347 		if (error != 0)
3348 			break;
3349 		error = vn_lock(vp, LK_EXCLUSIVE);
3350 		if (error != 0) {
3351 			vn_finished_write(mp);
3352 			break;
3353 		}
3354 #ifdef AUDIT
3355 		if (!audited_vnode1) {
3356 			AUDIT_ARG_VNODE1(vp);
3357 			audited_vnode1 = 1;
3358 		}
3359 #endif
3360 #ifdef MAC
3361 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3362 		if (error == 0)
3363 #endif
3364 			error = VOP_ALLOCATE(vp, &offset, &len);
3365 		VOP_UNLOCK(vp);
3366 		vn_finished_write(mp);
3367 
3368 		if (olen + ooffset != offset + len) {
3369 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3370 			    ooffset, olen, offset, len);
3371 		}
3372 		if (error != 0 || len == 0)
3373 			break;
3374 		KASSERT(olen > len, ("Iteration did not make progress?"));
3375 		maybe_yield();
3376 	}
3377 
3378 	return (error);
3379 }
3380 
3381 static u_long vn_lock_pair_pause_cnt;
3382 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3383     &vn_lock_pair_pause_cnt, 0,
3384     "Count of vn_lock_pair deadlocks");
3385 
3386 u_int vn_lock_pair_pause_max;
3387 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3388     &vn_lock_pair_pause_max, 0,
3389     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3390 
3391 static void
3392 vn_lock_pair_pause(const char *wmesg)
3393 {
3394 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3395 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3396 }
3397 
3398 /*
3399  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3400  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3401  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3402  * can be NULL.
3403  *
3404  * The function returns with both vnodes exclusively locked, and
3405  * guarantees that it does not create lock order reversal with other
3406  * threads during its execution.  Both vnodes could be unlocked
3407  * temporary (and reclaimed).
3408  */
3409 void
3410 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3411     bool vp2_locked)
3412 {
3413 	int error;
3414 
3415 	if (vp1 == NULL && vp2 == NULL)
3416 		return;
3417 	if (vp1 != NULL) {
3418 		if (vp1_locked)
3419 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3420 		else
3421 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3422 	} else {
3423 		vp1_locked = true;
3424 	}
3425 	if (vp2 != NULL) {
3426 		if (vp2_locked)
3427 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3428 		else
3429 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3430 	} else {
3431 		vp2_locked = true;
3432 	}
3433 	if (!vp1_locked && !vp2_locked) {
3434 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3435 		vp1_locked = true;
3436 	}
3437 
3438 	for (;;) {
3439 		if (vp1_locked && vp2_locked)
3440 			break;
3441 		if (vp1_locked && vp2 != NULL) {
3442 			if (vp1 != NULL) {
3443 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3444 				    __FILE__, __LINE__);
3445 				if (error == 0)
3446 					break;
3447 				VOP_UNLOCK(vp1);
3448 				vp1_locked = false;
3449 				vn_lock_pair_pause("vlp1");
3450 			}
3451 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3452 			vp2_locked = true;
3453 		}
3454 		if (vp2_locked && vp1 != NULL) {
3455 			if (vp2 != NULL) {
3456 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3457 				    __FILE__, __LINE__);
3458 				if (error == 0)
3459 					break;
3460 				VOP_UNLOCK(vp2);
3461 				vp2_locked = false;
3462 				vn_lock_pair_pause("vlp2");
3463 			}
3464 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3465 			vp1_locked = true;
3466 		}
3467 	}
3468 	if (vp1 != NULL)
3469 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3470 	if (vp2 != NULL)
3471 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3472 }
3473