xref: /freebsd/sys/kern/vfs_vnops.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
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 
43 #include <sys/cdefs.h>
44 #include "opt_hwpmc_hooks.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/disk.h>
49 #include <sys/fail.h>
50 #include <sys/fcntl.h>
51 #include <sys/file.h>
52 #include <sys/kdb.h>
53 #include <sys/ktr.h>
54 #include <sys/stat.h>
55 #include <sys/priv.h>
56 #include <sys/proc.h>
57 #include <sys/limits.h>
58 #include <sys/lock.h>
59 #include <sys/mman.h>
60 #include <sys/mount.h>
61 #include <sys/mutex.h>
62 #include <sys/namei.h>
63 #include <sys/vnode.h>
64 #include <sys/dirent.h>
65 #include <sys/bio.h>
66 #include <sys/buf.h>
67 #include <sys/filio.h>
68 #include <sys/resourcevar.h>
69 #include <sys/rwlock.h>
70 #include <sys/prng.h>
71 #include <sys/sx.h>
72 #include <sys/sleepqueue.h>
73 #include <sys/sysctl.h>
74 #include <sys/ttycom.h>
75 #include <sys/conf.h>
76 #include <sys/syslog.h>
77 #include <sys/unistd.h>
78 #include <sys/user.h>
79 #include <sys/ktrace.h>
80 
81 #include <security/audit/audit.h>
82 #include <security/mac/mac_framework.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_extern.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vnode_pager.h>
92 
93 #ifdef HWPMC_HOOKS
94 #include <sys/pmckern.h>
95 #endif
96 
97 static fo_rdwr_t	vn_read;
98 static fo_rdwr_t	vn_write;
99 static fo_rdwr_t	vn_io_fault;
100 static fo_truncate_t	vn_truncate;
101 static fo_ioctl_t	vn_ioctl;
102 static fo_poll_t	vn_poll;
103 static fo_kqfilter_t	vn_kqfilter;
104 static fo_close_t	vn_closefile;
105 static fo_mmap_t	vn_mmap;
106 static fo_fallocate_t	vn_fallocate;
107 static fo_fspacectl_t	vn_fspacectl;
108 
109 struct 	fileops vnops = {
110 	.fo_read = vn_io_fault,
111 	.fo_write = vn_io_fault,
112 	.fo_truncate = vn_truncate,
113 	.fo_ioctl = vn_ioctl,
114 	.fo_poll = vn_poll,
115 	.fo_kqfilter = vn_kqfilter,
116 	.fo_stat = vn_statfile,
117 	.fo_close = vn_closefile,
118 	.fo_chmod = vn_chmod,
119 	.fo_chown = vn_chown,
120 	.fo_sendfile = vn_sendfile,
121 	.fo_seek = vn_seek,
122 	.fo_fill_kinfo = vn_fill_kinfo,
123 	.fo_mmap = vn_mmap,
124 	.fo_fallocate = vn_fallocate,
125 	.fo_fspacectl = vn_fspacectl,
126 	.fo_cmp = vn_cmp,
127 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
128 };
129 
130 const u_int io_hold_cnt = 16;
131 static int vn_io_fault_enable = 1;
132 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
133     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
134 static int vn_io_fault_prefault = 0;
135 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
136     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
137 static int vn_io_pgcache_read_enable = 1;
138 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
139     &vn_io_pgcache_read_enable, 0,
140     "Enable copying from page cache for reads, avoiding fs");
141 static u_long vn_io_faults_cnt;
142 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
143     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
144 
145 static int vfs_allow_read_dir = 0;
146 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
147     &vfs_allow_read_dir, 0,
148     "Enable read(2) of directory by root for filesystems that support it");
149 
150 /*
151  * Returns true if vn_io_fault mode of handling the i/o request should
152  * be used.
153  */
154 static bool
155 do_vn_io_fault(struct vnode *vp, struct uio *uio)
156 {
157 	struct mount *mp;
158 
159 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
160 	    (mp = vp->v_mount) != NULL &&
161 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
162 }
163 
164 /*
165  * Structure used to pass arguments to vn_io_fault1(), to do either
166  * file- or vnode-based I/O calls.
167  */
168 struct vn_io_fault_args {
169 	enum {
170 		VN_IO_FAULT_FOP,
171 		VN_IO_FAULT_VOP
172 	} kind;
173 	struct ucred *cred;
174 	int flags;
175 	union {
176 		struct fop_args_tag {
177 			struct file *fp;
178 			fo_rdwr_t *doio;
179 		} fop_args;
180 		struct vop_args_tag {
181 			struct vnode *vp;
182 		} vop_args;
183 	} args;
184 };
185 
186 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
187     struct vn_io_fault_args *args, struct thread *td);
188 
189 int
190 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
191 {
192 	struct thread *td = curthread;
193 
194 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
195 }
196 
197 static uint64_t
198 open2nameif(int fmode, u_int vn_open_flags)
199 {
200 	uint64_t res;
201 
202 	res = ISOPEN | LOCKLEAF;
203 	if ((fmode & O_RESOLVE_BENEATH) != 0)
204 		res |= RBENEATH;
205 	if ((fmode & O_EMPTY_PATH) != 0)
206 		res |= EMPTYPATH;
207 	if ((fmode & FREAD) != 0)
208 		res |= OPENREAD;
209 	if ((fmode & FWRITE) != 0)
210 		res |= OPENWRITE;
211 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
212 		res |= AUDITVNODE1;
213 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
214 		res |= NOCAPCHECK;
215 	if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
216 		res |= WANTIOCTLCAPS;
217 	return (res);
218 }
219 
220 /*
221  * Common code for vnode open operations via a name lookup.
222  * Lookup the vnode and invoke VOP_CREATE if needed.
223  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
224  *
225  * Note that this does NOT free nameidata for the successful case,
226  * due to the NDINIT being done elsewhere.
227  */
228 int
229 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
230     struct ucred *cred, struct file *fp)
231 {
232 	struct vnode *vp;
233 	struct mount *mp;
234 	struct vattr vat;
235 	struct vattr *vap = &vat;
236 	int fmode, error;
237 	bool first_open;
238 
239 restart:
240 	first_open = false;
241 	fmode = *flagp;
242 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
243 	    O_EXCL | O_DIRECTORY) ||
244 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
245 		return (EINVAL);
246 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
247 		ndp->ni_cnd.cn_nameiop = CREATE;
248 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
249 		/*
250 		 * Set NOCACHE to avoid flushing the cache when
251 		 * rolling in many files at once.
252 		 *
253 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
254 		 * exist despite NOCACHE.
255 		 */
256 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
257 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
258 			ndp->ni_cnd.cn_flags |= FOLLOW;
259 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
260 			bwillwrite();
261 		if ((error = namei(ndp)) != 0)
262 			return (error);
263 		if (ndp->ni_vp == NULL) {
264 			VATTR_NULL(vap);
265 			vap->va_type = VREG;
266 			vap->va_mode = cmode;
267 			if (fmode & O_EXCL)
268 				vap->va_vaflags |= VA_EXCLUSIVE;
269 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
270 				NDFREE_PNBUF(ndp);
271 				vput(ndp->ni_dvp);
272 				if ((error = vn_start_write(NULL, &mp,
273 				    V_XSLEEP | V_PCATCH)) != 0)
274 					return (error);
275 				NDREINIT(ndp);
276 				goto restart;
277 			}
278 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
279 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
280 #ifdef MAC
281 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
282 			    &ndp->ni_cnd, vap);
283 			if (error == 0)
284 #endif
285 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
286 				    &ndp->ni_cnd, vap);
287 			vp = ndp->ni_vp;
288 			if (error == 0 && (fmode & O_EXCL) != 0 &&
289 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
290 				VI_LOCK(vp);
291 				vp->v_iflag |= VI_FOPENING;
292 				VI_UNLOCK(vp);
293 				first_open = true;
294 			}
295 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
296 			    false);
297 			vn_finished_write(mp);
298 			if (error) {
299 				NDFREE_PNBUF(ndp);
300 				if (error == ERELOOKUP) {
301 					NDREINIT(ndp);
302 					goto restart;
303 				}
304 				return (error);
305 			}
306 			fmode &= ~O_TRUNC;
307 		} else {
308 			if (ndp->ni_dvp == ndp->ni_vp)
309 				vrele(ndp->ni_dvp);
310 			else
311 				vput(ndp->ni_dvp);
312 			ndp->ni_dvp = NULL;
313 			vp = ndp->ni_vp;
314 			if (fmode & O_EXCL) {
315 				error = EEXIST;
316 				goto bad;
317 			}
318 			if (vp->v_type == VDIR) {
319 				error = EISDIR;
320 				goto bad;
321 			}
322 			fmode &= ~O_CREAT;
323 		}
324 	} else {
325 		ndp->ni_cnd.cn_nameiop = LOOKUP;
326 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
327 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
328 		    FOLLOW;
329 		if ((fmode & FWRITE) == 0)
330 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
331 		if ((error = namei(ndp)) != 0)
332 			return (error);
333 		vp = ndp->ni_vp;
334 	}
335 	error = vn_open_vnode(vp, fmode, cred, curthread, fp);
336 	if (first_open) {
337 		VI_LOCK(vp);
338 		vp->v_iflag &= ~VI_FOPENING;
339 		wakeup(vp);
340 		VI_UNLOCK(vp);
341 	}
342 	if (error)
343 		goto bad;
344 	*flagp = fmode;
345 	return (0);
346 bad:
347 	NDFREE_PNBUF(ndp);
348 	vput(vp);
349 	*flagp = fmode;
350 	ndp->ni_vp = NULL;
351 	return (error);
352 }
353 
354 static int
355 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
356 {
357 	struct flock lf;
358 	int error, lock_flags, type;
359 
360 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
361 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
362 		return (0);
363 	KASSERT(fp != NULL, ("open with flock requires fp"));
364 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
365 		return (EOPNOTSUPP);
366 
367 	lock_flags = VOP_ISLOCKED(vp);
368 	VOP_UNLOCK(vp);
369 
370 	lf.l_whence = SEEK_SET;
371 	lf.l_start = 0;
372 	lf.l_len = 0;
373 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
374 	type = F_FLOCK;
375 	if ((fmode & FNONBLOCK) == 0)
376 		type |= F_WAIT;
377 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
378 		type |= F_FIRSTOPEN;
379 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
380 	if (error == 0)
381 		fp->f_flag |= FHASLOCK;
382 
383 	vn_lock(vp, lock_flags | LK_RETRY);
384 	return (error);
385 }
386 
387 /*
388  * Common code for vnode open operations once a vnode is located.
