xref: /freebsd/sys/kern/vfs_vnops.c (revision 77013d11e6483b970af25e13c9b892075742f7e5)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/prng.h>
74 #include <sys/sx.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
78 #include <sys/conf.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
81 #include <sys/user.h>
82 #include <sys/ktrace.h>
83 
84 #include <security/audit/audit.h>
85 #include <security/mac/mac_framework.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_extern.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pager.h>
94 
95 #ifdef HWPMC_HOOKS
96 #include <sys/pmckern.h>
97 #endif
98 
99 static fo_rdwr_t	vn_read;
100 static fo_rdwr_t	vn_write;
101 static fo_rdwr_t	vn_io_fault;
102 static fo_truncate_t	vn_truncate;
103 static fo_ioctl_t	vn_ioctl;
104 static fo_poll_t	vn_poll;
105 static fo_kqfilter_t	vn_kqfilter;
106 static fo_close_t	vn_closefile;
107 static fo_mmap_t	vn_mmap;
108 static fo_fallocate_t	vn_fallocate;
109 
110 struct 	fileops vnops = {
111 	.fo_read = vn_io_fault,
112 	.fo_write = vn_io_fault,
113 	.fo_truncate = vn_truncate,
114 	.fo_ioctl = vn_ioctl,
115 	.fo_poll = vn_poll,
116 	.fo_kqfilter = vn_kqfilter,
117 	.fo_stat = vn_statfile,
118 	.fo_close = vn_closefile,
119 	.fo_chmod = vn_chmod,
120 	.fo_chown = vn_chown,
121 	.fo_sendfile = vn_sendfile,
122 	.fo_seek = vn_seek,
123 	.fo_fill_kinfo = vn_fill_kinfo,
124 	.fo_mmap = vn_mmap,
125 	.fo_fallocate = vn_fallocate,
126 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
127 };
128 
129 const u_int io_hold_cnt = 16;
130 static int vn_io_fault_enable = 1;
131 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
132     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
133 static int vn_io_fault_prefault = 0;
134 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
135     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
136 static int vn_io_pgcache_read_enable = 1;
137 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
138     &vn_io_pgcache_read_enable, 0,
139     "Enable copying from page cache for reads, avoiding fs");
140 static u_long vn_io_faults_cnt;
141 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
142     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
143 
144 static int vfs_allow_read_dir = 0;
145 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
146     &vfs_allow_read_dir, 0,
147     "Enable read(2) of directory by root for filesystems that support it");
148 
149 /*
150  * Returns true if vn_io_fault mode of handling the i/o request should
151  * be used.
152  */
153 static bool
154 do_vn_io_fault(struct vnode *vp, struct uio *uio)
155 {
156 	struct mount *mp;
157 
158 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
159 	    (mp = vp->v_mount) != NULL &&
160 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
161 }
162 
163 /*
164  * Structure used to pass arguments to vn_io_fault1(), to do either
165  * file- or vnode-based I/O calls.
166  */
167 struct vn_io_fault_args {
168 	enum {
169 		VN_IO_FAULT_FOP,
170 		VN_IO_FAULT_VOP
171 	} kind;
172 	struct ucred *cred;
173 	int flags;
174 	union {
175 		struct fop_args_tag {
176 			struct file *fp;
177 			fo_rdwr_t *doio;
178 		} fop_args;
179 		struct vop_args_tag {
180 			struct vnode *vp;
181 		} vop_args;
182 	} args;
183 };
184 
185 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
186     struct vn_io_fault_args *args, struct thread *td);
187 
188 int
189 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
190 {
191 	struct thread *td = ndp->ni_cnd.cn_thread;
192 
193 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
194 }
195 
196 static uint64_t
197 open2nameif(int fmode, u_int vn_open_flags)
198 {
199 	uint64_t res;
200 
201 	res = ISOPEN | LOCKLEAF;
202 	if ((fmode & O_RESOLVE_BENEATH) != 0)
203 		res |= RBENEATH;
204 	if ((fmode & O_EMPTY_PATH) != 0)
205 		res |= EMPTYPATH;
206 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
207 		res |= AUDITVNODE1;
208 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
209 		res |= NOCAPCHECK;
210 	return (res);
211 }
212 
213 /*
214  * Common code for vnode open operations via a name lookup.
215  * Lookup the vnode and invoke VOP_CREATE if needed.
216  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
217  *
218  * Note that this does NOT free nameidata for the successful case,
219  * due to the NDINIT being done elsewhere.
220  */
221 int
222 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
223     struct ucred *cred, struct file *fp)
224 {
225 	struct vnode *vp;
226 	struct mount *mp;
227 	struct thread *td = ndp->ni_cnd.cn_thread;
228 	struct vattr vat;
229 	struct vattr *vap = &vat;
230 	int fmode, error;
231 	bool first_open;
232 
233 restart:
234 	first_open = false;
235 	fmode = *flagp;
236 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
237 	    O_EXCL | O_DIRECTORY) ||
238 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
239 		return (EINVAL);
240 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
241 		ndp->ni_cnd.cn_nameiop = CREATE;
242 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
243 		/*
244 		 * Set NOCACHE to avoid flushing the cache when
245 		 * rolling in many files at once.
246 		 *
247 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
248 		 * exist despite NOCACHE.
249 		 */
250 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
251 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
252 			ndp->ni_cnd.cn_flags |= FOLLOW;
253 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
254 			bwillwrite();
255 		if ((error = namei(ndp)) != 0)
256 			return (error);
257 		if (ndp->ni_vp == NULL) {
258 			VATTR_NULL(vap);
259 			vap->va_type = VREG;
260 			vap->va_mode = cmode;
261 			if (fmode & O_EXCL)
262 				vap->va_vaflags |= VA_EXCLUSIVE;
263 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
264 				NDFREE(ndp, NDF_ONLY_PNBUF);
265 				vput(ndp->ni_dvp);
266 				if ((error = vn_start_write(NULL, &mp,
267 				    V_XSLEEP | PCATCH)) != 0)
268 					return (error);
269 				NDREINIT(ndp);
270 				goto restart;
271 			}
272 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
273 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
274 #ifdef MAC
275 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
276 			    &ndp->ni_cnd, vap);
277 			if (error == 0)
278 #endif
279 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
280 				    &ndp->ni_cnd, vap);
281 			vp = ndp->ni_vp;
282 			if (error == 0 && (fmode & O_EXCL) != 0 &&
283 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
284 				VI_LOCK(vp);
285 				vp->v_iflag |= VI_FOPENING;
286 				VI_UNLOCK(vp);
287 				first_open = true;
288 			}
289 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
290 			    false);
291 			vn_finished_write(mp);
292 			if (error) {
293 				NDFREE(ndp, NDF_ONLY_PNBUF);
294 				if (error == ERELOOKUP) {
295 					NDREINIT(ndp);
296 					goto restart;
297 				}
298 				return (error);
299 			}
300 			fmode &= ~O_TRUNC;
301 		} else {
302 			if (ndp->ni_dvp == ndp->ni_vp)
303 				vrele(ndp->ni_dvp);
304 			else
305 				vput(ndp->ni_dvp);
306 			ndp->ni_dvp = NULL;
307 			vp = ndp->ni_vp;
308 			if (fmode & O_EXCL) {
309 				error = EEXIST;
310 				goto bad;
311 			}
312 			if (vp->v_type == VDIR) {
313 				error = EISDIR;
314 				goto bad;
315 			}
316 			fmode &= ~O_CREAT;
317 		}
318 	} else {
319 		ndp->ni_cnd.cn_nameiop = LOOKUP;
320 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
321 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
322 		    FOLLOW;
323 		if ((fmode & FWRITE) == 0)
324 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
325 		if ((error = namei(ndp)) != 0)
326 			return (error);
327 		vp = ndp->ni_vp;
328 	}
329 	error = vn_open_vnode(vp, fmode, cred, td, fp);
330 	if (first_open) {
331 		VI_LOCK(vp);
332 		vp->v_iflag &= ~VI_FOPENING;
333 		wakeup(vp);
334 		VI_UNLOCK(vp);
335 	}
336 	if (error)
337 		goto bad;
338 	*flagp = fmode;
339 	return (0);
340 bad:
341 	NDFREE(ndp, NDF_ONLY_PNBUF);
342 	vput(vp);
343 	*flagp = fmode;
344 	ndp->ni_vp = NULL;
345 	return (error);
346 }
347 
348 static int
349 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
350 {
351 	struct flock lf;
352 	int error, lock_flags, type;
353 
354 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
355 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
356 		return (0);
357 	KASSERT(fp != NULL, ("open with flock requires fp"));
358 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
359 		return (EOPNOTSUPP);
360 
361 	lock_flags = VOP_ISLOCKED(vp);
362 	VOP_UNLOCK(vp);
363 
364 	lf.l_whence = SEEK_SET;
365 	lf.l_start = 0;
366 	lf.l_len = 0;
367 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
368 	type = F_FLOCK;
369 	if ((fmode & FNONBLOCK) == 0)
370 		type |= F_WAIT;
371 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
372 		type |= F_FIRSTOPEN;
373 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
374 	if (error == 0)
375 		fp->f_flag |= FHASLOCK;
376 
377 	vn_lock(vp, lock_flags | LK_RETRY);
378 	return (error);
379 }
380 
381 /*
382  * Common code for vnode open operations once a vnode is located.
383  * Check permissions, and call the VOP_OPEN routine.
384  */
385 int
386 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
387     struct thread *td, struct file *fp)
388 {
389 	accmode_t accmode;
390 	int error;
391 
392 	if (vp->v_type == VLNK) {
393 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
394 			return (EMLINK);
395 	}
396 	if (vp->v_type == VSOCK)
397 		return (EOPNOTSUPP);
398 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
399 		return (ENOTDIR);
400 
401 	accmode = 0;
402 	if ((fmode & O_PATH) == 0) {
403 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
404 			if (vp->v_type == VDIR)
405 				return (EISDIR);
406 			accmode |= VWRITE;
407 		}
408 		if ((fmode & FREAD) != 0)
409 			accmode |= VREAD;
410 		if ((fmode & O_APPEND) && (fmode & FWRITE))
411 			accmode |= VAPPEND;
412 #ifdef MAC
413 		if ((fmode & O_CREAT) != 0)
414 			accmode |= VCREAT;
415 #endif
416 	}
417 	if ((fmode & FEXEC) != 0)
418 		accmode |= VEXEC;
419 #ifdef MAC
420 	if ((fmode & O_VERIFY) != 0)
421 		accmode |= VVERIFY;
422 	error = mac_vnode_check_open(cred, vp, accmode);
423 	if (error != 0)
424 		return (error);
425 
426 	accmode &= ~(VCREAT | VVERIFY);
427 #endif
428 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
429 		error = VOP_ACCESS(vp, accmode, cred, td);
430 		if (error != 0)
431 			return (error);
432 	}
433 	if ((fmode & O_PATH) != 0) {
434 		if (vp->v_type == VFIFO)
435 			error = EPIPE;
436 		else
437 			error = VOP_ACCESS(vp, VREAD, cred, td);
438 		if (error == 0)
439 			fp->f_flag |= FKQALLOWED;
440 		return (0);
441 	}
442 
443 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
444 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
445 	error = VOP_OPEN(vp, fmode, cred, td, fp);
446 	if (error != 0)
447 		return (error);
448 
449 	error = vn_open_vnode_advlock(vp, fmode, fp);
450 	if (error == 0 && (fmode & FWRITE) != 0) {
451 		error = VOP_ADD_WRITECOUNT(vp, 1);
452 		if (error == 0) {
453 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
454 			     __func__, vp, vp->v_writecount);
455 		}
456 	}
457 
458 	/*
459 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
460 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
461 	 */
462 	if (error != 0) {
463 		if (fp != NULL) {
464 			/*
465 			 * Arrange the call by having fdrop() to use
466 			 * vn_closefile().  This is to satisfy
467 			 * filesystems like devfs or tmpfs, which
468 			 * override fo_close().
469 			 */
470 			fp->f_flag |= FOPENFAILED;
471 			fp->f_vnode = vp;
472 			if (fp->f_ops == &badfileops) {
473 				fp->f_type = DTYPE_VNODE;
474 				fp->f_ops = &vnops;
475 			}
476 			vref(vp);
477 		} else {
478 			/*
479 			 * If there is no fp, due to kernel-mode open,
480 			 * we can call VOP_CLOSE() now.
481 			 */
482 			if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
483 			    !MNT_EXTENDED_SHARED(vp->v_mount) &&
484 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
485 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
486 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
487 			    cred, td);
488 		}
489 	}
490 
491 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
492 	return (error);
493 
494 }
495 
496 /*
497  * Check for write permissions on the specified vnode.
498  * Prototype text segments cannot be written.
499  * It is racy.
500  */
501 int
502 vn_writechk(struct vnode *vp)
503 {
504 
505 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
506 	/*
507 	 * If there's shared text associated with
508 	 * the vnode, try to free it up once.  If
509 	 * we fail, we can't allow writing.