389  * Check permissions, and call the VOP_OPEN routine.
390  */
391 int
392 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
393     struct thread *td, struct file *fp)
394 {
395 	accmode_t accmode;
396 	int error;
397 
398 	if (vp->v_type == VLNK) {
399 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
400 			return (EMLINK);
401 	}
402 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
403 		return (ENOTDIR);
404 
405 	accmode = 0;
406 	if ((fmode & O_PATH) == 0) {
407 		if (vp->v_type == VSOCK)
408 			return (EOPNOTSUPP);
409 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
410 			if (vp->v_type == VDIR)
411 				return (EISDIR);
412 			accmode |= VWRITE;
413 		}
414 		if ((fmode & FREAD) != 0)
415 			accmode |= VREAD;
416 		if ((fmode & O_APPEND) && (fmode & FWRITE))
417 			accmode |= VAPPEND;
418 #ifdef MAC
419 		if ((fmode & O_CREAT) != 0)
420 			accmode |= VCREAT;
421 #endif
422 	}
423 	if ((fmode & FEXEC) != 0)
424 		accmode |= VEXEC;
425 #ifdef MAC
426 	if ((fmode & O_VERIFY) != 0)
427 		accmode |= VVERIFY;
428 	error = mac_vnode_check_open(cred, vp, accmode);
429 	if (error != 0)
430 		return (error);
431 
432 	accmode &= ~(VCREAT | VVERIFY);
433 #endif
434 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
435 		error = VOP_ACCESS(vp, accmode, cred, td);
436 		if (error != 0)
437 			return (error);
438 	}
439 	if ((fmode & O_PATH) != 0) {
440 		if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
441 		    VOP_ACCESS(vp, VREAD, cred, td) == 0)
442 			fp->f_flag |= FKQALLOWED;
443 		return (0);
444 	}
445 
446 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
447 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
448 	error = VOP_OPEN(vp, fmode, cred, td, fp);
449 	if (error != 0)
450 		return (error);
451 
452 	error = vn_open_vnode_advlock(vp, fmode, fp);
453 	if (error == 0 && (fmode & FWRITE) != 0) {
454 		error = VOP_ADD_WRITECOUNT(vp, 1);
455 		if (error == 0) {
456 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
457 			     __func__, vp, vp->v_writecount);
458 		}
459 	}
460 
461 	/*
462 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
463 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
464 	 */
465 	if (error != 0) {
466 		if (fp != NULL) {
467 			/*
468 			 * Arrange the call by having fdrop() to use
469 			 * vn_closefile().  This is to satisfy
470 			 * filesystems like devfs or tmpfs, which
471 			 * override fo_close().
472 			 */
473 			fp->f_flag |= FOPENFAILED;
474 			fp->f_vnode = vp;
475 			if (fp->f_ops == &badfileops) {
476 				fp->f_type = DTYPE_VNODE;
477 				fp->f_ops = &vnops;
478 			}
479 			vref(vp);
480 		} else {
481 			/*
482 			 * If there is no fp, due to kernel-mode open,
483 			 * we can call VOP_CLOSE() now.
484 			 */
485 			if ((vp->v_type == VFIFO ||
486 			    !MNT_EXTENDED_SHARED(vp->v_mount)) &&
487 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
488 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
489 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
490 			    cred, td);
491 		}
492 	}
493 
494 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
495 	return (error);
496 
497 }
498 
499 /*
500  * Check for write permissions on the specified vnode.
501  * Prototype text segments cannot be written.
502  * It is racy.
503  */
504 int
505 vn_writechk(struct vnode *vp)
506 {
507 
508 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
509 	/*
510 	 * If there's shared text associated with
511 	 * the vnode, try to free it up once.  If
512 	 * we fail, we can't allow writing.
513 	 */
514 	if (VOP_IS_TEXT(vp))
515 		return (ETXTBSY);
516 
517 	return (0);
518 }
519 
520 /*
521  * Vnode close call
522  */
523 static int
524 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
525     struct thread *td, bool keep_ref)
526 {
527 	struct mount *mp;
528 	int error, lock_flags;
529 
530 	lock_flags = vp->v_type != VFIFO && MNT_EXTENDED_SHARED(vp->v_mount) ?
531 	    LK_SHARED : 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 | V_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 
832 static off_t
833 foffset_read(struct file *fp)
834 {
835 
836 	return (atomic_load_long(&fp->f_offset));
837 }
838 #else
839 off_t
840 foffset_lock(struct file *fp, int flags)
841 {
842 	struct mtx *mtxp;
843 	off_t res;
844 
845 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
846 
847 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
848 	mtx_lock(mtxp);
849 	if ((flags & FOF_NOLOCK) == 0) {
850 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
851 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
852 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
853 			    "vofflock", 0);
854 		}
855 		fp->f_vnread_flags |= FOFFSET_LOCKED;
856 	}
857 	res = fp->f_offset;
858 	mtx_unlock(mtxp);
859 	return (res);
860 }
861 
862 void
863 foffset_unlock(struct file *fp, off_t val, int flags)
864 {
865 	struct mtx *mtxp;
866 
867 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
868 
869 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
870 	mtx_lock(mtxp);
871 	if ((flags & FOF_NOUPDATE) == 0)
872 		fp->f_offset = val;
873 	if ((flags & FOF_NEXTOFF_R) != 0)
874 		fp->f_nextoff[UIO_READ] = val;
875 	if ((flags & FOF_NEXTOFF_W) != 0)
876 		fp->f_nextoff[UIO_WRITE] = val;
877 	if ((flags & FOF_NOLOCK) == 0) {
878 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
879 		    ("Lost FOFFSET_LOCKED"));
880 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
881 			wakeup(&fp->f_vnread_flags);
882 		fp->f_vnread_flags = 0;
883 	}
884 	mtx_unlock(mtxp);
885 }
886 
887 static off_t
888 foffset_read(struct file *fp)
889 {
890 
891 	return (foffset_lock(fp, FOF_NOLOCK));
892 }
893 #endif
894 
895 void
896 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
897 {
898 
899 	if ((flags & FOF_OFFSET) == 0)
900 		uio->uio_offset = foffset_lock(fp, flags);
901 }
902 
903 void
904 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
905 {
906 
907 	if ((flags & FOF_OFFSET) == 0)
908 		foffset_unlock(fp, uio->uio_offset, flags);
909 }
910 
911 static int
912 get_advice(struct file *fp, struct uio *uio)
913 {
914 	struct mtx *mtxp;
915 	int ret;
916 
917 	ret = POSIX_FADV_NORMAL;
918 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
919 		return (ret);
920 
921 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
922 	mtx_lock(mtxp);
923 	if (fp->f_advice != NULL &&
924 	    uio->uio_offset >= fp->f_advice->fa_start &&
925 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
926 		ret = fp->f_advice->fa_advice;
927 	mtx_unlock(mtxp);
928 	return (ret);
929 }
930 
931 static int
932 get_write_ioflag(struct file *fp)
933 {
934 	int ioflag;
935 	struct mount *mp;
936 	struct vnode *vp;
937 
938 	ioflag = 0;
939 	vp = fp->f_vnode;
940 	mp = atomic_load_ptr(&vp->v_mount);
941 
942 	if ((fp->f_flag & O_DIRECT) != 0)
943 		ioflag |= IO_DIRECT;
944 
945 	if ((fp->f_flag & O_FSYNC) != 0 ||
946 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
947 		ioflag |= IO_SYNC;
948 
949 	/*
950 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
951 	 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
952 	 * fall back to full O_SYNC behavior.
953 	 */
954 	if ((fp->f_flag & O_DSYNC) != 0)
955 		ioflag |= IO_SYNC | IO_DATASYNC;
956 
957 	return (ioflag);
958 }
959 
960 int
961 vn_read_from_obj(struct vnode *vp, struct uio *uio)
962 {
963 	vm_object_t obj;
964 	vm_page_t ma[io_hold_cnt + 2];
965 	off_t off, vsz;
966 	ssize_t resid;
967 	int error, i, j;
968 
969 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
970 	obj = atomic_load_ptr(&vp->v_object);
971 	if (obj == NULL)
972 		return (EJUSTRETURN);
973 
974 	/*
975 	 * Depends on type stability of vm_objects.
976 	 */
977 	vm_object_pip_add(obj, 1);
978 	if ((obj->flags & OBJ_DEAD) != 0) {
979 		/*
980 		 * Note that object might be already reused from the
981 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
982 		 * we recheck for DOOMED vnode state after all pages
983 		 * are busied, and retract then.
984 		 *
985 		 * But we check for OBJ_DEAD to ensure that we do not
986 		 * busy pages while vm_object_terminate_pages()
987 		 * processes the queue.
988 		 */
989 		error = EJUSTRETURN;
990 		goto out_pip;
991 	}
992 
993 	resid = uio->uio_resid;
994 	off = uio->uio_offset;
995 	for (i = 0; resid > 0; i++) {
996 		MPASS(i < io_hold_cnt + 2);
997 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
998 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
999 		    VM_ALLOC_NOWAIT);
1000 		if (ma[i] == NULL)
1001 			break;
1002 
1003 		/*
1004 		 * Skip invalid pages.  Valid mask can be partial only
1005 		 * at EOF, and we clip later.
1006 		 */
1007 		if (vm_page_none_valid(ma[i])) {
1008 			vm_page_sunbusy(ma[i]);
1009 			break;
1010 		}
1011 
1012 		resid -= PAGE_SIZE;
1013 		off += PAGE_SIZE;
1014 	}
1015 	if (i == 0) {
1016 		error = EJUSTRETURN;
1017 		goto out_pip;
1018 	}
1019 
1020 	/*
1021 	 * Check VIRF_DOOMED after we busied our pages.  Since
1022 	 * vgonel() terminates the vnode' vm_object, it cannot
1023 	 * process past pages busied by us.
1024 	 */
1025 	if (VN_IS_DOOMED(vp)) {
1026 		error = EJUSTRETURN;
1027 		goto out;
1028 	}
1029 
1030 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1031 	if (resid > uio->uio_resid)
1032 		resid = uio->uio_resid;
1033 
1034 	/*
1035 	 * Unlocked read of vnp_size is safe because truncation cannot
1036 	 * pass busied page.  But we load vnp_size into a local
1037 	 * variable so that possible concurrent extension does not
1038 	 * break calculation.
1039 	 */
1040 #if defined(__powerpc__) && !defined(__powerpc64__)
1041 	vsz = obj->un_pager.vnp.vnp_size;
1042 #else
1043 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1044 #endif
1045 	if (uio->uio_offset >= vsz) {
1046 		error = EJUSTRETURN;
1047 		goto out;
1048 	}
1049 	if (uio->uio_offset + resid > vsz)
1050 		resid = vsz - uio->uio_offset;
1051 
1052 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1053 
1054 out:
1055 	for (j = 0; j < i; j++) {
1056 		if (error == 0)
1057 			vm_page_reference(ma[j]);
1058 		vm_page_sunbusy(ma[j]);
1059 	}
1060 out_pip:
1061 	vm_object_pip_wakeup(obj);
1062 	if (error != 0)
1063 		return (error);
1064 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1065 }
1066 
1067 /*
1068  * File table vnode read routine.
1069  */
1070 static int
1071 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1072     struct thread *td)
1073 {
1074 	struct vnode *vp;
1075 	off_t orig_offset;
1076 	int error, ioflag;
1077 	int advice;
1078 
1079 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1080 	    uio->uio_td, td));
1081 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1082 	vp = fp->f_vnode;
1083 	ioflag = 0;
1084 	if (fp->f_flag & FNONBLOCK)
1085 		ioflag |= IO_NDELAY;
1086 	if (fp->f_flag & O_DIRECT)
1087 		ioflag |= IO_DIRECT;
1088 
1089 	/*
1090 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1091 	 * allows us to avoid unneeded work outright.
1092 	 */
1093 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1094 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1095 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1096 		if (error == 0) {
1097 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1098 			return (0);
1099 		}
1100 		if (error != EJUSTRETURN)
1101 			return (error);
1102 	}
1103 
1104 	advice = get_advice(fp, uio);
1105 	vn_lock(vp, LK_SHARED | LK_RETRY);
1106 
1107 	switch (advice) {
1108 	case POSIX_FADV_NORMAL:
1109 	case POSIX_FADV_SEQUENTIAL:
1110 	case POSIX_FADV_NOREUSE:
1111 		ioflag |= sequential_heuristic(uio, fp);
1112 		break;
1113 	case POSIX_FADV_RANDOM:
1114 		/* Disable read-ahead for random I/O. */
1115 		break;
1116 	}
1117 	orig_offset = uio->uio_offset;
1118 
1119 #ifdef MAC
1120 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1121 	if (error == 0)
1122 #endif
1123 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1124 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1125 	VOP_UNLOCK(vp);
1126 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1127 	    orig_offset != uio->uio_offset)
1128 		/*
1129 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1130 		 * for the backing file after a POSIX_FADV_NOREUSE
1131 		 * read(2).
1132 		 */
1133 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1134 		    POSIX_FADV_DONTNEED);
1135 	return (error);
1136 }
1137 
1138 /*
1139  * File table vnode write routine.
1140  */
1141 static int
1142 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1143     struct thread *td)
1144 {
1145 	struct vnode *vp;
1146 	struct mount *mp;
1147 	off_t orig_offset;
1148 	int error, ioflag;
1149 	int advice;
1150 	bool need_finished_write;
1151 
1152 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1153 	    uio->uio_td, td));
1154 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1155 	vp = fp->f_vnode;
1156 	if (vp->v_type == VREG)
1157 		bwillwrite();
1158 	ioflag = IO_UNIT;
1159 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1160 		ioflag |= IO_APPEND;
1161 	if ((fp->f_flag & FNONBLOCK) != 0)
1162 		ioflag |= IO_NDELAY;
1163 	ioflag |= get_write_ioflag(fp);
1164 
1165 	mp = NULL;
1166 	need_finished_write = false;
1167 	if (vp->v_type != VCHR) {
1168 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1169 		if (error != 0)
1170 			goto unlock;
1171 		need_finished_write = true;
1172 	}
1173 
1174 	advice = get_advice(fp, uio);
1175 
1176 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1177 	switch (advice) {
1178 	case POSIX_FADV_NORMAL:
1179 	case POSIX_FADV_SEQUENTIAL:
1180 	case POSIX_FADV_NOREUSE:
1181 		ioflag |= sequential_heuristic(uio, fp);
1182 		break;
1183 	case POSIX_FADV_RANDOM:
1184 		/* XXX: Is this correct? */
1185 		break;
1186 	}
1187 	orig_offset = uio->uio_offset;
1188 
1189 #ifdef MAC
1190 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1191 	if (error == 0)
1192 #endif
1193 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1194 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1195 	VOP_UNLOCK(vp);
1196 	if (need_finished_write)
1197 		vn_finished_write(mp);
1198 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1199 	    orig_offset != uio->uio_offset)
1200 		/*
1201 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1202 		 * for the backing file after a POSIX_FADV_NOREUSE
1203 		 * write(2).