510 	 */
511 	if (VOP_IS_TEXT(vp))
512 		return (ETXTBSY);
513 
514 	return (0);
515 }
516 
517 /*
518  * Vnode close call
519  */
520 static int
521 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
522     struct thread *td, bool keep_ref)
523 {
524 	struct mount *mp;
525 	int error, lock_flags;
526 
527 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
528 	    MNT_EXTENDED_SHARED(vp->v_mount))
529 		lock_flags = LK_SHARED;
530 	else
531 		lock_flags = LK_EXCLUSIVE;
532 
533 	vn_start_write(vp, &mp, V_WAIT);
534 	vn_lock(vp, lock_flags | LK_RETRY);
535 	AUDIT_ARG_VNODE1(vp);
536 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
537 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
538 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
539 		    __func__, vp, vp->v_writecount);
540 	}
541 	error = VOP_CLOSE(vp, flags, file_cred, td);
542 	if (keep_ref)
543 		VOP_UNLOCK(vp);
544 	else
545 		vput(vp);
546 	vn_finished_write(mp);
547 	return (error);
548 }
549 
550 int
551 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
552     struct thread *td)
553 {
554 
555 	return (vn_close1(vp, flags, file_cred, td, false));
556 }
557 
558 /*
559  * Heuristic to detect sequential operation.
560  */
561 static int
562 sequential_heuristic(struct uio *uio, struct file *fp)
563 {
564 	enum uio_rw rw;
565 
566 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
567 
568 	rw = uio->uio_rw;
569 	if (fp->f_flag & FRDAHEAD)
570 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
571 
572 	/*
573 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
574 	 * that the first I/O is normally considered to be slightly
575 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
576 	 * unless previous seeks have reduced f_seqcount to 0, in which
577 	 * case offset 0 is not special.
578 	 */
579 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
580 	    uio->uio_offset == fp->f_nextoff[rw]) {
581 		/*
582 		 * f_seqcount is in units of fixed-size blocks so that it
583 		 * depends mainly on the amount of sequential I/O and not
584 		 * much on the number of sequential I/O's.  The fixed size
585 		 * of 16384 is hard-coded here since it is (not quite) just
586 		 * a magic size that works well here.  This size is more
587 		 * closely related to the best I/O size for real disks than
588 		 * to any block size used by software.
589 		 */
590 		if (uio->uio_resid >= IO_SEQMAX * 16384)
591 			fp->f_seqcount[rw] = IO_SEQMAX;
592 		else {
593 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
594 			if (fp->f_seqcount[rw] > IO_SEQMAX)
595 				fp->f_seqcount[rw] = IO_SEQMAX;
596 		}
597 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
598 	}
599 
600 	/* Not sequential.  Quickly draw-down sequentiality. */
601 	if (fp->f_seqcount[rw] > 1)
602 		fp->f_seqcount[rw] = 1;
603 	else
604 		fp->f_seqcount[rw] = 0;
605 	return (0);
606 }
607 
608 /*
609  * Package up an I/O request on a vnode into a uio and do it.
610  */
611 int
612 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
613     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
614     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
615 {
616 	struct uio auio;
617 	struct iovec aiov;
618 	struct mount *mp;
619 	struct ucred *cred;
620 	void *rl_cookie;
621 	struct vn_io_fault_args args;
622 	int error, lock_flags;
623 
624 	if (offset < 0 && vp->v_type != VCHR)
625 		return (EINVAL);
626 	auio.uio_iov = &aiov;
627 	auio.uio_iovcnt = 1;
628 	aiov.iov_base = base;
629 	aiov.iov_len = len;
630 	auio.uio_resid = len;
631 	auio.uio_offset = offset;
632 	auio.uio_segflg = segflg;
633 	auio.uio_rw = rw;
634 	auio.uio_td = td;
635 	error = 0;
636 
637 	if ((ioflg & IO_NODELOCKED) == 0) {
638 		if ((ioflg & IO_RANGELOCKED) == 0) {
639 			if (rw == UIO_READ) {
640 				rl_cookie = vn_rangelock_rlock(vp, offset,
641 				    offset + len);
642 			} else if ((ioflg & IO_APPEND) != 0) {
643 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
644 			} else {
645 				rl_cookie = vn_rangelock_wlock(vp, offset,
646 				    offset + len);
647 			}
648 		} else
649 			rl_cookie = NULL;
650 		mp = NULL;
651 		if (rw == UIO_WRITE) {
652 			if (vp->v_type != VCHR &&
653 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
654 			    != 0)
655 				goto out;
656 			lock_flags = vn_lktype_write(mp, vp);
657 		} else
658 			lock_flags = LK_SHARED;
659 		vn_lock(vp, lock_flags | LK_RETRY);
660 	} else
661 		rl_cookie = NULL;
662 
663 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
664 #ifdef MAC
665 	if ((ioflg & IO_NOMACCHECK) == 0) {
666 		if (rw == UIO_READ)
667 			error = mac_vnode_check_read(active_cred, file_cred,
668 			    vp);
669 		else
670 			error = mac_vnode_check_write(active_cred, file_cred,
671 			    vp);
672 	}
673 #endif
674 	if (error == 0) {
675 		if (file_cred != NULL)
676 			cred = file_cred;
677 		else
678 			cred = active_cred;
679 		if (do_vn_io_fault(vp, &auio)) {
680 			args.kind = VN_IO_FAULT_VOP;
681 			args.cred = cred;
682 			args.flags = ioflg;
683 			args.args.vop_args.vp = vp;
684 			error = vn_io_fault1(vp, &auio, &args, td);
685 		} else if (rw == UIO_READ) {
686 			error = VOP_READ(vp, &auio, ioflg, cred);
687 		} else /* if (rw == UIO_WRITE) */ {
688 			error = VOP_WRITE(vp, &auio, ioflg, cred);
689 		}
690 	}
691 	if (aresid)
692 		*aresid = auio.uio_resid;
693 	else
694 		if (auio.uio_resid && error == 0)
695 			error = EIO;
696 	if ((ioflg & IO_NODELOCKED) == 0) {
697 		VOP_UNLOCK(vp);
698 		if (mp != NULL)
699 			vn_finished_write(mp);
700 	}
701  out:
702 	if (rl_cookie != NULL)
703 		vn_rangelock_unlock(vp, rl_cookie);
704 	return (error);
705 }
706 
707 /*
708  * Package up an I/O request on a vnode into a uio and do it.  The I/O
709  * request is split up into smaller chunks and we try to avoid saturating
710  * the buffer cache while potentially holding a vnode locked, so we
711  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
712  * to give other processes a chance to lock the vnode (either other processes
713  * core'ing the same binary, or unrelated processes scanning the directory).
714  */
715 int
716 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
717     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
718     struct ucred *file_cred, size_t *aresid, struct thread *td)
719 {
720 	int error = 0;
721 	ssize_t iaresid;
722 
723 	do {
724 		int chunk;
725 
726 		/*
727 		 * Force `offset' to a multiple of MAXBSIZE except possibly
728 		 * for the first chunk, so that filesystems only need to
729 		 * write full blocks except possibly for the first and last
730 		 * chunks.
731 		 */
732 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
733 
734 		if (chunk > len)
735 			chunk = len;
736 		if (rw != UIO_READ && vp->v_type == VREG)
737 			bwillwrite();
738 		iaresid = 0;
739 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
740 		    ioflg, active_cred, file_cred, &iaresid, td);
741 		len -= chunk;	/* aresid calc already includes length */
742 		if (error)
743 			break;
744 		offset += chunk;
745 		base = (char *)base + chunk;
746 		kern_yield(PRI_USER);
747 	} while (len);
748 	if (aresid)
749 		*aresid = len + iaresid;
750 	return (error);
751 }
752 
753 #if OFF_MAX <= LONG_MAX
754 off_t
755 foffset_lock(struct file *fp, int flags)
756 {
757 	volatile short *flagsp;
758 	off_t res;
759 	short state;
760 
761 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
762 
763 	if ((flags & FOF_NOLOCK) != 0)
764 		return (atomic_load_long(&fp->f_offset));
765 
766 	/*
767 	 * According to McKusick the vn lock was protecting f_offset here.
768 	 * It is now protected by the FOFFSET_LOCKED flag.
769 	 */
770 	flagsp = &fp->f_vnread_flags;
771 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
772 		return (atomic_load_long(&fp->f_offset));
773 
774 	sleepq_lock(&fp->f_vnread_flags);
775 	state = atomic_load_16(flagsp);
776 	for (;;) {
777 		if ((state & FOFFSET_LOCKED) == 0) {
778 			if (!atomic_fcmpset_acq_16(flagsp, &state,
779 			    FOFFSET_LOCKED))
780 				continue;
781 			break;
782 		}
783 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
784 			if (!atomic_fcmpset_acq_16(flagsp, &state,
785 			    state | FOFFSET_LOCK_WAITING))
786 				continue;
787 		}
788 		DROP_GIANT();
789 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
790 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
791 		PICKUP_GIANT();
792 		sleepq_lock(&fp->f_vnread_flags);
793 		state = atomic_load_16(flagsp);
794 	}
795 	res = atomic_load_long(&fp->f_offset);
796 	sleepq_release(&fp->f_vnread_flags);
797 	return (res);
798 }
799 
800 void
801 foffset_unlock(struct file *fp, off_t val, int flags)
802 {
803 	volatile short *flagsp;
804 	short state;
805 
806 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
807 
808 	if ((flags & FOF_NOUPDATE) == 0)
809 		atomic_store_long(&fp->f_offset, val);
810 	if ((flags & FOF_NEXTOFF_R) != 0)
811 		fp->f_nextoff[UIO_READ] = val;
812 	if ((flags & FOF_NEXTOFF_W) != 0)
813 		fp->f_nextoff[UIO_WRITE] = val;
814 
815 	if ((flags & FOF_NOLOCK) != 0)
816 		return;
817 
818 	flagsp = &fp->f_vnread_flags;
819 	state = atomic_load_16(flagsp);
820 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
821 	    atomic_cmpset_rel_16(flagsp, state, 0))
822 		return;
823 
824 	sleepq_lock(&fp->f_vnread_flags);
825 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
826 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
827 	fp->f_vnread_flags = 0;
828 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
829 	sleepq_release(&fp->f_vnread_flags);
830 }
831 #else
832 off_t
833 foffset_lock(struct file *fp, int flags)
834 {
835 	struct mtx *mtxp;
836 	off_t res;
837 
838 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
839 
840 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
841 	mtx_lock(mtxp);
842 	if ((flags & FOF_NOLOCK) == 0) {
843 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
844 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
845 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
846 			    "vofflock", 0);
847 		}
848 		fp->f_vnread_flags |= FOFFSET_LOCKED;
849 	}
850 	res = fp->f_offset;
851 	mtx_unlock(mtxp);
852 	return (res);
853 }
854 
855 void
856 foffset_unlock(struct file *fp, off_t val, int flags)
857 {
858 	struct mtx *mtxp;
859 
860 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
861 
862 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
863 	mtx_lock(mtxp);
864 	if ((flags & FOF_NOUPDATE) == 0)
865 		fp->f_offset = val;
866 	if ((flags & FOF_NEXTOFF_R) != 0)
867 		fp->f_nextoff[UIO_READ] = val;
868 	if ((flags & FOF_NEXTOFF_W) != 0)
869 		fp->f_nextoff[UIO_WRITE] = val;
870 	if ((flags & FOF_NOLOCK) == 0) {
871 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
872 		    ("Lost FOFFSET_LOCKED"));
873 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
874 			wakeup(&fp->f_vnread_flags);
875 		fp->f_vnread_flags = 0;
876 	}
877 	mtx_unlock(mtxp);
878 }
879 #endif
880 
881 void
882 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
883 {
884 
885 	if ((flags & FOF_OFFSET) == 0)
886 		uio->uio_offset = foffset_lock(fp, flags);
887 }
888 
889 void
890 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
891 {
892 
893 	if ((flags & FOF_OFFSET) == 0)
894 		foffset_unlock(fp, uio->uio_offset, flags);
895 }
896 
897 static int
898 get_advice(struct file *fp, struct uio *uio)
899 {
900 	struct mtx *mtxp;
901 	int ret;
902 
903 	ret = POSIX_FADV_NORMAL;
904 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
905 		return (ret);
906 
907 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
908 	mtx_lock(mtxp);
909 	if (fp->f_advice != NULL &&
910 	    uio->uio_offset >= fp->f_advice->fa_start &&
911 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
912 		ret = fp->f_advice->fa_advice;
913 	mtx_unlock(mtxp);
914 	return (ret);
915 }
916 
917 int
918 vn_read_from_obj(struct vnode *vp, struct uio *uio)
919 {
920 	vm_object_t obj;
921 	vm_page_t ma[io_hold_cnt + 2];
922 	off_t off, vsz;
923 	ssize_t resid;
924 	int error, i, j;
925 
926 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
927 	obj = atomic_load_ptr(&vp->v_object);
928 	if (obj == NULL)
929 		return (EJUSTRETURN);
930 
931 	/*
932 	 * Depends on type stability of vm_objects.
933 	 */
934 	vm_object_pip_add(obj, 1);
935 	if ((obj->flags & OBJ_DEAD) != 0) {
936 		/*
937 		 * Note that object might be already reused from the
938 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
939 		 * we recheck for DOOMED vnode state after all pages
940 		 * are busied, and retract then.