1204 		 */
1205 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1206 		    POSIX_FADV_DONTNEED);
1207 unlock:
1208 	return (error);
1209 }
1210 
1211 /*
1212  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1213  * prevent the following deadlock:
1214  *
1215  * Assume that the thread A reads from the vnode vp1 into userspace
1216  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1217  * currently not resident, then system ends up with the call chain
1218  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1219  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1220  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1221  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1222  * backed by the pages of vnode vp1, and some page in buf2 is not
1223  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1224  *
1225  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1226  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1227  * Instead, it first tries to do the whole range i/o with pagefaults
1228  * disabled. If all pages in the i/o buffer are resident and mapped,
1229  * VOP will succeed (ignoring the genuine filesystem errors).
1230  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1231  * i/o in chunks, with all pages in the chunk prefaulted and held
1232  * using vm_fault_quick_hold_pages().
1233  *
1234  * Filesystems using this deadlock avoidance scheme should use the
1235  * array of the held pages from uio, saved in the curthread->td_ma,
1236  * instead of doing uiomove().  A helper function
1237  * vn_io_fault_uiomove() converts uiomove request into
1238  * uiomove_fromphys() over td_ma array.
1239  *
1240  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1241  * make the current i/o request atomic with respect to other i/os and
1242  * truncations.
1243  */
1244 
1245 /*
1246  * Decode vn_io_fault_args and perform the corresponding i/o.
1247  */
1248 static int
1249 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1250     struct thread *td)
1251 {
1252 	int error, save;
1253 
1254 	error = 0;
1255 	save = vm_fault_disable_pagefaults();
1256 	switch (args->kind) {
1257 	case VN_IO_FAULT_FOP:
1258 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1259 		    uio, args->cred, args->flags, td);
1260 		break;
1261 	case VN_IO_FAULT_VOP:
1262 		if (uio->uio_rw == UIO_READ) {
1263 			error = VOP_READ(args->args.vop_args.vp, uio,
1264 			    args->flags, args->cred);
1265 		} else if (uio->uio_rw == UIO_WRITE) {
1266 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1267 			    args->flags, args->cred);
1268 		}
1269 		break;
1270 	default:
1271 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1272 		    args->kind, uio->uio_rw);
1273 	}
1274 	vm_fault_enable_pagefaults(save);
1275 	return (error);
1276 }
1277 
1278 static int
1279 vn_io_fault_touch(char *base, const struct uio *uio)
1280 {
1281 	int r;
1282 
1283 	r = fubyte(base);
1284 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1285 		return (EFAULT);
1286 	return (0);
1287 }
1288 
1289 static int
1290 vn_io_fault_prefault_user(const struct uio *uio)
1291 {
1292 	char *base;
1293 	const struct iovec *iov;
1294 	size_t len;
1295 	ssize_t resid;
1296 	int error, i;
1297 
1298 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1299 	    ("vn_io_fault_prefault userspace"));
1300 
1301 	error = i = 0;
1302 	iov = uio->uio_iov;
1303 	resid = uio->uio_resid;
1304 	base = iov->iov_base;
1305 	len = iov->iov_len;
1306 	while (resid > 0) {
1307 		error = vn_io_fault_touch(base, uio);
1308 		if (error != 0)
1309 			break;
1310 		if (len < PAGE_SIZE) {
1311 			if (len != 0) {
1312 				error = vn_io_fault_touch(base + len - 1, uio);
1313 				if (error != 0)
1314 					break;
1315 				resid -= len;
1316 			}
1317 			if (++i >= uio->uio_iovcnt)
1318 				break;
1319 			iov = uio->uio_iov + i;
1320 			base = iov->iov_base;
1321 			len = iov->iov_len;
1322 		} else {
1323 			len -= PAGE_SIZE;
1324 			base += PAGE_SIZE;
1325 			resid -= PAGE_SIZE;
1326 		}
1327 	}
1328 	return (error);
1329 }
1330 
1331 /*
1332  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1333  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1334  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1335  * into args and call vn_io_fault1() to handle faults during the user
1336  * mode buffer accesses.
1337  */
1338 static int
1339 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1340     struct thread *td)
1341 {
1342 	vm_page_t ma[io_hold_cnt + 2];
1343 	struct uio *uio_clone, short_uio;
1344 	struct iovec short_iovec[1];
1345 	vm_page_t *prev_td_ma;
1346 	vm_prot_t prot;
1347 	vm_offset_t addr, end;
1348 	size_t len, resid;
1349 	ssize_t adv;
1350 	int error, cnt, saveheld, prev_td_ma_cnt;
1351 
1352 	if (vn_io_fault_prefault) {
1353 		error = vn_io_fault_prefault_user(uio);
1354 		if (error != 0)
1355 			return (error); /* Or ignore ? */
1356 	}
1357 
1358 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1359 
1360 	/*
1361 	 * The UFS follows IO_UNIT directive and replays back both
1362 	 * uio_offset and uio_resid if an error is encountered during the
1363 	 * operation.  But, since the iovec may be already advanced,
1364 	 * uio is still in an inconsistent state.
1365 	 *
1366 	 * Cache a copy of the original uio, which is advanced to the redo
1367 	 * point using UIO_NOCOPY below.
1368 	 */
1369 	uio_clone = cloneuio(uio);
1370 	resid = uio->uio_resid;
1371 
1372 	short_uio.uio_segflg = UIO_USERSPACE;
1373 	short_uio.uio_rw = uio->uio_rw;
1374 	short_uio.uio_td = uio->uio_td;
1375 
1376 	error = vn_io_fault_doio(args, uio, td);
1377 	if (error != EFAULT)
1378 		goto out;
1379 
1380 	atomic_add_long(&vn_io_faults_cnt, 1);
1381 	uio_clone->uio_segflg = UIO_NOCOPY;
1382 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1383 	uio_clone->uio_segflg = uio->uio_segflg;
1384 
1385 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1386 	prev_td_ma = td->td_ma;
1387 	prev_td_ma_cnt = td->td_ma_cnt;
1388 
1389 	while (uio_clone->uio_resid != 0) {
1390 		len = uio_clone->uio_iov->iov_len;
1391 		if (len == 0) {
1392 			KASSERT(uio_clone->uio_iovcnt >= 1,
1393 			    ("iovcnt underflow"));
1394 			uio_clone->uio_iov++;
1395 			uio_clone->uio_iovcnt--;
1396 			continue;
1397 		}
1398 		if (len > ptoa(io_hold_cnt))
1399 			len = ptoa(io_hold_cnt);
1400 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1401 		end = round_page(addr + len);
1402 		if (end < addr) {
1403 			error = EFAULT;
1404 			break;
1405 		}
1406 		/*
1407 		 * A perfectly misaligned address and length could cause
1408 		 * both the start and the end of the chunk to use partial
1409 		 * page.  +2 accounts for such a situation.
1410 		 */
1411 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1412 		    addr, len, prot, ma, io_hold_cnt + 2);
1413 		if (cnt == -1) {
1414 			error = EFAULT;
1415 			break;
1416 		}
1417 		short_uio.uio_iov = &short_iovec[0];
1418 		short_iovec[0].iov_base = (void *)addr;
1419 		short_uio.uio_iovcnt = 1;
1420 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1421 		short_uio.uio_offset = uio_clone->uio_offset;
1422 		td->td_ma = ma;
1423 		td->td_ma_cnt = cnt;
1424 
1425 		error = vn_io_fault_doio(args, &short_uio, td);
1426 		vm_page_unhold_pages(ma, cnt);
1427 		adv = len - short_uio.uio_resid;
1428 
1429 		uio_clone->uio_iov->iov_base =
1430 		    (char *)uio_clone->uio_iov->iov_base + adv;
1431 		uio_clone->uio_iov->iov_len -= adv;
1432 		uio_clone->uio_resid -= adv;
1433 		uio_clone->uio_offset += adv;
1434 
1435 		uio->uio_resid -= adv;
1436 		uio->uio_offset += adv;
1437 
1438 		if (error != 0 || adv == 0)
1439 			break;
1440 	}
1441 	td->td_ma = prev_td_ma;
1442 	td->td_ma_cnt = prev_td_ma_cnt;
1443 	curthread_pflags_restore(saveheld);
1444 out:
1445 	freeuio(uio_clone);
1446 	return (error);
1447 }
1448 
1449 static int
1450 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1451     int flags, struct thread *td)
1452 {
1453 	fo_rdwr_t *doio;
1454 	struct vnode *vp;
1455 	void *rl_cookie;
1456 	struct vn_io_fault_args args;
1457 	int error;
1458 	bool do_io_fault, do_rangelock;
1459 
1460 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1461 	vp = fp->f_vnode;
1462 
1463 	/*
1464 	 * The ability to read(2) on a directory has historically been
1465 	 * allowed for all users, but this can and has been the source of
1466 	 * at least one security issue in the past.  As such, it is now hidden
1467 	 * away behind a sysctl for those that actually need it to use it, and
1468 	 * restricted to root when it's turned on to make it relatively safe to
1469 	 * leave on for longer sessions of need.
1470 	 */
1471 	if (vp->v_type == VDIR) {
1472 		KASSERT(uio->uio_rw == UIO_READ,
1473 		    ("illegal write attempted on a directory"));
1474 		if (!vfs_allow_read_dir)
1475 			return (EISDIR);
1476 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1477 			return (EISDIR);
1478 	}
1479 
1480 	do_io_fault = do_vn_io_fault(vp, uio);
1481 	do_rangelock = do_io_fault || (vn_irflag_read(vp) & VIRF_PGREAD) != 0;
1482 	foffset_lock_uio(fp, uio, flags);
1483 	if (do_rangelock) {
1484 		if (uio->uio_rw == UIO_READ) {
1485 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1486 			    uio->uio_offset + uio->uio_resid);
1487 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1488 		    (flags & FOF_OFFSET) == 0) {
1489 			/* For appenders, punt and lock the whole range. */
1490 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1491 		} else {
1492 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1493 			    uio->uio_offset + uio->uio_resid);
1494 		}
1495 	}
1496 	if (do_io_fault) {
1497 		args.kind = VN_IO_FAULT_FOP;
1498 		args.args.fop_args.fp = fp;
1499 		args.args.fop_args.doio = doio;
1500 		args.cred = active_cred;
1501 		args.flags = flags | FOF_OFFSET;
1502 		error = vn_io_fault1(vp, uio, &args, td);
1503 	} else {
1504 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1505 	}
1506 	if (do_rangelock)
1507 		vn_rangelock_unlock(vp, rl_cookie);
1508 	foffset_unlock_uio(fp, uio, flags);
1509 	return (error);
1510 }
1511 
1512 /*
1513  * Helper function to perform the requested uiomove operation using
1514  * the held pages for io->uio_iov[0].iov_base buffer instead of
1515  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1516  * instead of iov_base prevents page faults that could occur due to
1517  * pmap_collect() invalidating the mapping created by
1518  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1519  * object cleanup revoking the write access from page mappings.
1520  *
1521  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1522  * instead of plain uiomove().
1523  */
1524 int
1525 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1526 {
1527 	struct uio transp_uio;
1528 	struct iovec transp_iov[1];
1529 	struct thread *td;
1530 	size_t adv;
1531 	int error, pgadv;
1532 
1533 	td = curthread;
1534 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1535 	    uio->uio_segflg != UIO_USERSPACE)
1536 		return (uiomove(data, xfersize, uio));
1537 
1538 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1539 	transp_iov[0].iov_base = data;
1540 	transp_uio.uio_iov = &transp_iov[0];
1541 	transp_uio.uio_iovcnt = 1;
1542 	if (xfersize > uio->uio_resid)
1543 		xfersize = uio->uio_resid;
1544 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1545 	transp_uio.uio_offset = 0;
1546 	transp_uio.uio_segflg = UIO_SYSSPACE;
1547 	/*
1548 	 * Since transp_iov points to data, and td_ma page array
1549 	 * corresponds to original uio->uio_iov, we need to invert the
1550 	 * direction of the i/o operation as passed to
1551 	 * uiomove_fromphys().
1552 	 */
1553 	switch (uio->uio_rw) {
1554 	case UIO_WRITE:
1555 		transp_uio.uio_rw = UIO_READ;
1556 		break;
1557 	case UIO_READ:
1558 		transp_uio.uio_rw = UIO_WRITE;
1559 		break;
1560 	}
1561 	transp_uio.uio_td = uio->uio_td;
1562 	error = uiomove_fromphys(td->td_ma,
1563 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1564 	    xfersize, &transp_uio);
1565 	adv = xfersize - transp_uio.uio_resid;
1566 	pgadv =
1567 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1568 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1569 	td->td_ma += pgadv;
1570 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1571 	    pgadv));
1572 	td->td_ma_cnt -= pgadv;
1573 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1574 	uio->uio_iov->iov_len -= adv;
1575 	uio->uio_resid -= adv;
1576 	uio->uio_offset += adv;
1577 	return (error);
1578 }
1579 
1580 int
1581 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1582     struct uio *uio)
1583 {
1584 	struct thread *td;
1585 	vm_offset_t iov_base;
1586 	int cnt, pgadv;
1587 
1588 	td = curthread;
1589 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1590 	    uio->uio_segflg != UIO_USERSPACE)
1591 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1592 
1593 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1594 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1595 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1596 	switch (uio->uio_rw) {
1597 	case UIO_WRITE:
1598 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1599 		    offset, cnt);
1600 		break;
1601 	case UIO_READ:
1602 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1603 		    cnt);
1604 		break;
1605 	}
1606 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1607 	td->td_ma += pgadv;
1608 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1609 	    pgadv));
1610 	td->td_ma_cnt -= pgadv;
1611 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1612 	uio->uio_iov->iov_len -= cnt;
1613 	uio->uio_resid -= cnt;
1614 	uio->uio_offset += cnt;
1615 	return (0);
1616 }
1617 
1618 /*
1619  * File table truncate routine.
1620  */
1621 static int
1622 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1623     struct thread *td)
1624 {
1625 	struct mount *mp;
1626 	struct vnode *vp;
1627 	void *rl_cookie;
1628 	int error;
1629 
1630 	vp = fp->f_vnode;
1631 
1632 retry:
1633 	/*
1634 	 * Lock the whole range for truncation.  Otherwise split i/o
1635 	 * might happen partly before and partly after the truncation.
1636 	 */
1637 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1638 	error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1639 	if (error)
1640 		goto out1;
1641 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1642 	AUDIT_ARG_VNODE1(vp);
1643 	if (vp->v_type == VDIR) {
1644 		error = EISDIR;
1645 		goto out;
1646 	}
1647 #ifdef MAC
1648 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1649 	if (error)
1650 		goto out;
1651 #endif
1652 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1653 	    fp->f_cred);
1654 out:
1655 	VOP_UNLOCK(vp);
1656 	vn_finished_write(mp);
1657 out1:
1658 	vn_rangelock_unlock(vp, rl_cookie);
1659 	if (error == ERELOOKUP)
1660 		goto retry;
1661 	return (error);
1662 }
1663 
1664 /*
1665  * Truncate a file that is already locked.
1666  */
1667 int
1668 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1669     struct ucred *cred)
1670 {
1671 	struct vattr vattr;
1672 	int error;
1673 
1674 	error = VOP_ADD_WRITECOUNT(vp, 1);
1675 	if (error == 0) {
1676 		VATTR_NULL(&vattr);
1677 		vattr.va_size = length;
1678 		if (sync)
1679 			vattr.va_vaflags |= VA_SYNC;
1680 		error = VOP_SETATTR(vp, &vattr, cred);
1681 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1682 	}
1683 	return (error);
1684 }
1685 
1686 /*
1687  * File table vnode stat routine.
1688  */
1689 int
1690 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred)
1691 {
1692 	struct vnode *vp = fp->f_vnode;
1693 	int error;
1694 
1695 	vn_lock(vp, LK_SHARED | LK_RETRY);
1696 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1697 	VOP_UNLOCK(vp);
1698 
1699 	return (error);
1700 }
1701 
1702 /*
1703  * File table vnode ioctl routine.
1704  */
1705 static int
1706 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1707     struct thread *td)
1708 {
1709 	struct vnode *vp;
1710 	struct fiobmap2_arg *bmarg;
1711 	off_t size;
1712 	int error;
1713 
1714 	vp = fp->f_vnode;
1715 	switch (vp->v_type) {
1716 	case VDIR:
1717 	case VREG:
1718 		switch (com) {
1719 		case FIONREAD:
1720 			error = vn_getsize(vp, &size, active_cred);
1721 			if (error == 0)
1722 				*(int *)data = size - fp->f_offset;
1723 			return (error);
1724 		case FIOBMAP2:
1725 			bmarg = (struct fiobmap2_arg *)data;
1726 			vn_lock(vp, LK_SHARED | LK_RETRY);
1727 #ifdef MAC
1728 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1729 			    vp);
1730 			if (error == 0)
1731 #endif
1732 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1733 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1734 			VOP_UNLOCK(vp);
1735 			return (error);
1736 		case FIONBIO:
1737 		case FIOASYNC:
1738 			return (0);
1739 		default:
1740 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1741 			    active_cred, td));
1742 		}
1743 		break;
1744 	case VCHR:
1745 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1746 		    active_cred, td));
1747 	default:
1748 		return (ENOTTY);
1749 	}
1750 }
1751 
1752 /*
1753  * File table vnode poll routine.