941 		 *
942 		 * But we check for OBJ_DEAD to ensure that we do not
943 		 * busy pages while vm_object_terminate_pages()
944 		 * processes the queue.
945 		 */
946 		error = EJUSTRETURN;
947 		goto out_pip;
948 	}
949 
950 	resid = uio->uio_resid;
951 	off = uio->uio_offset;
952 	for (i = 0; resid > 0; i++) {
953 		MPASS(i < io_hold_cnt + 2);
954 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
955 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
956 		    VM_ALLOC_NOWAIT);
957 		if (ma[i] == NULL)
958 			break;
959 
960 		/*
961 		 * Skip invalid pages.  Valid mask can be partial only
962 		 * at EOF, and we clip later.
963 		 */
964 		if (vm_page_none_valid(ma[i])) {
965 			vm_page_sunbusy(ma[i]);
966 			break;
967 		}
968 
969 		resid -= PAGE_SIZE;
970 		off += PAGE_SIZE;
971 	}
972 	if (i == 0) {
973 		error = EJUSTRETURN;
974 		goto out_pip;
975 	}
976 
977 	/*
978 	 * Check VIRF_DOOMED after we busied our pages.  Since
979 	 * vgonel() terminates the vnode' vm_object, it cannot
980 	 * process past pages busied by us.
981 	 */
982 	if (VN_IS_DOOMED(vp)) {
983 		error = EJUSTRETURN;
984 		goto out;
985 	}
986 
987 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
988 	if (resid > uio->uio_resid)
989 		resid = uio->uio_resid;
990 
991 	/*
992 	 * Unlocked read of vnp_size is safe because truncation cannot
993 	 * pass busied page.  But we load vnp_size into a local
994 	 * variable so that possible concurrent extension does not
995 	 * break calculation.
996 	 */
997 #if defined(__powerpc__) && !defined(__powerpc64__)
998 	vsz = obj->un_pager.vnp.vnp_size;
999 #else
1000 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1001 #endif
1002 	if (uio->uio_offset >= vsz) {
1003 		error = EJUSTRETURN;
1004 		goto out;
1005 	}
1006 	if (uio->uio_offset + resid > vsz)
1007 		resid = vsz - uio->uio_offset;
1008 
1009 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1010 
1011 out:
1012 	for (j = 0; j < i; j++) {
1013 		if (error == 0)
1014 			vm_page_reference(ma[j]);
1015 		vm_page_sunbusy(ma[j]);
1016 	}
1017 out_pip:
1018 	vm_object_pip_wakeup(obj);
1019 	if (error != 0)
1020 		return (error);
1021 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1022 }
1023 
1024 /*
1025  * File table vnode read routine.
1026  */
1027 static int
1028 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1029     struct thread *td)
1030 {
1031 	struct vnode *vp;
1032 	off_t orig_offset;
1033 	int error, ioflag;
1034 	int advice;
1035 
1036 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1037 	    uio->uio_td, td));
1038 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1039 	vp = fp->f_vnode;
1040 	ioflag = 0;
1041 	if (fp->f_flag & FNONBLOCK)
1042 		ioflag |= IO_NDELAY;
1043 	if (fp->f_flag & O_DIRECT)
1044 		ioflag |= IO_DIRECT;
1045 
1046 	/*
1047 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1048 	 * allows us to avoid unneeded work outright.
1049 	 */
1050 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1051 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1052 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1053 		if (error == 0) {
1054 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1055 			return (0);
1056 		}
1057 		if (error != EJUSTRETURN)
1058 			return (error);
1059 	}
1060 
1061 	advice = get_advice(fp, uio);
1062 	vn_lock(vp, LK_SHARED | LK_RETRY);
1063 
1064 	switch (advice) {
1065 	case POSIX_FADV_NORMAL:
1066 	case POSIX_FADV_SEQUENTIAL:
1067 	case POSIX_FADV_NOREUSE:
1068 		ioflag |= sequential_heuristic(uio, fp);
1069 		break;
1070 	case POSIX_FADV_RANDOM:
1071 		/* Disable read-ahead for random I/O. */
1072 		break;
1073 	}
1074 	orig_offset = uio->uio_offset;
1075 
1076 #ifdef MAC
1077 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1078 	if (error == 0)
1079 #endif
1080 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1081 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1082 	VOP_UNLOCK(vp);
1083 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1084 	    orig_offset != uio->uio_offset)
1085 		/*
1086 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1087 		 * for the backing file after a POSIX_FADV_NOREUSE
1088 		 * read(2).
1089 		 */
1090 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1091 		    POSIX_FADV_DONTNEED);
1092 	return (error);
1093 }
1094 
1095 /*
1096  * File table vnode write routine.
1097  */
1098 static int
1099 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1100     struct thread *td)
1101 {
1102 	struct vnode *vp;
1103 	struct mount *mp;
1104 	off_t orig_offset;
1105 	int error, ioflag;
1106 	int advice;
1107 	bool need_finished_write;
1108 
1109 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1110 	    uio->uio_td, td));
1111 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1112 	vp = fp->f_vnode;
1113 	if (vp->v_type == VREG)
1114 		bwillwrite();
1115 	ioflag = IO_UNIT;
1116 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1117 		ioflag |= IO_APPEND;
1118 	if (fp->f_flag & FNONBLOCK)
1119 		ioflag |= IO_NDELAY;
1120 	if (fp->f_flag & O_DIRECT)
1121 		ioflag |= IO_DIRECT;
1122 
1123 	mp = atomic_load_ptr(&vp->v_mount);
1124 	if ((fp->f_flag & O_FSYNC) ||
1125 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS)))
1126 		ioflag |= IO_SYNC;
1127 
1128 	/*
1129 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1130 	 * implementations that don't understand IO_DATASYNC fall back to full
1131 	 * O_SYNC behavior.
1132 	 */
1133 	if (fp->f_flag & O_DSYNC)
1134 		ioflag |= IO_SYNC | IO_DATASYNC;
1135 	mp = NULL;
1136 	need_finished_write = false;
1137 	if (vp->v_type != VCHR) {
1138 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1139 		if (error != 0)
1140 			goto unlock;
1141 		need_finished_write = true;
1142 	}
1143 
1144 	advice = get_advice(fp, uio);
1145 
1146 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1147 	switch (advice) {
1148 	case POSIX_FADV_NORMAL:
1149 	case POSIX_FADV_SEQUENTIAL:
1150 	case POSIX_FADV_NOREUSE:
1151 		ioflag |= sequential_heuristic(uio, fp);
1152 		break;
1153 	case POSIX_FADV_RANDOM:
1154 		/* XXX: Is this correct? */
1155 		break;
1156 	}
1157 	orig_offset = uio->uio_offset;
1158 
1159 #ifdef MAC
1160 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1161 	if (error == 0)
1162 #endif
1163 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1164 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1165 	VOP_UNLOCK(vp);
1166 	if (need_finished_write)
1167 		vn_finished_write(mp);
1168 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1169 	    orig_offset != uio->uio_offset)
1170 		/*
1171 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1172 		 * for the backing file after a POSIX_FADV_NOREUSE
1173 		 * write(2).
1174 		 */
1175 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1176 		    POSIX_FADV_DONTNEED);
1177 unlock:
1178 	return (error);
1179 }
1180 
1181 /*
1182  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1183  * prevent the following deadlock:
1184  *
1185  * Assume that the thread A reads from the vnode vp1 into userspace
1186  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1187  * currently not resident, then system ends up with the call chain
1188  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1189  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1190  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1191  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1192  * backed by the pages of vnode vp1, and some page in buf2 is not
1193  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1194  *
1195  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1196  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1197  * Instead, it first tries to do the whole range i/o with pagefaults
1198  * disabled. If all pages in the i/o buffer are resident and mapped,
1199  * VOP will succeed (ignoring the genuine filesystem errors).
1200  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1201  * i/o in chunks, with all pages in the chunk prefaulted and held
1202  * using vm_fault_quick_hold_pages().
1203  *
1204  * Filesystems using this deadlock avoidance scheme should use the
1205  * array of the held pages from uio, saved in the curthread->td_ma,
1206  * instead of doing uiomove().  A helper function
1207  * vn_io_fault_uiomove() converts uiomove request into
1208  * uiomove_fromphys() over td_ma array.
1209  *
1210  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1211  * make the current i/o request atomic with respect to other i/os and
1212  * truncations.
1213  */
1214 
1215 /*
1216  * Decode vn_io_fault_args and perform the corresponding i/o.
1217  */
1218 static int
1219 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1220     struct thread *td)
1221 {
1222 	int error, save;
1223 
1224 	error = 0;
1225 	save = vm_fault_disable_pagefaults();
1226 	switch (args->kind) {
1227 	case VN_IO_FAULT_FOP:
1228 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1229 		    uio, args->cred, args->flags, td);
1230 		break;
1231 	case VN_IO_FAULT_VOP:
1232 		if (uio->uio_rw == UIO_READ) {
1233 			error = VOP_READ(args->args.vop_args.vp, uio,
1234 			    args->flags, args->cred);
1235 		} else if (uio->uio_rw == UIO_WRITE) {
1236 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1237 			    args->flags, args->cred);
1238 		}
1239 		break;
1240 	default:
1241 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1242 		    args->kind, uio->uio_rw);
1243 	}
1244 	vm_fault_enable_pagefaults(save);
1245 	return (error);
1246 }
1247 
1248 static int
1249 vn_io_fault_touch(char *base, const struct uio *uio)
1250 {
1251 	int r;
1252 
1253 	r = fubyte(base);
1254 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1255 		return (EFAULT);
1256 	return (0);
1257 }
1258 
1259 static int
1260 vn_io_fault_prefault_user(const struct uio *uio)
1261 {
1262 	char *base;
1263 	const struct iovec *iov;
1264 	size_t len;
1265 	ssize_t resid;
1266 	int error, i;
1267 
1268 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1269 	    ("vn_io_fault_prefault userspace"));
1270 
1271 	error = i = 0;
1272 	iov = uio->uio_iov;
1273 	resid = uio->uio_resid;
1274 	base = iov->iov_base;
1275 	len = iov->iov_len;
1276 	while (resid > 0) {
1277 		error = vn_io_fault_touch(base, uio);
1278 		if (error != 0)
1279 			break;
1280 		if (len < PAGE_SIZE) {
1281 			if (len != 0) {
1282 				error = vn_io_fault_touch(base + len - 1, uio);
1283 				if (error != 0)
1284 					break;
1285 				resid -= len;
1286 			}
1287 			if (++i >= uio->uio_iovcnt)
1288 				break;
1289 			iov = uio->uio_iov + i;
1290 			base = iov->iov_base;
1291 			len = iov->iov_len;
1292 		} else {
1293 			len -= PAGE_SIZE;
1294 			base += PAGE_SIZE;
1295 			resid -= PAGE_SIZE;
1296 		}
1297 	}
1298 	return (error);
1299 }
1300 
1301 /*
1302  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1303  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1304  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1305  * into args and call vn_io_fault1() to handle faults during the user
1306  * mode buffer accesses.
1307  */
1308 static int
1309 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1310     struct thread *td)
1311 {
1312 	vm_page_t ma[io_hold_cnt + 2];
1313 	struct uio *uio_clone, short_uio;
1314 	struct iovec short_iovec[1];
1315 	vm_page_t *prev_td_ma;
1316 	vm_prot_t prot;
1317 	vm_offset_t addr, end;
1318 	size_t len, resid;
1319 	ssize_t adv;
1320 	int error, cnt, saveheld, prev_td_ma_cnt;
1321 
1322 	if (vn_io_fault_prefault) {
1323 		error = vn_io_fault_prefault_user(uio);
1324 		if (error != 0)
1325 			return (error); /* Or ignore ? */
1326 	}
1327 
1328 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1329 
1330 	/*
1331 	 * The UFS follows IO_UNIT directive and replays back both
1332 	 * uio_offset and uio_resid if an error is encountered during the
1333 	 * operation.  But, since the iovec may be already advanced,
1334 	 * uio is still in an inconsistent state.
1335 	 *
1336 	 * Cache a copy of the original uio, which is advanced to the redo
1337 	 * point using UIO_NOCOPY below.
1338 	 */
1339 	uio_clone = cloneuio(uio);
1340 	resid = uio->uio_resid;
1341 
1342 	short_uio.uio_segflg = UIO_USERSPACE;
1343 	short_uio.uio_rw = uio->uio_rw;
1344 	short_uio.uio_td = uio->uio_td;
1345 
1346 	error = vn_io_fault_doio(args, uio, td);
1347 	if (error != EFAULT)
1348 		goto out;
1349 
1350 	atomic_add_long(&vn_io_faults_cnt, 1);
1351 	uio_clone->uio_segflg = UIO_NOCOPY;
1352 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1353 	uio_clone->uio_segflg = uio->uio_segflg;
1354 
1355 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1356 	prev_td_ma = td->td_ma;
1357 	prev_td_ma_cnt = td->td_ma_cnt;
1358 
1359 	while (uio_clone->uio_resid != 0) {
1360 		len = uio_clone->uio_iov->iov_len;
1361 		if (len == 0) {
1362 			KASSERT(uio_clone->uio_iovcnt >= 1,
1363 			    ("iovcnt underflow"));
1364 			uio_clone->uio_iov++;
1365 			uio_clone->uio_iovcnt--;
1366 			continue;
1367 		}
1368 		if (len > ptoa(io_hold_cnt))
1369 			len = ptoa(io_hold_cnt);
1370 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1371 		end = round_page(addr + len);
1372 		if (end < addr) {
1373 			error = EFAULT;
1374 			break;
1375 		}
1376 		cnt = atop(end - trunc_page(addr));
1377 		/*
1378 		 * A perfectly misaligned address and length could cause
1379 		 * both the start and the end of the chunk to use partial
1380 		 * page.  +2 accounts for such a situation.
1381 		 */
1382 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1383 		    addr, len, prot, ma, io_hold_cnt + 2);
1384 		if (cnt == -1) {
1385 			error = EFAULT;
1386 			break;
1387 		}
1388 		short_uio.uio_iov = &short_iovec[0];
1389 		short_iovec[0].iov_base = (void *)addr;
1390 		short_uio.uio_iovcnt = 1;
1391 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1392 		short_uio.uio_offset = uio_clone->uio_offset;
1393 		td->td_ma = ma;
1394 		td->td_ma_cnt = cnt;
1395 
1396 		error = vn_io_fault_doio(args, &short_uio, td);
1397 		vm_page_unhold_pages(ma, cnt);
1398 		adv = len - short_uio.uio_resid;
1399 
1400 		uio_clone->uio_iov->iov_base =
1401 		    (char *)uio_clone->uio_iov->iov_base + adv;
1402 		uio_clone->uio_iov->iov_len -= adv;
1403 		uio_clone->uio_resid -= adv;
1404 		uio_clone->uio_offset += adv;
1405 
1406 		uio->uio_resid -= adv;
1407 		uio->uio_offset += adv;
1408 
1409 		if (error != 0 || adv == 0)
1410 			break;
1411 	}
1412 	td->td_ma = prev_td_ma;
1413 	td->td_ma_cnt = prev_td_ma_cnt;
1414 	curthread_pflags_restore(saveheld);
1415 out:
1416 	free(uio_clone, M_IOV);
1417 	return (error);
1418 }
1419 
1420 static int
1421 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1422     int flags, struct thread *td)
1423 {
1424 	fo_rdwr_t *doio;
1425 	struct vnode *vp;
1426 	void *rl_cookie;
1427 	struct vn_io_fault_args args;
1428 	int error;
1429 
1430 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1431 	vp = fp->f_vnode;
1432 
1433 	/*
1434 	 * The ability to read(2) on a directory has historically been
1435 	 * allowed for all users, but this can and has been the source of
1436 	 * at least one security issue in the past.  As such, it is now hidden
1437 	 * away behind a sysctl for those that actually need it to use it, and
1438 	 * restricted to root when it's turned on to make it relatively safe to
1439 	 * leave on for longer sessions of need.
1440 	 */
1441 	if (vp->v_type == VDIR) {
1442 		KASSERT(uio->uio_rw == UIO_READ,
1443 		    ("illegal write attempted on a directory"));
1444 		if (!vfs_allow_read_dir)
1445 			return (EISDIR);
1446 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1447 			return (EISDIR);
1448 	}
1449 
1450 	foffset_lock_uio(fp, uio, flags);
1451 	if (do_vn_io_fault(vp, uio)) {
1452 		args.kind = VN_IO_FAULT_FOP;
1453 		args.args.fop_args.fp = fp;
1454 		args.args.fop_args.doio = doio;
1455 		args.cred = active_cred;
1456 		args.flags = flags | FOF_OFFSET;
1457 		if (uio->uio_rw == UIO_READ) {
1458 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1459 			    uio->uio_offset + uio->uio_resid);
1460 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1461 		    (flags & FOF_OFFSET) == 0) {
1462 			/* For appenders, punt and lock the whole range. */
1463 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1464 		} else {
1465 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1466 			    uio->uio_offset + uio->uio_resid);
1467 		}
1468 		error = vn_io_fault1(vp, uio, &args, td);
1469 		vn_rangelock_unlock(vp, rl_cookie);
1470 	} else {
1471 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1472 	}
1473 	foffset_unlock_uio(fp, uio, flags);
1474 	return (error);
1475 }
1476 
1477 /*
1478  * Helper function to perform the requested uiomove operation using
1479  * the held pages for io->uio_iov[0].iov_base buffer instead of
1480  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1481  * instead of iov_base prevents page faults that could occur due to
1482  * pmap_collect() invalidating the mapping created by
1483  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1484  * object cleanup revoking the write access from page mappings.
1485  *
1486  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1487  * instead of plain uiomove().
1488  */
1489 int
1490 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1491 {
1492 	struct uio transp_uio;
1493 	struct iovec transp_iov[1];
1494 	struct thread *td;
1495 	size_t adv;
1496 	int error, pgadv;
1497 
1498 	td = curthread;
1499 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1500 	    uio->uio_segflg != UIO_USERSPACE)
1501 		return (uiomove(data, xfersize, uio));
1502 
1503 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1504 	transp_iov[0].iov_base = data;
1505 	transp_uio.uio_iov = &transp_iov[0];
1506 	transp_uio.uio_iovcnt = 1;
1507 	if (xfersize > uio->uio_resid)
1508 		xfersize = uio->uio_resid;
1509 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1510 	transp_uio.uio_offset = 0;
1511 	transp_uio.uio_segflg = UIO_SYSSPACE;
1512 	/*
1513 	 * Since transp_iov points to data, and td_ma page array
1514 	 * corresponds to original uio->uio_iov, we need to invert the
1515 	 * direction of the i/o operation as passed to
1516 	 * uiomove_fromphys().
1517 	 */
1518 	switch (uio->uio_rw) {
1519 	case UIO_WRITE:
1520 		transp_uio.uio_rw = UIO_READ;
1521 		break;
1522 	case UIO_READ:
1523 		transp_uio.uio_rw = UIO_WRITE;
1524 		break;
1525 	}
1526 	transp_uio.uio_td = uio->uio_td;
1527 	error = uiomove_fromphys(td->td_ma,
1528 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1529 	    xfersize, &transp_uio);
1530 	adv = xfersize - transp_uio.uio_resid;
1531 	pgadv =
1532 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1533 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1534 	td->td_ma += pgadv;
1535 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1536 	    pgadv));
1537 	td->td_ma_cnt -= pgadv;
1538 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1539 	uio->uio_iov->iov_len -= adv;
1540 	uio->uio_resid -= adv;
1541 	uio->uio_offset += adv;
1542 	return (error);
1543 }
1544 
1545 int
1546 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1547     struct uio *uio)
1548 {
1549 	struct thread *td;
1550 	vm_offset_t iov_base;
1551 	int cnt, pgadv;
1552 
1553 	td = curthread;
1554 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1555 	    uio->uio_segflg != UIO_USERSPACE)
1556 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1557 
1558 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1559 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1560 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1561 	switch (uio->uio_rw) {
1562 	case UIO_WRITE:
1563 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1564 		    offset, cnt);
1565 		break;
1566 	case UIO_READ:
1567 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1568 		    cnt);
1569 		break;
1570 	}
1571 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1572 	td->td_ma += pgadv;
1573 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1574 	    pgadv));
1575 	td->td_ma_cnt -= pgadv;
1576 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1577 	uio->uio_iov->iov_len -= cnt;
1578 	uio->uio_resid -= cnt;
1579 	uio->uio_offset += cnt;
1580 	return (0);
1581 }
1582 
1583 /*
1584  * File table truncate routine.
1585  */
1586 static int
1587 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1588     struct thread *td)
1589 {
1590 	struct mount *mp;
1591 	struct vnode *vp;
1592 	void *rl_cookie;
1593 	int error;
1594 
1595 	vp = fp->f_vnode;
1596 
1597 retry:
1598 	/*
1599 	 * Lock the whole range for truncation.  Otherwise split i/o
1600 	 * might happen partly before and partly after the truncation.
1601 	 */
1602 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1603 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1604 	if (error)
1605 		goto out1;
1606 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1607 	AUDIT_ARG_VNODE1(vp);
1608 	if (vp->v_type == VDIR) {
1609 		error = EISDIR;
1610 		goto out;
1611 	}
1612 #ifdef MAC
1613 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1614 	if (error)
1615 		goto out;
1616 #endif
1617 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1618 	    fp->f_cred);
1619 out:
1620 	VOP_UNLOCK(vp);
1621 	vn_finished_write(mp);
1622 out1:
1623 	vn_rangelock_unlock(vp, rl_cookie);
1624 	if (error == ERELOOKUP)
1625 		goto retry;
1626 	return (error);
1627 }
1628 
1629 /*
1630  * Truncate a file that is already locked.
1631  */
1632 int
1633 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1634     struct ucred *cred)
1635 {
1636 	struct vattr vattr;
1637 	int error;
1638 
1639 	error = VOP_ADD_WRITECOUNT(vp, 1);
1640 	if (error == 0) {
1641 		VATTR_NULL(&vattr);
1642 		vattr.va_size = length;
1643 		if (sync)
1644 			vattr.va_vaflags |= VA_SYNC;
1645 		error = VOP_SETATTR(vp, &vattr, cred);
1646 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1647 	}
1648 	return (error);
1649 }
1650 
1651 /*
1652  * File table vnode stat routine.
1653  */
1654 int
1655 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1656     struct thread *td)
1657 {
1658 	struct vnode *vp = fp->f_vnode;
1659 	int error;
1660 
1661 	vn_lock(vp, LK_SHARED | LK_RETRY);
1662 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1663 	VOP_UNLOCK(vp);
1664 
1665 	return (error);
1666 }
1667 
1668 /*
1669  * File table vnode ioctl routine.
1670  */
1671 static int
1672 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1673     struct thread *td)
1674 {
1675 	struct vattr vattr;
1676 	struct vnode *vp;
1677 	struct fiobmap2_arg *bmarg;
1678 	int error;
1679 
1680 	vp = fp->f_vnode;
1681 	switch (vp->v_type) {
1682 	case VDIR:
1683 	case VREG:
1684 		switch (com) {
1685 		case FIONREAD:
1686 			vn_lock(vp, LK_SHARED | LK_RETRY);
1687 			error = VOP_GETATTR(vp, &vattr, active_cred);
1688 			VOP_UNLOCK(vp);
1689 			if (error == 0)
1690 				*(int *)data = vattr.va_size - fp->f_offset;
1691 			return (error);
1692 		case FIOBMAP2:
1693 			bmarg = (struct fiobmap2_arg *)data;
1694 			vn_lock(vp, LK_SHARED | LK_RETRY);
1695 #ifdef MAC
1696 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1697 			    vp);
1698 			if (error == 0)
1699 #endif
1700 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1701 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1702 			VOP_UNLOCK(vp);
1703 			return (error);
1704 		case FIONBIO:
1705 		case FIOASYNC:
1706 			return (0);
1707 		default:
1708 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1709 			    active_cred, td));
1710 		}
1711 		break;
1712 	case VCHR:
1713 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1714 		    active_cred, td));
1715 	default:
1716 		return (ENOTTY);
1717 	}
1718 }
1719 
1720 /*
1721  * File table vnode poll routine.
1722  */
1723 static int
1724 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1725     struct thread *td)
1726 {
1727 	struct vnode *vp;
1728 	int error;
1729 
1730 	vp = fp->f_vnode;
1731 #if defined(MAC) || defined(AUDIT)
1732 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1733 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1734 		AUDIT_ARG_VNODE1(vp);
1735 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1736 		VOP_UNLOCK(vp);
1737 		if (error != 0)
1738 			return (error);
1739 	}
1740 #endif
1741 	error = VOP_POLL(vp, events, fp->f_cred, td);
1742 	return (error);
1743 }
1744 
1745 /*
1746  * Acquire the requested lock and then check for validity.  LK_RETRY
1747  * permits vn_lock to return doomed vnodes.
1748  */
1749 static int __noinline
1750 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1751     int error)
1752 {
1753 
1754 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1755 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1756 
1757 	if (error == 0)
1758 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1759 
1760 	if ((flags & LK_RETRY) == 0) {
1761 		if (error == 0) {
1762 			VOP_UNLOCK(vp);
1763 			error = ENOENT;
1764 		}
1765 		return (error);
1766 	}
1767 
1768 	/*
1769 	 * LK_RETRY case.
1770 	 *
1771 	 * Nothing to do if we got the lock.
1772 	 */
1773 	if (error == 0)
1774 		return (0);
1775 
1776 	/*
1777 	 * Interlock was dropped by the call in _vn_lock.
1778 	 */
1779 	flags &= ~LK_INTERLOCK;
1780 	do {
1781 		error = VOP_LOCK1(vp, flags, file, line);
1782 	} while (error != 0);
1783 	return (0);
1784 }
1785 
1786 int
1787 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1788 {
1789 	int error;
1790 
1791 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1792 	    ("vn_lock: no locktype (%d passed)", flags));
1793 	VNPASS(vp->v_holdcnt > 0, vp);
1794 	error = VOP_LOCK1(vp, flags, file, line);
1795 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1796 		return (_vn_lock_fallback(vp, flags, file, line, error));
1797 	return (0);
1798 }
1799 
1800 /*
1801  * File table vnode close routine.