1754  */
1755 static int
1756 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1757     struct thread *td)
1758 {
1759 	struct vnode *vp;
1760 	int error;
1761 
1762 	vp = fp->f_vnode;
1763 #if defined(MAC) || defined(AUDIT)
1764 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1765 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1766 		AUDIT_ARG_VNODE1(vp);
1767 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1768 		VOP_UNLOCK(vp);
1769 		if (error != 0)
1770 			return (error);
1771 	}
1772 #endif
1773 	error = VOP_POLL(vp, events, fp->f_cred, td);
1774 	return (error);
1775 }
1776 
1777 /*
1778  * Acquire the requested lock and then check for validity.  LK_RETRY
1779  * permits vn_lock to return doomed vnodes.
1780  */
1781 static int __noinline
1782 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1783     int error)
1784 {
1785 
1786 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1787 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1788 
1789 	if (error == 0)
1790 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1791 
1792 	if ((flags & LK_RETRY) == 0) {
1793 		if (error == 0) {
1794 			VOP_UNLOCK(vp);
1795 			error = ENOENT;
1796 		}
1797 		return (error);
1798 	}
1799 
1800 	/*
1801 	 * LK_RETRY case.
1802 	 *
1803 	 * Nothing to do if we got the lock.
1804 	 */
1805 	if (error == 0)
1806 		return (0);
1807 
1808 	/*
1809 	 * Interlock was dropped by the call in _vn_lock.
1810 	 */
1811 	flags &= ~LK_INTERLOCK;
1812 	do {
1813 		error = VOP_LOCK1(vp, flags, file, line);
1814 	} while (error != 0);
1815 	return (0);
1816 }
1817 
1818 int
1819 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1820 {
1821 	int error;
1822 
1823 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1824 	    ("vn_lock: no locktype (%d passed)", flags));
1825 	VNPASS(vp->v_holdcnt > 0, vp);
1826 	error = VOP_LOCK1(vp, flags, file, line);
1827 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1828 		return (_vn_lock_fallback(vp, flags, file, line, error));
1829 	return (0);
1830 }
1831 
1832 /*
1833  * File table vnode close routine.
1834  */
1835 static int
1836 vn_closefile(struct file *fp, struct thread *td)
1837 {
1838 	struct vnode *vp;
1839 	struct flock lf;
1840 	int error;
1841 	bool ref;
1842 
1843 	vp = fp->f_vnode;
1844 	fp->f_ops = &badfileops;
1845 	ref = (fp->f_flag & FHASLOCK) != 0;
1846 
1847 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1848 
1849 	if (__predict_false(ref)) {
1850 		lf.l_whence = SEEK_SET;
1851 		lf.l_start = 0;
1852 		lf.l_len = 0;
1853 		lf.l_type = F_UNLCK;
1854 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1855 		vrele(vp);
1856 	}
1857 	return (error);
1858 }
1859 
1860 /*
1861  * Preparing to start a filesystem write operation. If the operation is
1862  * permitted, then we bump the count of operations in progress and
1863  * proceed. If a suspend request is in progress, we wait until the
1864  * suspension is over, and then proceed.
1865  */
1866 static int
1867 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1868 {
1869 	struct mount_pcpu *mpcpu;
1870 	int error, mflags;
1871 
1872 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1873 	    vfs_op_thread_enter(mp, mpcpu)) {
1874 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1875 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1876 		vfs_op_thread_exit(mp, mpcpu);
1877 		return (0);
1878 	}
1879 
1880 	if (mplocked)
1881 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1882 	else
1883 		MNT_ILOCK(mp);
1884 
1885 	error = 0;
1886 
1887 	/*
1888 	 * Check on status of suspension.
1889 	 */
1890 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1891 	    mp->mnt_susp_owner != curthread) {
1892 		mflags = 0;
1893 		if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
1894 			if (flags & V_PCATCH)
1895 				mflags |= PCATCH;
1896 		}
1897 		mflags |= (PUSER - 1);
1898 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1899 			if ((flags & V_NOWAIT) != 0) {
1900 				error = EWOULDBLOCK;
1901 				goto unlock;
1902 			}
1903 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1904 			    "suspfs", 0);
1905 			if (error != 0)
1906 				goto unlock;
1907 		}
1908 	}
1909 	if ((flags & V_XSLEEP) != 0)
1910 		goto unlock;
1911 	mp->mnt_writeopcount++;
1912 unlock:
1913 	if (error != 0 || (flags & V_XSLEEP) != 0)
1914 		MNT_REL(mp);
1915 	MNT_IUNLOCK(mp);
1916 	return (error);
1917 }
1918 
1919 int
1920 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1921 {
1922 	struct mount *mp;
1923 	int error;
1924 
1925 	KASSERT((flags & ~V_VALID_FLAGS) == 0,
1926 	    ("%s: invalid flags passed %d\n", __func__, flags));
1927 
1928 	error = 0;
1929 	/*
1930 	 * If a vnode is provided, get and return the mount point that
1931 	 * to which it will write.
1932 	 */
1933 	if (vp != NULL) {
1934 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1935 			*mpp = NULL;
1936 			if (error != EOPNOTSUPP)
1937 				return (error);
1938 			return (0);
1939 		}
1940 	}
1941 	if ((mp = *mpp) == NULL)
1942 		return (0);
1943 
1944 	/*
1945 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1946 	 * a vfs_ref().
1947 	 * As long as a vnode is not provided we need to acquire a
1948 	 * refcount for the provided mountpoint too, in order to
1949 	 * emulate a vfs_ref().
1950 	 */
1951 	if (vp == NULL)
1952 		vfs_ref(mp);
1953 
1954 	error = vn_start_write_refed(mp, flags, false);
1955 	if (error != 0 && (flags & V_NOWAIT) == 0)
1956 		*mpp = NULL;
1957 	return (error);
1958 }
1959 
1960 /*
1961  * Secondary suspension. Used by operations such as vop_inactive
1962  * routines that are needed by the higher level functions. These
1963  * are allowed to proceed until all the higher level functions have
1964  * completed (indicated by mnt_writeopcount dropping to zero). At that
1965  * time, these operations are halted until the suspension is over.
1966  */
1967 int
1968 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1969 {
1970 	struct mount *mp;
1971 	int error, mflags;
1972 
1973 	KASSERT((flags & (~V_VALID_FLAGS | V_XSLEEP)) == 0,
1974 	    ("%s: invalid flags passed %d\n", __func__, flags));
1975 
1976  retry:
1977 	if (vp != NULL) {
1978 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1979 			*mpp = NULL;
1980 			if (error != EOPNOTSUPP)
1981 				return (error);
1982 			return (0);
1983 		}
1984 	}
1985 	/*
1986 	 * If we are not suspended or have not yet reached suspended
1987 	 * mode, then let the operation proceed.
1988 	 */
1989 	if ((mp = *mpp) == NULL)
1990 		return (0);
1991 
1992 	/*
1993 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1994 	 * a vfs_ref().
1995 	 * As long as a vnode is not provided we need to acquire a
1996 	 * refcount for the provided mountpoint too, in order to
1997 	 * emulate a vfs_ref().
1998 	 */
1999 	MNT_ILOCK(mp);
2000 	if (vp == NULL)
2001 		MNT_REF(mp);
2002 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
2003 		mp->mnt_secondary_writes++;
2004 		mp->mnt_secondary_accwrites++;
2005 		MNT_IUNLOCK(mp);
2006 		return (0);
2007 	}
2008 	if ((flags & V_NOWAIT) != 0) {
2009 		MNT_REL(mp);
2010 		MNT_IUNLOCK(mp);
2011 		*mpp = NULL;
2012 		return (EWOULDBLOCK);
2013 	}
2014 	/*
2015 	 * Wait for the suspension to finish.
2016 	 */
2017 	mflags = 0;
2018 	if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
2019 		if ((flags & V_PCATCH) != 0)
2020 			mflags |= PCATCH;
2021 	}
2022 	mflags |= (PUSER - 1) | PDROP;
2023 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0);
2024 	vfs_rel(mp);
2025 	if (error == 0)
2026 		goto retry;
2027 	*mpp = NULL;
2028 	return (error);
2029 }
2030 
2031 /*
2032  * Filesystem write operation has completed. If we are suspending and this
2033  * operation is the last one, notify the suspender that the suspension is
2034  * now in effect.