1802  */
1803 static int
1804 vn_closefile(struct file *fp, struct thread *td)
1805 {
1806 	struct vnode *vp;
1807 	struct flock lf;
1808 	int error;
1809 	bool ref;
1810 
1811 	vp = fp->f_vnode;
1812 	fp->f_ops = &badfileops;
1813 	ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1814 
1815 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1816 
1817 	if (__predict_false(ref)) {
1818 		lf.l_whence = SEEK_SET;
1819 		lf.l_start = 0;
1820 		lf.l_len = 0;
1821 		lf.l_type = F_UNLCK;
1822 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1823 		vrele(vp);
1824 	}
1825 	return (error);
1826 }
1827 
1828 /*
1829  * Preparing to start a filesystem write operation. If the operation is
1830  * permitted, then we bump the count of operations in progress and
1831  * proceed. If a suspend request is in progress, we wait until the
1832  * suspension is over, and then proceed.
1833  */
1834 static int
1835 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1836 {
1837 	struct mount_pcpu *mpcpu;
1838 	int error, mflags;
1839 
1840 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1841 	    vfs_op_thread_enter(mp, mpcpu)) {
1842 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1843 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1844 		vfs_op_thread_exit(mp, mpcpu);
1845 		return (0);
1846 	}
1847 
1848 	if (mplocked)
1849 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1850 	else
1851 		MNT_ILOCK(mp);
1852 
1853 	error = 0;
1854 
1855 	/*
1856 	 * Check on status of suspension.
1857 	 */
1858 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1859 	    mp->mnt_susp_owner != curthread) {
1860 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1861 		    (flags & PCATCH) : 0) | (PUSER - 1);
1862 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1863 			if (flags & V_NOWAIT) {
1864 				error = EWOULDBLOCK;
1865 				goto unlock;
1866 			}
1867 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1868 			    "suspfs", 0);
1869 			if (error)
1870 				goto unlock;
1871 		}
1872 	}
1873 	if (flags & V_XSLEEP)
1874 		goto unlock;
1875 	mp->mnt_writeopcount++;
1876 unlock:
1877 	if (error != 0 || (flags & V_XSLEEP) != 0)
1878 		MNT_REL(mp);
1879 	MNT_IUNLOCK(mp);
1880 	return (error);
1881 }
1882 
1883 int
1884 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1885 {
1886 	struct mount *mp;
1887 	int error;
1888 
1889 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1890 	    ("V_MNTREF requires mp"));
1891 
1892 	error = 0;
1893 	/*
1894 	 * If a vnode is provided, get and return the mount point that
1895 	 * to which it will write.
1896 	 */
1897 	if (vp != NULL) {
1898 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1899 			*mpp = NULL;
1900 			if (error != EOPNOTSUPP)
1901 				return (error);
1902 			return (0);
1903 		}
1904 	}
1905 	if ((mp = *mpp) == NULL)
1906 		return (0);
1907 
1908 	/*
1909 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1910 	 * a vfs_ref().
1911 	 * As long as a vnode is not provided we need to acquire a
1912 	 * refcount for the provided mountpoint too, in order to
1913 	 * emulate a vfs_ref().
1914 	 */
1915 	if (vp == NULL && (flags & V_MNTREF) == 0)
1916 		vfs_ref(mp);
1917 
1918 	return (vn_start_write_refed(mp, flags, false));
1919 }
1920 
1921 /*
1922  * Secondary suspension. Used by operations such as vop_inactive
1923  * routines that are needed by the higher level functions. These
1924  * are allowed to proceed until all the higher level functions have
1925  * completed (indicated by mnt_writeopcount dropping to zero). At that
1926  * time, these operations are halted until the suspension is over.
1927  */
1928 int
1929 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1930 {
1931 	struct mount *mp;
1932 	int error;
1933 
1934 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1935 	    ("V_MNTREF requires mp"));
1936 
1937  retry:
1938 	if (vp != NULL) {
1939 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1940 			*mpp = NULL;
1941 			if (error != EOPNOTSUPP)
1942 				return (error);
1943 			return (0);
1944 		}
1945 	}
1946 	/*
1947 	 * If we are not suspended or have not yet reached suspended
1948 	 * mode, then let the operation proceed.
1949 	 */
1950 	if ((mp = *mpp) == NULL)
1951 		return (0);
1952 
1953 	/*
1954 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1955 	 * a vfs_ref().
1956 	 * As long as a vnode is not provided we need to acquire a
1957 	 * refcount for the provided mountpoint too, in order to
1958 	 * emulate a vfs_ref().
1959 	 */
1960 	MNT_ILOCK(mp);
1961 	if (vp == NULL && (flags & V_MNTREF) == 0)
1962 		MNT_REF(mp);
1963 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1964 		mp->mnt_secondary_writes++;
1965 		mp->mnt_secondary_accwrites++;
1966 		MNT_IUNLOCK(mp);
1967 		return (0);
1968 	}
1969 	if (flags & V_NOWAIT) {
1970 		MNT_REL(mp);
1971 		MNT_IUNLOCK(mp);
1972 		return (EWOULDBLOCK);
1973 	}
1974 	/*
1975 	 * Wait for the suspension to finish.
1976 	 */
1977 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1978 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1979 	    "suspfs", 0);
1980 	vfs_rel(mp);
1981 	if (error == 0)
1982 		goto retry;
1983 	return (error);
1984 }
1985 
1986 /*
1987  * Filesystem write operation has completed. If we are suspending and this
1988  * operation is the last one, notify the suspender that the suspension is
1989  * now in effect.
1990  */
1991 void
1992 vn_finished_write(struct mount *mp)
1993 {
1994 	struct mount_pcpu *mpcpu;
1995 	int c;
1996 
1997 	if (mp == NULL)
1998 		return;
1999 
2000 	if (vfs_op_thread_enter(mp, mpcpu)) {
2001 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2002 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2003 		vfs_op_thread_exit(mp, mpcpu);
2004 		return;
2005 	}
2006 
2007 	MNT_ILOCK(mp);
2008 	vfs_assert_mount_counters(mp);
2009 	MNT_REL(mp);
2010 	c = --mp->mnt_writeopcount;
2011 	if (mp->mnt_vfs_ops == 0) {
2012 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2013 		MNT_IUNLOCK(mp);
2014 		return;
2015 	}
2016 	if (c < 0)
2017 		vfs_dump_mount_counters(mp);
2018 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2019 		wakeup(&mp->mnt_writeopcount);
2020 	MNT_IUNLOCK(mp);
2021 }
2022 
2023 /*
2024  * Filesystem secondary write operation has completed. If we are
2025  * suspending and this operation is the last one, notify the suspender
2026  * that the suspension is now in effect.
2027  */
2028 void
2029 vn_finished_secondary_write(struct mount *mp)
2030 {
2031 	if (mp == NULL)
2032 		return;
2033 	MNT_ILOCK(mp);
2034 	MNT_REL(mp);
2035 	mp->mnt_secondary_writes--;
2036 	if (mp->mnt_secondary_writes < 0)
2037 		panic("vn_finished_secondary_write: neg cnt");
2038 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2039 	    mp->mnt_secondary_writes <= 0)
2040 		wakeup(&mp->mnt_secondary_writes);
2041 	MNT_IUNLOCK(mp);
2042 }
2043 
2044 /*
2045  * Request a filesystem to suspend write operations.
2046  */
2047 int
2048 vfs_write_suspend(struct mount *mp, int flags)
2049 {
2050 	int error;
2051 
2052 	vfs_op_enter(mp);
2053 
2054 	MNT_ILOCK(mp);
2055 	vfs_assert_mount_counters(mp);
2056 	if (mp->mnt_susp_owner == curthread) {
2057 		vfs_op_exit_locked(mp);
2058 		MNT_IUNLOCK(mp);
2059 		return (EALREADY);
2060 	}
2061 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2062 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2063 
2064 	/*
2065 	 * Unmount holds a write reference on the mount point.  If we
2066 	 * own busy reference and drain for writers, we deadlock with
2067 	 * the reference draining in the unmount path.  Callers of
2068 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2069 	 * vfs_busy() reference is owned and caller is not in the
2070 	 * unmount context.
2071 	 */
2072 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2073 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2074 		vfs_op_exit_locked(mp);
2075 		MNT_IUNLOCK(mp);
2076 		return (EBUSY);
2077 	}
2078 
2079 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2080 	mp->mnt_susp_owner = curthread;
2081 	if (mp->mnt_writeopcount > 0)
2082 		(void) msleep(&mp->mnt_writeopcount,
2083 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2084 	else
2085 		MNT_IUNLOCK(mp);
2086 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2087 		vfs_write_resume(mp, 0);
2088 		/* vfs_write_resume does vfs_op_exit() for us */
2089 	}
2090 	return (error);
2091 }
2092 
2093 /*
2094  * Request a filesystem to resume write operations.
2095  */
2096 void
2097 vfs_write_resume(struct mount *mp, int flags)
2098 {
2099 
2100 	MNT_ILOCK(mp);
2101 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2102 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2103 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2104 				       MNTK_SUSPENDED);
2105 		mp->mnt_susp_owner = NULL;
2106 		wakeup(&mp->mnt_writeopcount);
2107 		wakeup(&mp->mnt_flag);
2108 		curthread->td_pflags &= ~TDP_IGNSUSP;
2109 		if ((flags & VR_START_WRITE) != 0) {
2110 			MNT_REF(mp);
2111 			mp->mnt_writeopcount++;
2112 		}
2113 		MNT_IUNLOCK(mp);
2114 		if ((flags & VR_NO_SUSPCLR) == 0)
2115 			VFS_SUSP_CLEAN(mp);
2116 		vfs_op_exit(mp);
2117 	} else if ((flags & VR_START_WRITE) != 0) {
2118 		MNT_REF(mp);
2119 		vn_start_write_refed(mp, 0, true);
2120 	} else {
2121 		MNT_IUNLOCK(mp);
2122 	}
2123 }
2124 
2125 /*
2126  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2127  * methods.
2128  */
2129 int
2130 vfs_write_suspend_umnt(struct mount *mp)
2131 {
2132 	int error;
2133 
2134 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2135 	    ("vfs_write_suspend_umnt: recursed"));
2136 
2137 	/* dounmount() already called vn_start_write(). */
2138 	for (;;) {
2139 		vn_finished_write(mp);
2140 		error = vfs_write_suspend(mp, 0);
2141 		if (error != 0) {
2142 			vn_start_write(NULL, &mp, V_WAIT);
2143 			return (error);
2144 		}
2145 		MNT_ILOCK(mp);
2146 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2147 			break;
2148 		MNT_IUNLOCK(mp);
2149 		vn_start_write(NULL, &mp, V_WAIT);
2150 	}
2151 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2152 	wakeup(&mp->mnt_flag);
2153 	MNT_IUNLOCK(mp);
2154 	curthread->td_pflags |= TDP_IGNSUSP;
2155 	return (0);
2156 }
2157 
2158 /*
2159  * Implement kqueues for files by translating it to vnode operation.
2160  */
2161 static int
2162 vn_kqfilter(struct file *fp, struct knote *kn)
2163 {
2164 
2165 	return (VOP_KQFILTER(fp->f_vnode, kn));
2166 }
2167 
2168 int
2169 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2170 {
2171 	if ((fp->f_flag & FKQALLOWED) == 0)
2172 		return (EBADF);
2173 	return (vn_kqfilter(fp, kn));
2174 }
2175 
2176 /*
2177  * Simplified in-kernel wrapper calls for extended attribute access.
2178  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2179  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2180  */
2181 int
2182 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2183     const char *attrname, int *buflen, char *buf, struct thread *td)
2184 {
2185 	struct uio	auio;
2186 	struct iovec	iov;
2187 	int	error;
2188 
2189 	iov.iov_len = *buflen;
2190 	iov.iov_base = buf;
2191 
2192 	auio.uio_iov = &iov;
2193 	auio.uio_iovcnt = 1;
2194 	auio.uio_rw = UIO_READ;
2195 	auio.uio_segflg = UIO_SYSSPACE;
2196 	auio.uio_td = td;
2197 	auio.uio_offset = 0;
2198 	auio.uio_resid = *buflen;
2199 
2200 	if ((ioflg & IO_NODELOCKED) == 0)
2201 		vn_lock(vp, LK_SHARED | LK_RETRY);
2202 
2203 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2204 
2205 	/* authorize attribute retrieval as kernel */
2206 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2207 	    td);
2208 
2209 	if ((ioflg & IO_NODELOCKED) == 0)
2210 		VOP_UNLOCK(vp);
2211 
2212 	if (error == 0) {
2213 		*buflen = *buflen - auio.uio_resid;
2214 	}
2215 
2216 	return (error);
2217 }
2218 
2219 /*
2220  * XXX failure mode if partially written?
2221  */
2222 int
2223 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2224     const char *attrname, int buflen, char *buf, struct thread *td)
2225 {
2226 	struct uio	auio;
2227 	struct iovec	iov;
2228 	struct mount	*mp;
2229 	int	error;
2230 
2231 	iov.iov_len = buflen;
2232 	iov.iov_base = buf;
2233 
2234 	auio.uio_iov = &iov;
2235 	auio.uio_iovcnt = 1;
2236 	auio.uio_rw = UIO_WRITE;
2237 	auio.uio_segflg = UIO_SYSSPACE;
2238 	auio.uio_td = td;
2239 	auio.uio_offset = 0;
2240 	auio.uio_resid = buflen;
2241 
2242 	if ((ioflg & IO_NODELOCKED) == 0) {
2243 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2244 			return (error);
2245 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2246 	}
2247 
2248 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2249 
2250 	/* authorize attribute setting as kernel */
2251 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2252 
2253 	if ((ioflg & IO_NODELOCKED) == 0) {
2254 		vn_finished_write(mp);
2255 		VOP_UNLOCK(vp);
2256 	}
2257 
2258 	return (error);
2259 }
2260 
2261 int
2262 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2263     const char *attrname, struct thread *td)
2264 {
2265 	struct mount	*mp;
2266 	int	error;
2267 
2268 	if ((ioflg & IO_NODELOCKED) == 0) {
2269 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2270 			return (error);
2271 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2272 	}
2273 
2274 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2275 
2276 	/* authorize attribute removal as kernel */
2277 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2278 	if (error == EOPNOTSUPP)
2279 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2280 		    NULL, td);
2281 
2282 	if ((ioflg & IO_NODELOCKED) == 0) {
2283 		vn_finished_write(mp);
2284 		VOP_UNLOCK(vp);
2285 	}
2286 
2287 	return (error);
2288 }
2289 
2290 static int
2291 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2292     struct vnode **rvp)
2293 {
2294 
2295 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2296 }
2297 
2298 int
2299 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2300 {
2301 
2302 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2303 	    lkflags, rvp));
2304 }
2305 
2306 int
2307 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2308     int lkflags, struct vnode **rvp)
2309 {
2310 	struct mount *mp;
2311 	int ltype, error;
2312 
2313 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2314 	mp = vp->v_mount;
2315 	ltype = VOP_ISLOCKED(vp);
2316 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2317 	    ("vn_vget_ino: vp not locked"));
2318 	error = vfs_busy(mp, MBF_NOWAIT);
2319 	if (error != 0) {
2320 		vfs_ref(mp);
2321 		VOP_UNLOCK(vp);
2322 		error = vfs_busy(mp, 0);
2323 		vn_lock(vp, ltype | LK_RETRY);
2324 		vfs_rel(mp);
2325 		if (error != 0)
2326 			return (ENOENT);
2327 		if (VN_IS_DOOMED(vp)) {
2328 			vfs_unbusy(mp);
2329 			return (ENOENT);
2330 		}
2331 	}
2332 	VOP_UNLOCK(vp);
2333 	error = alloc(mp, alloc_arg, lkflags, rvp);
2334 	vfs_unbusy(mp);
2335 	if (error != 0 || *rvp != vp)
2336 		vn_lock(vp, ltype | LK_RETRY);
2337 	if (VN_IS_DOOMED(vp)) {
2338 		if (error == 0) {
2339 			if (*rvp == vp)
2340 				vunref(vp);
2341 			else
2342 				vput(*rvp);
2343 		}
2344 		error = ENOENT;
2345 	}
2346 	return (error);
2347 }
2348 
2349 int
2350 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2351     struct thread *td)
2352 {
2353 	off_t lim;
2354 	bool ktr_write;
2355 
2356 	if (td == NULL)
2357 		return (0);
2358 
2359 	/*
2360 	 * There are conditions where the limit is to be ignored.
2361 	 * However, since it is almost never reached, check it first.
2362 	 */
2363 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2364 	lim = lim_cur(td, RLIMIT_FSIZE);
2365 	if (__predict_false(ktr_write))
2366 		lim = td->td_ktr_io_lim;
2367 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2368 		return (0);
2369 
2370 	/*
2371 	 * The limit is reached.
2372 	 */
2373 	if (vp->v_type != VREG ||
2374 	    (td->td_pflags2 & TDP2_ACCT) != 0)
2375 		return (0);
2376 
2377 	if (!ktr_write || ktr_filesize_limit_signal) {
2378 		PROC_LOCK(td->td_proc);
2379 		kern_psignal(td->td_proc, SIGXFSZ);
2380 		PROC_UNLOCK(td->td_proc);
2381 	}
2382 	return (EFBIG);
2383 }
2384 
2385 int
2386 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2387     struct thread *td)
2388 {
2389 	struct vnode *vp;
2390 
2391 	vp = fp->f_vnode;
2392 #ifdef AUDIT
2393 	vn_lock(vp, LK_SHARED | LK_RETRY);
2394 	AUDIT_ARG_VNODE1(vp);
2395 	VOP_UNLOCK(vp);
2396 #endif
2397 	return (setfmode(td, active_cred, vp, mode));
2398 }
2399 
2400 int
2401 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2402     struct thread *td)
2403 {
2404 	struct vnode *vp;
2405 
2406 	vp = fp->f_vnode;
2407 #ifdef AUDIT
2408 	vn_lock(vp, LK_SHARED | LK_RETRY);
2409 	AUDIT_ARG_VNODE1(vp);
2410 	VOP_UNLOCK(vp);
2411 #endif
2412 	return (setfown(td, active_cred, vp, uid, gid));
2413 }
2414 
2415 void
2416 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2417 {
2418 	vm_object_t object;
2419 
2420 	if ((object = vp->v_object) == NULL)
2421 		return;
2422 	VM_OBJECT_WLOCK(object);
2423 	vm_object_page_remove(object, start, end, 0);
2424 	VM_OBJECT_WUNLOCK(object);
2425 }
2426 
2427 int
2428 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2429 {
2430 	struct vattr va;
2431 	daddr_t bn, bnp;
2432 	uint64_t bsize;
2433 	off_t noff;
2434 	int error;
2435 
2436 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2437 	    ("Wrong command %lu", cmd));
2438 
2439 	if (vn_lock(vp, LK_SHARED) != 0)
2440 		return (EBADF);
2441 	if (vp->v_type != VREG) {
2442 		error = ENOTTY;
2443 		goto unlock;
2444 	}
2445 	error = VOP_GETATTR(vp, &va, cred);
2446 	if (error != 0)
2447 		goto unlock;
2448 	noff = *off;
2449 	if (noff >= va.va_size) {
2450 		error = ENXIO;
2451 		goto unlock;
2452 	}
2453 	bsize = vp->v_mount->mnt_stat.f_iosize;
2454 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2455 	    noff % bsize) {
2456 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2457 		if (error == EOPNOTSUPP) {
2458 			error = ENOTTY;
2459 			goto unlock;
2460 		}
2461 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2462 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2463 			noff = bn * bsize;
2464 			if (noff < *off)
2465 				noff = *off;
2466 			goto unlock;
2467 		}
2468 	}
2469 	if (noff > va.va_size)
2470 		noff = va.va_size;
2471 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2472 	if (cmd == FIOSEEKDATA)
2473 		error = ENXIO;
2474 unlock:
2475 	VOP_UNLOCK(vp);
2476 	if (error == 0)
2477 		*off = noff;
2478 	return (error);
2479 }
2480 
2481 int
2482 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2483 {
2484 	struct ucred *cred;
2485 	struct vnode *vp;
2486 	struct vattr vattr;
2487 	off_t foffset, size;
2488 	int error, noneg;
2489 
2490 	cred = td->td_ucred;
2491 	vp = fp->f_vnode;
2492 	foffset = foffset_lock(fp, 0);
2493 	noneg = (vp->v_type != VCHR);
2494 	error = 0;
2495 	switch (whence) {
2496 	case L_INCR:
2497 		if (noneg &&
2498 		    (foffset < 0 ||
2499 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2500 			error = EOVERFLOW;
2501 			break;
2502 		}
2503 		offset += foffset;
2504 		break;
2505 	case L_XTND:
2506 		vn_lock(vp, LK_SHARED | LK_RETRY);
2507 		error = VOP_GETATTR(vp, &vattr, cred);
2508 		VOP_UNLOCK(vp);
2509 		if (error)
2510 			break;
2511 
2512 		/*
2513 		 * If the file references a disk device, then fetch
2514 		 * the media size and use that to determine the ending
2515 		 * offset.
2516 		 */
2517 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2518 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2519 			vattr.va_size = size;
2520 		if (noneg &&
2521 		    (vattr.va_size > OFF_MAX ||
2522 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2523 			error = EOVERFLOW;
2524 			break;
2525 		}
2526 		offset += vattr.va_size;
2527 		break;
2528 	case L_SET:
2529 		break;
2530 	case SEEK_DATA:
2531 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2532 		if (error == ENOTTY)
2533 			error = EINVAL;
2534 		break;
2535 	case SEEK_HOLE:
2536 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2537 		if (error == ENOTTY)
2538 			error = EINVAL;
2539 		break;
2540 	default:
2541 		error = EINVAL;
2542 	}
2543 	if (error == 0 && noneg && offset < 0)
2544 		error = EINVAL;
2545 	if (error != 0)
2546 		goto drop;
2547 	VFS_KNOTE_UNLOCKED(vp, 0);
2548 	td->td_uretoff.tdu_off = offset;
2549 drop:
2550 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2551 	return (error);
2552 }
2553 
2554 int
2555 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2556     struct thread *td)
2557 {
2558 	int error;
2559 
2560 	/*
2561 	 * Grant permission if the caller is the owner of the file, or
2562 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2563 	 * on the file.  If the time pointer is null, then write
2564 	 * permission on the file is also sufficient.
2565 	 *
2566 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2567 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2568 	 * will be allowed to set the times [..] to the current
2569 	 * server time.
2570 	 */
2571 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2572 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2573 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2574 	return (error);
2575 }
2576 
2577 int
2578 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2579 {
2580 	struct vnode *vp;
2581 	int error;
2582 
2583 	if (fp->f_type == DTYPE_FIFO)
2584 		kif->kf_type = KF_TYPE_FIFO;
2585 	else
2586 		kif->kf_type = KF_TYPE_VNODE;
2587 	vp = fp->f_vnode;
2588 	vref(vp);
2589 	FILEDESC_SUNLOCK(fdp);
2590 	error = vn_fill_kinfo_vnode(vp, kif);
2591 	vrele(vp);
2592 	FILEDESC_SLOCK(fdp);
2593 	return (error);
2594 }
2595 
2596 static inline void
2597 vn_fill_junk(struct kinfo_file *kif)
2598 {
2599 	size_t len, olen;
2600 
2601 	/*
2602 	 * Simulate vn_fullpath returning changing values for a given
2603 	 * vp during e.g. coredump.
2604 	 */
2605 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2606 	olen = strlen(kif->kf_path);
2607 	if (len < olen)
2608 		strcpy(&kif->kf_path[len - 1], "$");
2609 	else
2610 		for (; olen < len; olen++)
2611 			strcpy(&kif->kf_path[olen], "A");
2612 }
2613 
2614 int
2615 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2616 {
2617 	struct vattr va;
2618 	char *fullpath, *freepath;
2619 	int error;
2620 
2621 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2622 	freepath = NULL;
2623 	fullpath = "-";
2624 	error = vn_fullpath(vp, &fullpath, &freepath);
2625 	if (error == 0) {
2626 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2627 	}
2628 	if (freepath != NULL)
2629 		free(freepath, M_TEMP);
2630 
2631 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2632 		vn_fill_junk(kif);
2633 	);
2634 
2635 	/*
2636 	 * Retrieve vnode attributes.
2637 	 */
2638 	va.va_fsid = VNOVAL;
2639 	va.va_rdev = NODEV;
2640 	vn_lock(vp, LK_SHARED | LK_RETRY);
2641 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2642 	VOP_UNLOCK(vp);
2643 	if (error != 0)
2644 		return (error);
2645 	if (va.va_fsid != VNOVAL)
2646 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2647 	else
2648 		kif->kf_un.kf_file.kf_file_fsid =
2649 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2650 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2651 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2652 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2653 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2654 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2655 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2656 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2657 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2658 	return (0);
2659 }
2660 
2661 int
2662 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2663     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2664     struct thread *td)
2665 {
2666 #ifdef HWPMC_HOOKS
2667 	struct pmckern_map_in pkm;
2668 #endif
2669 	struct mount *mp;
2670 	struct vnode *vp;
2671 	vm_object_t object;
2672 	vm_prot_t maxprot;
2673 	boolean_t writecounted;
2674 	int error;
2675 
2676 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2677     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2678 	/*
2679 	 * POSIX shared-memory objects are defined to have
2680 	 * kernel persistence, and are not defined to support
2681 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2682 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2683 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2684 	 * flag to request this behavior.
2685 	 */
2686 	if ((fp->f_flag & FPOSIXSHM) != 0)
2687 		flags |= MAP_NOSYNC;
2688 #endif
2689 	vp = fp->f_vnode;
2690 
2691 	/*
2692 	 * Ensure that file and memory protections are
2693 	 * compatible.  Note that we only worry about
2694 	 * writability if mapping is shared; in this case,
2695 	 * current and max prot are dictated by the open file.
2696 	 * XXX use the vnode instead?  Problem is: what
2697 	 * credentials do we use for determination? What if
2698 	 * proc does a setuid?