2035  */
2036 void
2037 vn_finished_write(struct mount *mp)
2038 {
2039 	struct mount_pcpu *mpcpu;
2040 	int c;
2041 
2042 	if (mp == NULL)
2043 		return;
2044 
2045 	if (vfs_op_thread_enter(mp, mpcpu)) {
2046 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2047 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2048 		vfs_op_thread_exit(mp, mpcpu);
2049 		return;
2050 	}
2051 
2052 	MNT_ILOCK(mp);
2053 	vfs_assert_mount_counters(mp);
2054 	MNT_REL(mp);
2055 	c = --mp->mnt_writeopcount;
2056 	if (mp->mnt_vfs_ops == 0) {
2057 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2058 		MNT_IUNLOCK(mp);
2059 		return;
2060 	}
2061 	if (c < 0)
2062 		vfs_dump_mount_counters(mp);
2063 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2064 		wakeup(&mp->mnt_writeopcount);
2065 	MNT_IUNLOCK(mp);
2066 }
2067 
2068 /*
2069  * Filesystem secondary write operation has completed. If we are
2070  * suspending and this operation is the last one, notify the suspender
2071  * that the suspension is now in effect.
2072  */
2073 void
2074 vn_finished_secondary_write(struct mount *mp)
2075 {
2076 	if (mp == NULL)
2077 		return;
2078 	MNT_ILOCK(mp);
2079 	MNT_REL(mp);
2080 	mp->mnt_secondary_writes--;
2081 	if (mp->mnt_secondary_writes < 0)
2082 		panic("vn_finished_secondary_write: neg cnt");
2083 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2084 	    mp->mnt_secondary_writes <= 0)
2085 		wakeup(&mp->mnt_secondary_writes);
2086 	MNT_IUNLOCK(mp);
2087 }
2088 
2089 /*
2090  * Request a filesystem to suspend write operations.
2091  */
2092 int
2093 vfs_write_suspend(struct mount *mp, int flags)
2094 {
2095 	int error;
2096 
2097 	vfs_op_enter(mp);
2098 
2099 	MNT_ILOCK(mp);
2100 	vfs_assert_mount_counters(mp);
2101 	if (mp->mnt_susp_owner == curthread) {
2102 		vfs_op_exit_locked(mp);
2103 		MNT_IUNLOCK(mp);
2104 		return (EALREADY);
2105 	}
2106 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2107 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2108 
2109 	/*
2110 	 * Unmount holds a write reference on the mount point.  If we
2111 	 * own busy reference and drain for writers, we deadlock with
2112 	 * the reference draining in the unmount path.  Callers of
2113 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2114 	 * vfs_busy() reference is owned and caller is not in the
2115 	 * unmount context.
2116 	 */
2117 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2118 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2119 		vfs_op_exit_locked(mp);
2120 		MNT_IUNLOCK(mp);
2121 		return (EBUSY);
2122 	}
2123 
2124 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2125 	mp->mnt_susp_owner = curthread;
2126 	if (mp->mnt_writeopcount > 0)
2127 		(void) msleep(&mp->mnt_writeopcount,
2128 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2129 	else
2130 		MNT_IUNLOCK(mp);
2131 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2132 		vfs_write_resume(mp, 0);
2133 		/* vfs_write_resume does vfs_op_exit() for us */
2134 	}
2135 	return (error);
2136 }
2137 
2138 /*
2139  * Request a filesystem to resume write operations.
2140  */
2141 void
2142 vfs_write_resume(struct mount *mp, int flags)
2143 {
2144 
2145 	MNT_ILOCK(mp);
2146 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2147 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2148 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2149 				       MNTK_SUSPENDED);
2150 		mp->mnt_susp_owner = NULL;
2151 		wakeup(&mp->mnt_writeopcount);
2152 		wakeup(&mp->mnt_flag);
2153 		curthread->td_pflags &= ~TDP_IGNSUSP;
2154 		if ((flags & VR_START_WRITE) != 0) {
2155 			MNT_REF(mp);
2156 			mp->mnt_writeopcount++;
2157 		}
2158 		MNT_IUNLOCK(mp);
2159 		if ((flags & VR_NO_SUSPCLR) == 0)
2160 			VFS_SUSP_CLEAN(mp);
2161 		vfs_op_exit(mp);
2162 	} else if ((flags & VR_START_WRITE) != 0) {
2163 		MNT_REF(mp);
2164 		vn_start_write_refed(mp, 0, true);
2165 	} else {
2166 		MNT_IUNLOCK(mp);
2167 	}
2168 }
2169 
2170 /*
2171  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2172  * methods.
2173  */
2174 int
2175 vfs_write_suspend_umnt(struct mount *mp)
2176 {
2177 	int error;
2178 
2179 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2180 	    ("vfs_write_suspend_umnt: recursed"));
2181 
2182 	/* dounmount() already called vn_start_write(). */
2183 	for (;;) {
2184 		vn_finished_write(mp);
2185 		error = vfs_write_suspend(mp, 0);
2186 		if (error != 0) {
2187 			vn_start_write(NULL, &mp, V_WAIT);
2188 			return (error);
2189 		}
2190 		MNT_ILOCK(mp);
2191 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2192 			break;
2193 		MNT_IUNLOCK(mp);
2194 		vn_start_write(NULL, &mp, V_WAIT);
2195 	}
2196 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2197 	wakeup(&mp->mnt_flag);
2198 	MNT_IUNLOCK(mp);
2199 	curthread->td_pflags |= TDP_IGNSUSP;
2200 	return (0);
2201 }
2202 
2203 /*
2204  * Implement kqueues for files by translating it to vnode operation.
2205  */
2206 static int
2207 vn_kqfilter(struct file *fp, struct knote *kn)
2208 {
2209 
2210 	return (VOP_KQFILTER(fp->f_vnode, kn));
2211 }
2212 
2213 int
2214 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2215 {
2216 	if ((fp->f_flag & FKQALLOWED) == 0)
2217 		return (EBADF);
2218 	return (vn_kqfilter(fp, kn));
2219 }
2220 
2221 /*
2222  * Simplified in-kernel wrapper calls for extended attribute access.
2223  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2224  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2225  */
2226 int
2227 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2228     const char *attrname, int *buflen, char *buf, struct thread *td)
2229 {
2230 	struct uio	auio;
2231 	struct iovec	iov;
2232 	int	error;
2233 
2234 	iov.iov_len = *buflen;
2235 	iov.iov_base = buf;
2236 
2237 	auio.uio_iov = &iov;
2238 	auio.uio_iovcnt = 1;
2239 	auio.uio_rw = UIO_READ;
2240 	auio.uio_segflg = UIO_SYSSPACE;
2241 	auio.uio_td = td;
2242 	auio.uio_offset = 0;
2243 	auio.uio_resid = *buflen;
2244 
2245 	if ((ioflg & IO_NODELOCKED) == 0)
2246 		vn_lock(vp, LK_SHARED | LK_RETRY);
2247 
2248 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2249 
2250 	/* authorize attribute retrieval as kernel */
2251 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2252 	    td);
2253 
2254 	if ((ioflg & IO_NODELOCKED) == 0)
2255 		VOP_UNLOCK(vp);
2256 
2257 	if (error == 0) {
2258 		*buflen = *buflen - auio.uio_resid;
2259 	}
2260 
2261 	return (error);
2262 }
2263 
2264 /*
2265  * XXX failure mode if partially written?