2699 	 */
2700 	mp = vp->v_mount;
2701 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2702 		maxprot = VM_PROT_NONE;
2703 		if ((prot & VM_PROT_EXECUTE) != 0)
2704 			return (EACCES);
2705 	} else
2706 		maxprot = VM_PROT_EXECUTE;
2707 	if ((fp->f_flag & FREAD) != 0)
2708 		maxprot |= VM_PROT_READ;
2709 	else if ((prot & VM_PROT_READ) != 0)
2710 		return (EACCES);
2711 
2712 	/*
2713 	 * If we are sharing potential changes via MAP_SHARED and we
2714 	 * are trying to get write permission although we opened it
2715 	 * without asking for it, bail out.
2716 	 */
2717 	if ((flags & MAP_SHARED) != 0) {
2718 		if ((fp->f_flag & FWRITE) != 0)
2719 			maxprot |= VM_PROT_WRITE;
2720 		else if ((prot & VM_PROT_WRITE) != 0)
2721 			return (EACCES);
2722 	} else {
2723 		maxprot |= VM_PROT_WRITE;
2724 		cap_maxprot |= VM_PROT_WRITE;
2725 	}
2726 	maxprot &= cap_maxprot;
2727 
2728 	/*
2729 	 * For regular files and shared memory, POSIX requires that
2730 	 * the value of foff be a legitimate offset within the data
2731 	 * object.  In particular, negative offsets are invalid.
2732 	 * Blocking negative offsets and overflows here avoids
2733 	 * possible wraparound or user-level access into reserved
2734 	 * ranges of the data object later.  In contrast, POSIX does
2735 	 * not dictate how offsets are used by device drivers, so in
2736 	 * the case of a device mapping a negative offset is passed
2737 	 * on.
2738 	 */
2739 	if (
2740 #ifdef _LP64
2741 	    size > OFF_MAX ||
2742 #endif
2743 	    foff > OFF_MAX - size)
2744 		return (EINVAL);
2745 
2746 	writecounted = FALSE;
2747 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2748 	    &foff, &object, &writecounted);
2749 	if (error != 0)
2750 		return (error);
2751 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2752 	    foff, writecounted, td);
2753 	if (error != 0) {
2754 		/*
2755 		 * If this mapping was accounted for in the vnode's
2756 		 * writecount, then undo that now.
2757 		 */
2758 		if (writecounted)
2759 			vm_pager_release_writecount(object, 0, size);
2760 		vm_object_deallocate(object);
2761 	}
2762 #ifdef HWPMC_HOOKS
2763 	/* Inform hwpmc(4) if an executable is being mapped. */
2764 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2765 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2766 			pkm.pm_file = vp;
2767 			pkm.pm_address = (uintptr_t) *addr;
2768 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2769 		}
2770 	}
2771 #endif
2772 	return (error);
2773 }
2774 
2775 void
2776 vn_fsid(struct vnode *vp, struct vattr *va)
2777 {
2778 	fsid_t *f;
2779 
2780 	f = &vp->v_mount->mnt_stat.f_fsid;
2781 	va->va_fsid = (uint32_t)f->val[1];
2782 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2783 	va->va_fsid += (uint32_t)f->val[0];
2784 }
2785 
2786 int
2787 vn_fsync_buf(struct vnode *vp, int waitfor)
2788 {
2789 	struct buf *bp, *nbp;
2790 	struct bufobj *bo;
2791 	struct mount *mp;
2792 	int error, maxretry;
2793 
2794 	error = 0;
2795 	maxretry = 10000;     /* large, arbitrarily chosen */
2796 	mp = NULL;
2797 	if (vp->v_type == VCHR) {
2798 		VI_LOCK(vp);
2799 		mp = vp->v_rdev->si_mountpt;
2800 		VI_UNLOCK(vp);
2801 	}
2802 	bo = &vp->v_bufobj;
2803 	BO_LOCK(bo);
2804 loop1:
2805 	/*
2806 	 * MARK/SCAN initialization to avoid infinite loops.
2807 	 */
2808         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2809 		bp->b_vflags &= ~BV_SCANNED;
2810 		bp->b_error = 0;
2811 	}
2812 
2813 	/*
2814 	 * Flush all dirty buffers associated with a vnode.
2815 	 */
2816 loop2:
2817 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2818 		if ((bp->b_vflags & BV_SCANNED) != 0)
2819 			continue;
2820 		bp->b_vflags |= BV_SCANNED;
2821 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2822 			if (waitfor != MNT_WAIT)
2823 				continue;
2824 			if (BUF_LOCK(bp,
2825 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2826 			    BO_LOCKPTR(bo)) != 0) {
2827 				BO_LOCK(bo);
2828 				goto loop1;
2829 			}
2830 			BO_LOCK(bo);
2831 		}
2832 		BO_UNLOCK(bo);
2833 		KASSERT(bp->b_bufobj == bo,
2834 		    ("bp %p wrong b_bufobj %p should be %p",
2835 		    bp, bp->b_bufobj, bo));
2836 		if ((bp->b_flags & B_DELWRI) == 0)
2837 			panic("fsync: not dirty");
2838 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2839 			vfs_bio_awrite(bp);
2840 		} else {
2841 			bremfree(bp);
2842 			bawrite(bp);
2843 		}
2844 		if (maxretry < 1000)
2845 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2846 		BO_LOCK(bo);
2847 		goto loop2;
2848 	}
2849 
2850 	/*
2851 	 * If synchronous the caller expects us to completely resolve all
2852 	 * dirty buffers in the system.  Wait for in-progress I/O to
2853 	 * complete (which could include background bitmap writes), then
2854 	 * retry if dirty blocks still exist.
2855 	 */
2856 	if (waitfor == MNT_WAIT) {
2857 		bufobj_wwait(bo, 0, 0);
2858 		if (bo->bo_dirty.bv_cnt > 0) {
2859 			/*
2860 			 * If we are unable to write any of these buffers
2861 			 * then we fail now rather than trying endlessly
2862 			 * to write them out.
2863 			 */
2864 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2865 				if ((error = bp->b_error) != 0)
2866 					break;
2867 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2868 			    (error == 0 && --maxretry >= 0))
2869 				goto loop1;
2870 			if (error == 0)
2871 				error = EAGAIN;
2872 		}
2873 	}
2874 	BO_UNLOCK(bo);
2875 	if (error != 0)
2876 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2877 
2878 	return (error);
2879 }
2880 
2881 /*
2882  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2883  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2884  * to do the actual copy.
2885  * vn_generic_copy_file_range() is factored out, so it can be called
2886  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2887  * different file systems.
2888  */
2889 int
2890 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2891     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2892     struct ucred *outcred, struct thread *fsize_td)
2893 {
2894 	int error;
2895 	size_t len;
2896 	uint64_t uval;
2897 
2898 	len = *lenp;
2899 	*lenp = 0;		/* For error returns. */
2900 	error = 0;
2901 
2902 	/* Do some sanity checks on the arguments. */
2903 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2904 		error = EISDIR;
2905 	else if (*inoffp < 0 || *outoffp < 0 ||
2906 	    invp->v_type != VREG || outvp->v_type != VREG)
2907 		error = EINVAL;
2908 	if (error != 0)
2909 		goto out;
2910 
2911 	/* Ensure offset + len does not wrap around. */
2912 	uval = *inoffp;
2913 	uval += len;
2914 	if (uval > INT64_MAX)
2915 		len = INT64_MAX - *inoffp;
2916 	uval = *outoffp;
2917 	uval += len;
2918 	if (uval > INT64_MAX)
2919 		len = INT64_MAX - *outoffp;
2920 	if (len == 0)
2921 		goto out;
2922 
2923 	/*
2924 	 * If the two vnode are for the same file system, call
2925 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2926 	 * which can handle copies across multiple file systems.
2927 	 */
2928 	*lenp = len;
2929 	if (invp->v_mount == outvp->v_mount)
2930 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2931 		    lenp, flags, incred, outcred, fsize_td);
2932 	else
2933 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2934 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2935 out:
2936 	return (error);
2937 }
2938 
2939 /*
2940  * Test len bytes of data starting at dat for all bytes == 0.
2941  * Return true if all bytes are zero, false otherwise.
2942  * Expects dat to be well aligned.
2943  */
2944 static bool
2945 mem_iszero(void *dat, int len)
2946 {
2947 	int i;
2948 	const u_int *p;
2949 	const char *cp;
2950 
2951 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2952 		if (len >= sizeof(*p)) {
2953 			if (*p != 0)
2954 				return (false);
2955 		} else {
2956 			cp = (const char *)p;
2957 			for (i = 0; i < len; i++, cp++)
2958 				if (*cp != '\0')
2959 					return (false);
2960 		}
2961 	}
2962 	return (true);
2963 }
2964 
2965 /*
2966  * Look for a hole in the output file and, if found, adjust *outoffp
2967  * and *xferp to skip past the hole.
2968  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2969  * to be written as 0's upon return.
2970  */
2971 static off_t
2972 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2973     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2974 {
2975 	int error;
2976 	off_t delta;
2977 
2978 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2979 		*dataoffp = *outoffp;
2980 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2981 		    curthread);
2982 		if (error == 0) {
2983 			*holeoffp = *dataoffp;
2984 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2985 			    curthread);
2986 		}
2987 		if (error != 0 || *holeoffp == *dataoffp) {
2988 			/*
2989 			 * Since outvp is unlocked, it may be possible for
2990 			 * another thread to do a truncate(), lseek(), write()
2991 			 * creating a hole at startoff between the above
2992 			 * VOP_IOCTL() calls, if the other thread does not do
2993 			 * rangelocking.
2994 			 * If that happens, *holeoffp == *dataoffp and finding
2995 			 * the hole has failed, so disable vn_skip_hole().
2996 			 */
2997 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2998 			return (xfer2);
2999 		}
3000 		KASSERT(*dataoffp >= *outoffp,
3001 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3002 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3003 		KASSERT(*holeoffp > *dataoffp,
3004 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3005 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3006 	}
3007 
3008 	/*
3009 	 * If there is a hole before the data starts, advance *outoffp and
3010 	 * *xferp past the hole.
3011 	 */
3012 	if (*dataoffp > *outoffp) {
3013 		delta = *dataoffp - *outoffp;
3014 		if (delta >= *xferp) {
3015 			/* Entire *xferp is a hole. */
3016 			*outoffp += *xferp;
3017 			*xferp = 0;
3018 			return (0);
3019 		}
3020 		*xferp -= delta;
3021 		*outoffp += delta;
3022 		xfer2 = MIN(xfer2, *xferp);
3023 	}
3024 
3025 	/*
3026 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3027 	 * that the write ends at the start of the hole.
3028 	 * *holeoffp should always be greater than *outoffp, but for the
3029 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3030 	 * value.
3031 	 */
3032 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3033 		xfer2 = *holeoffp - *outoffp;
3034 	return (xfer2);
3035 }
3036 
3037 /*
3038  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3039  * dat is a maximum of blksize in length and can be written repeatedly in
3040  * the chunk.
3041  * If growfile == true, just grow the file via vn_truncate_locked() instead
3042  * of doing actual writes.
3043  * If checkhole == true, a hole is being punched, so skip over any hole
3044  * already in the output file.
3045  */
3046 static int
3047 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3048     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3049 {
3050 	struct mount *mp;
3051 	off_t dataoff, holeoff, xfer2;
3052 	int error;
3053 
3054 	/*
3055 	 * Loop around doing writes of blksize until write has been completed.
3056 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3057 	 * done for each iteration, since the xfer argument can be very
3058 	 * large if there is a large hole to punch in the output file.
3059 	 */
3060 	error = 0;
3061 	holeoff = 0;
3062 	do {
3063 		xfer2 = MIN(xfer, blksize);
3064 		if (checkhole) {
3065 			/*
3066 			 * Punching a hole.  Skip writing if there is
3067 			 * already a hole in the output file.
3068 			 */
3069 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3070 			    &dataoff, &holeoff, cred);
3071 			if (xfer == 0)
3072 				break;
3073 			if (holeoff < 0)
3074 				checkhole = false;
3075 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3076 			    (intmax_t)xfer2));
3077 		}
3078 		bwillwrite();
3079 		mp = NULL;
3080 		error = vn_start_write(outvp, &mp, V_WAIT);
3081 		if (error != 0)
3082 			break;
3083 		if (growfile) {
3084 			error = vn_lock(outvp, LK_EXCLUSIVE);
3085 			if (error == 0) {
3086 				error = vn_truncate_locked(outvp, outoff + xfer,
3087 				    false, cred);
3088 				VOP_UNLOCK(outvp);
3089 			}
3090 		} else {
3091 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3092 			if (error == 0) {
3093 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3094 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3095 				    curthread->td_ucred, cred, NULL, curthread);
3096 				outoff += xfer2;
3097 				xfer -= xfer2;
3098 				VOP_UNLOCK(outvp);
3099 			}
3100 		}
3101 		if (mp != NULL)
3102 			vn_finished_write(mp);
3103 	} while (!growfile && xfer > 0 && error == 0);
3104 	return (error);
3105 }
3106 
3107 /*
3108  * Copy a byte range of one file to another.  This function can handle the
3109  * case where invp and outvp are on different file systems.
3110  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3111  * is no better file system specific way to do it.
3112  */
3113 int
3114 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3115     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3116     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3117 {
3118 	struct vattr va, inva;
3119 	struct mount *mp;
3120 	struct uio io;
3121 	off_t startoff, endoff, xfer, xfer2;
3122 	u_long blksize;
3123 	int error, interrupted;
3124 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3125 	ssize_t aresid;
3126 	size_t copylen, len, rem, savlen;
3127 	char *dat;
3128 	long holein, holeout;
3129 
3130 	holein = holeout = 0;
3131 	savlen = len = *lenp;
3132 	error = 0;
3133 	interrupted = 0;
3134 	dat = NULL;
3135 
3136 	error = vn_lock(invp, LK_SHARED);
3137 	if (error != 0)
3138 		goto out;
3139 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3140 		holein = 0;
3141 	if (holein > 0)
3142 		error = VOP_GETATTR(invp, &inva, incred);
3143 	VOP_UNLOCK(invp);
3144 	if (error != 0)
3145 		goto out;
3146 
3147 	mp = NULL;
3148 	error = vn_start_write(outvp, &mp, V_WAIT);
3149 	if (error == 0)
3150 		error = vn_lock(outvp, LK_EXCLUSIVE);
3151 	if (error == 0) {
3152 		/*
3153 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3154 		 * now that outvp is locked.
3155 		 */
3156 		if (fsize_td != NULL) {
3157 			io.uio_offset = *outoffp;
3158 			io.uio_resid = len;
3159 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3160 			if (error != 0)
3161 				error = EFBIG;
3162 		}
3163 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3164 			holeout = 0;
3165 		/*
3166 		 * Holes that are past EOF do not need to be written as a block
3167 		 * of zero bytes.  So, truncate the output file as far as
3168 		 * possible and then use va.va_size to decide if writing 0
3169 		 * bytes is necessary in the loop below.
3170 		 */
3171 		if (error == 0)
3172 			error = VOP_GETATTR(outvp, &va, outcred);
3173 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3174 		    *outoffp + len) {
3175 #ifdef MAC
3176 			error = mac_vnode_check_write(curthread->td_ucred,
3177 			    outcred, outvp);
3178 			if (error == 0)
3179 #endif
3180 				error = vn_truncate_locked(outvp, *outoffp,
3181 				    false, outcred);
3182 			if (error == 0)
3183 				va.va_size = *outoffp;
3184 		}
3185 		VOP_UNLOCK(outvp);
3186 	}
3187 	if (mp != NULL)
3188 		vn_finished_write(mp);
3189 	if (error != 0)
3190 		goto out;
3191 
3192 	/*
3193 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3194 	 * If hole sizes aren't available, set the blksize to the larger
3195 	 * f_iosize of invp and outvp.
3196 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3197 	 * This value is clipped at 4Kbytes and 1Mbyte.
3198 	 */
3199 	blksize = MAX(holein, holeout);
3200 
3201 	/* Clip len to end at an exact multiple of hole size. */
3202 	if (blksize > 1) {
3203 		rem = *inoffp % blksize;
3204 		if (rem > 0)
3205 			rem = blksize - rem;
3206 		if (len > rem && len - rem > blksize)
3207 			len = savlen = rounddown(len - rem, blksize) + rem;
3208 	}
3209 
3210 	if (blksize <= 1)
3211 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3212 		    outvp->v_mount->mnt_stat.f_iosize);
3213 	if (blksize < 4096)
3214 		blksize = 4096;
3215 	else if (blksize > 1024 * 1024)
3216 		blksize = 1024 * 1024;
3217 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3218 
3219 	/*
3220 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3221 	 * to find holes.  Otherwise, just scan the read block for all 0s
3222 	 * in the inner loop where the data copying is done.
3223 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3224 	 * support holes on the server, but do not support FIOSEEKHOLE.
3225 	 */
3226 	holetoeof = eof = false;
3227 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3228 		endoff = 0;			/* To shut up compilers. */
3229 		cantseek = true;
3230 		startoff = *inoffp;
3231 		copylen = len;
3232 
3233 		/*
3234 		 * Find the next data area.  If there is just a hole to EOF,
3235 		 * FIOSEEKDATA should fail with ENXIO.
3236 		 * (I do not know if any file system will report a hole to
3237 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3238 		 *  will fail for those file systems.)
3239 		 *
3240 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3241 		 * the code just falls through to the inner copy loop.
3242 		 */
3243 		error = EINVAL;
3244 		if (holein > 0) {
3245 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3246 			    incred, curthread);
3247 			if (error == ENXIO) {
3248 				startoff = endoff = inva.va_size;
3249 				eof = holetoeof = true;
3250 				error = 0;
3251 			}
3252 		}
3253 		if (error == 0 && !holetoeof) {
3254 			endoff = startoff;
3255 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3256 			    incred, curthread);
3257 			/*
3258 			 * Since invp is unlocked, it may be possible for
3259 			 * another thread to do a truncate(), lseek(), write()
3260 			 * creating a hole at startoff between the above
3261 			 * VOP_IOCTL() calls, if the other thread does not do
3262 			 * rangelocking.
3263 			 * If that happens, startoff == endoff and finding
3264 			 * the hole has failed, so set an error.
3265 			 */
3266 			if (error == 0 && startoff == endoff)
3267 				error = EINVAL; /* Any error. Reset to 0. */
3268 		}
3269 		if (error == 0) {
3270 			if (startoff > *inoffp) {
3271 				/* Found hole before data block. */
3272 				xfer = MIN(startoff - *inoffp, len);
3273 				if (*outoffp < va.va_size) {
3274 					/* Must write 0s to punch hole. */
3275 					xfer2 = MIN(va.va_size - *outoffp,
3276 					    xfer);
3277 					memset(dat, 0, MIN(xfer2, blksize));
3278 					error = vn_write_outvp(outvp, dat,
3279 					    *outoffp, xfer2, blksize, false,
3280 					    holeout > 0, outcred);
3281 				}
3282 
3283 				if (error == 0 && *outoffp + xfer >
3284 				    va.va_size && (xfer == len || holetoeof)) {
3285 					/* Grow output file (hole at end). */
3286 					error = vn_write_outvp(outvp, dat,
3287 					    *outoffp, xfer, blksize, true,
3288 					    false, outcred);
3289 				}
3290 				if (error == 0) {
3291 					*inoffp += xfer;
3292 					*outoffp += xfer;
3293 					len -= xfer;
3294 					if (len < savlen)
3295 						interrupted = sig_intr();
3296 				}
3297 			}
3298 			copylen = MIN(len, endoff - startoff);
3299 			cantseek = false;
3300 		} else {
3301 			cantseek = true;
3302 			startoff = *inoffp;
3303 			copylen = len;
3304 			error = 0;
3305 		}
3306 
3307 		xfer = blksize;
3308 		if (cantseek) {
3309 			/*
3310 			 * Set first xfer to end at a block boundary, so that
3311 			 * holes are more likely detected in the loop below via
3312 			 * the for all bytes 0 method.
3313 			 */
3314 			xfer -= (*inoffp % blksize);
3315 		}
3316 		/* Loop copying the data block. */
3317 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3318 			if (copylen < xfer)
3319 				xfer = copylen;
3320 			error = vn_lock(invp, LK_SHARED);
3321 			if (error != 0)
3322 				goto out;
3323 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3324 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3325 			    curthread->td_ucred, incred, &aresid,
3326 			    curthread);
3327 			VOP_UNLOCK(invp);
3328 			lastblock = false;
3329 			if (error == 0 && aresid > 0) {
3330 				/* Stop the copy at EOF on the input file. */
3331 				xfer -= aresid;
3332 				eof = true;
3333 				lastblock = true;
3334 			}
3335 			if (error == 0) {
3336 				/*
3337 				 * Skip the write for holes past the initial EOF
3338 				 * of the output file, unless this is the last
3339 				 * write of the output file at EOF.
3340 				 */
3341 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3342 				    false;
3343 				if (xfer == len)
3344 					lastblock = true;
3345 				if (!cantseek || *outoffp < va.va_size ||
3346 				    lastblock || !readzeros)
3347 					error = vn_write_outvp(outvp, dat,
3348 					    *outoffp, xfer, blksize,
3349 					    readzeros && lastblock &&
3350 					    *outoffp >= va.va_size, false,
3351 					    outcred);
3352 				if (error == 0) {
3353 					*inoffp += xfer;
3354 					startoff += xfer;
3355 					*outoffp += xfer;
3356 					copylen -= xfer;
3357 					len -= xfer;
3358 					if (len < savlen)
3359 						interrupted = sig_intr();
3360 				}
3361 			}
3362 			xfer = blksize;
3363 		}
3364 	}
3365 out:
3366 	*lenp = savlen - len;
3367 	free(dat, M_TEMP);
3368 	return (error);
3369 }
3370 
3371 static int
3372 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3373 {
3374 	struct mount *mp;
3375 	struct vnode *vp;
3376 	off_t olen, ooffset;
3377 	int error;
3378 #ifdef AUDIT
3379 	int audited_vnode1 = 0;
3380 #endif
3381 
3382 	vp = fp->f_vnode;
3383 	if (vp->v_type != VREG)
3384 		return (ENODEV);
3385 
3386 	/* Allocating blocks may take a long time, so iterate. */
3387 	for (;;) {
3388 		olen = len;
3389 		ooffset = offset;
3390 
3391 		bwillwrite();
3392 		mp = NULL;
3393 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3394 		if (error != 0)
3395 			break;
3396 		error = vn_lock(vp, LK_EXCLUSIVE);
3397 		if (error != 0) {
3398 			vn_finished_write(mp);
3399 			break;
3400 		}
3401 #ifdef AUDIT
3402 		if (!audited_vnode1) {
3403 			AUDIT_ARG_VNODE1(vp);
3404 			audited_vnode1 = 1;
3405 		}
3406 #endif
3407 #ifdef MAC
3408 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3409 		if (error == 0)
3410 #endif
3411 			error = VOP_ALLOCATE(vp, &offset, &len);
3412 		VOP_UNLOCK(vp);
3413 		vn_finished_write(mp);
3414 
3415 		if (olen + ooffset != offset + len) {
3416 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3417 			    ooffset, olen, offset, len);
3418 		}
3419 		if (error != 0 || len == 0)
3420 			break;
3421 		KASSERT(olen > len, ("Iteration did not make progress?"));
3422 		maybe_yield();
3423 	}
3424 
3425 	return (error);
3426 }
3427 
3428 static u_long vn_lock_pair_pause_cnt;
3429 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3430     &vn_lock_pair_pause_cnt, 0,
3431     "Count of vn_lock_pair deadlocks");
3432 
3433 u_int vn_lock_pair_pause_max;
3434 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3435     &vn_lock_pair_pause_max, 0,
3436     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3437 
3438 static void
3439 vn_lock_pair_pause(const char *wmesg)
3440 {
3441 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3442 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3443 }
3444 
3445 /*
3446  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3447  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3448  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3449  * can be NULL.
3450  *
3451  * The function returns with both vnodes exclusively locked, and
3452  * guarantees that it does not create lock order reversal with other
3453  * threads during its execution.  Both vnodes could be unlocked
3454  * temporary (and reclaimed).
3455  */
3456 void
3457 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3458     bool vp2_locked)
3459 {
3460 	int error;
3461 
3462 	if (vp1 == NULL && vp2 == NULL)
3463 		return;
3464 	if (vp1 != NULL) {
3465 		if (vp1_locked)
3466 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3467 		else
3468 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3469 	} else {
3470 		vp1_locked = true;
3471 	}
3472 	if (vp2 != NULL) {
3473 		if (vp2_locked)
3474 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3475 		else
3476 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3477 	} else {
3478 		vp2_locked = true;
3479 	}
3480 	if (!vp1_locked && !vp2_locked) {
3481 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3482 		vp1_locked = true;
3483 	}
3484 
3485 	for (;;) {
3486 		if (vp1_locked && vp2_locked)
3487 			break;
3488 		if (vp1_locked && vp2 != NULL) {
3489 			if (vp1 != NULL) {
3490 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3491 				    __FILE__, __LINE__);
3492 				if (error == 0)
3493 					break;
3494 				VOP_UNLOCK(vp1);
3495 				vp1_locked = false;
3496 				vn_lock_pair_pause("vlp1");
3497 			}
3498 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3499 			vp2_locked = true;
3500 		}
3501 		if (vp2_locked && vp1 != NULL) {
3502 			if (vp2 != NULL) {
3503 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3504 				    __FILE__, __LINE__);
3505 				if (error == 0)
3506 					break;
3507 				VOP_UNLOCK(vp2);
3508 				vp2_locked = false;
3509 				vn_lock_pair_pause("vlp2");
3510 			}
3511 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3512 			vp1_locked = true;
3513 		}
3514 	}
3515 	if (vp1 != NULL)
3516 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3517 	if (vp2 != NULL)
3518 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3519 }
3520 
3521 int
3522 vn_lktype_write(struct mount *mp, struct vnode *vp)
3523 {
3524 	if (MNT_SHARED_WRITES(mp) ||
3525 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3526 		return (LK_SHARED);
3527 	return (LK_EXCLUSIVE);
3528 }
3529