2266  */
2267 int
2268 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2269     const char *attrname, int buflen, char *buf, struct thread *td)
2270 {
2271 	struct uio	auio;
2272 	struct iovec	iov;
2273 	struct mount	*mp;
2274 	int	error;
2275 
2276 	iov.iov_len = buflen;
2277 	iov.iov_base = buf;
2278 
2279 	auio.uio_iov = &iov;
2280 	auio.uio_iovcnt = 1;
2281 	auio.uio_rw = UIO_WRITE;
2282 	auio.uio_segflg = UIO_SYSSPACE;
2283 	auio.uio_td = td;
2284 	auio.uio_offset = 0;
2285 	auio.uio_resid = buflen;
2286 
2287 	if ((ioflg & IO_NODELOCKED) == 0) {
2288 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2289 			return (error);
2290 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2291 	}
2292 
2293 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2294 
2295 	/* authorize attribute setting as kernel */
2296 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2297 
2298 	if ((ioflg & IO_NODELOCKED) == 0) {
2299 		vn_finished_write(mp);
2300 		VOP_UNLOCK(vp);
2301 	}
2302 
2303 	return (error);
2304 }
2305 
2306 int
2307 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2308     const char *attrname, struct thread *td)
2309 {
2310 	struct mount	*mp;
2311 	int	error;
2312 
2313 	if ((ioflg & IO_NODELOCKED) == 0) {
2314 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2315 			return (error);
2316 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2317 	}
2318 
2319 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2320 
2321 	/* authorize attribute removal as kernel */
2322 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2323 	if (error == EOPNOTSUPP)
2324 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2325 		    NULL, td);
2326 
2327 	if ((ioflg & IO_NODELOCKED) == 0) {
2328 		vn_finished_write(mp);
2329 		VOP_UNLOCK(vp);
2330 	}
2331 
2332 	return (error);
2333 }
2334 
2335 static int
2336 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2337     struct vnode **rvp)
2338 {
2339 
2340 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2341 }
2342 
2343 int
2344 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2345 {
2346 
2347 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2348 	    lkflags, rvp));
2349 }
2350 
2351 int
2352 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2353     int lkflags, struct vnode **rvp)
2354 {
2355 	struct mount *mp;
2356 	int ltype, error;
2357 
2358 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2359 	mp = vp->v_mount;
2360 	ltype = VOP_ISLOCKED(vp);
2361 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2362 	    ("vn_vget_ino: vp not locked"));
2363 	error = vfs_busy(mp, MBF_NOWAIT);
2364 	if (error != 0) {
2365 		vfs_ref(mp);
2366 		VOP_UNLOCK(vp);
2367 		error = vfs_busy(mp, 0);
2368 		vn_lock(vp, ltype | LK_RETRY);
2369 		vfs_rel(mp);
2370 		if (error != 0)
2371 			return (ENOENT);
2372 		if (VN_IS_DOOMED(vp)) {
2373 			vfs_unbusy(mp);
2374 			return (ENOENT);
2375 		}
2376 	}
2377 	VOP_UNLOCK(vp);
2378 	error = alloc(mp, alloc_arg, lkflags, rvp);
2379 	vfs_unbusy(mp);
2380 	if (error != 0 || *rvp != vp)
2381 		vn_lock(vp, ltype | LK_RETRY);
2382 	if (VN_IS_DOOMED(vp)) {
2383 		if (error == 0) {
2384 			if (*rvp == vp)
2385 				vunref(vp);
2386 			else
2387 				vput(*rvp);
2388 		}
2389 		error = ENOENT;
2390 	}
2391 	return (error);
2392 }
2393 
2394 static void
2395 vn_send_sigxfsz(struct proc *p)
2396 {
2397 	PROC_LOCK(p);
2398 	kern_psignal(p, SIGXFSZ);
2399 	PROC_UNLOCK(p);
2400 }
2401 
2402 int
2403 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2404 {
2405 	if (size <= lim_cur(td, RLIMIT_FSIZE))
2406 		return (0);
2407 	vn_send_sigxfsz(td->td_proc);
2408 	return (EFBIG);
2409 }
2410 
2411 static int
2412 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2413     bool adj, struct thread *td)
2414 {
2415 	off_t lim;
2416 	bool ktr_write;
2417 
2418 	if (vp->v_type != VREG)
2419 		return (0);
2420 
2421 	/*
2422 	 * Handle file system maximum file size.
2423 	 */
2424 	if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2425 		if (!adj || uio->uio_offset >= maxfsz)
2426 			return (EFBIG);
2427 		uio->uio_resid = maxfsz - uio->uio_offset;
2428 	}
2429 
2430 	/*
2431 	 * This is kernel write (e.g. vnode_pager) or accounting
2432 	 * write, ignore limit.
2433 	 */
2434 	if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2435 		return (0);
2436 
2437 	/*
2438 	 * Calculate file size limit.
2439 	 */
2440 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2441 	lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2442 	    lim_cur(td, RLIMIT_FSIZE);
2443 
2444 	/*
2445 	 * Is the limit reached?
2446 	 */
2447 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2448 		return (0);
2449 
2450 	/*
2451 	 * Prepared filesystems can handle writes truncated to the
2452 	 * file size limit.
2453 	 */
2454 	if (adj && (uoff_t)uio->uio_offset < lim) {
2455 		uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2456 		return (0);
2457 	}
2458 
2459 	if (!ktr_write || ktr_filesize_limit_signal)
2460 		vn_send_sigxfsz(td->td_proc);
2461 	return (EFBIG);
2462 }
2463 
2464 /*
2465  * Helper for VOP_WRITE() implementations, the common code to
2466  * handle maximum supported file size on the filesystem, and
2467  * RLIMIT_FSIZE, except for special writes from accounting subsystem
2468  * and ktrace.
2469  *
2470  * For maximum file size (maxfsz argument):
2471  * - return EFBIG if uio_offset is beyond it
2472  * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2473  *
2474  * For RLIMIT_FSIZE:
2475  * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2476  * - otherwise, clamp uio_resid if write would extend file beyond limit.
2477  *
2478  * If clamping occured, the adjustment for uio_resid is stored in
2479  * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2480  * from the VOP.
2481  */
2482 int
2483 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2484     ssize_t *resid_adj, struct thread *td)
2485 {
2486 	ssize_t resid_orig;
2487 	int error;
2488 	bool adj;
2489 
2490 	resid_orig = uio->uio_resid;
2491 	adj = resid_adj != NULL;
2492 	error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2493 	if (adj)
2494 		*resid_adj = resid_orig - uio->uio_resid;
2495 	return (error);
2496 }
2497 
2498 void
2499 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2500 {
2501 	uio->uio_resid += resid_adj;
2502 }
2503 
2504 int
2505 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2506     struct thread *td)
2507 {
2508 	return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2509 	    td));
2510 }
2511 
2512 int
2513 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2514     struct thread *td)
2515 {
2516 	struct vnode *vp;
2517 
2518 	vp = fp->f_vnode;
2519 #ifdef AUDIT
2520 	vn_lock(vp, LK_SHARED | LK_RETRY);
2521 	AUDIT_ARG_VNODE1(vp);
2522 	VOP_UNLOCK(vp);
2523 #endif
2524 	return (setfmode(td, active_cred, vp, mode));
2525 }
2526 
2527 int
2528 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2529     struct thread *td)
2530 {
2531 	struct vnode *vp;
2532 
2533 	vp = fp->f_vnode;
2534 #ifdef AUDIT
2535 	vn_lock(vp, LK_SHARED | LK_RETRY);
2536 	AUDIT_ARG_VNODE1(vp);
2537 	VOP_UNLOCK(vp);
2538 #endif
2539 	return (setfown(td, active_cred, vp, uid, gid));
2540 }
2541 
2542 /*
2543  * Remove pages in the range ["start", "end") from the vnode's VM object.  If
2544  * "end" is 0, then the range extends to the end of the object.
2545  */
2546 void
2547 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2548 {
2549 	vm_object_t object;
2550 
2551 	if ((object = vp->v_object) == NULL)
2552 		return;
2553 	VM_OBJECT_WLOCK(object);
2554 	vm_object_page_remove(object, start, end, 0);
2555 	VM_OBJECT_WUNLOCK(object);
2556 }
2557 
2558 /*
2559  * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2560  * mapped into any process' address space.  Filesystems may use this in
2561  * preference to vn_pages_remove() to avoid blocking on pages busied in
2562  * preparation for a VOP_GETPAGES.
2563  */
2564 void
2565 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2566 {
2567 	vm_object_t object;
2568 
2569 	if ((object = vp->v_object) == NULL)
2570 		return;
2571 	VM_OBJECT_WLOCK(object);
2572 	vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2573 	VM_OBJECT_WUNLOCK(object);
2574 }
2575 
2576 int
2577 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2578     struct ucred *cred)
2579 {
2580 	off_t size;
2581 	daddr_t bn, bnp;
2582 	uint64_t bsize;
2583 	off_t noff;
2584 	int error;
2585 
2586 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2587 	    ("%s: Wrong command %lu", __func__, cmd));
2588 	ASSERT_VOP_ELOCKED(vp, "vn_bmap_seekhole_locked");
2589 
2590 	if (vp->v_type != VREG) {
2591 		error = ENOTTY;
2592 		goto out;
2593 	}
2594 	error = vn_getsize_locked(vp, &size, cred);
2595 	if (error != 0)
2596 		goto out;
2597 	noff = *off;
2598 	if (noff < 0 || noff >= size) {
2599 		error = ENXIO;
2600 		goto out;
2601 	}
2602 
2603 	/* See the comment in ufs_bmap_seekdata(). */
2604 	vnode_pager_clean_sync(vp);
2605 
2606 	bsize = vp->v_mount->mnt_stat.f_iosize;
2607 	for (bn = noff / bsize; noff < size; bn++, noff += bsize -
2608 	    noff % bsize) {
2609 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2610 		if (error == EOPNOTSUPP) {
2611 			error = ENOTTY;
2612 			goto out;
2613 		}
2614 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2615 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2616 			noff = bn * bsize;
2617 			if (noff < *off)
2618 				noff = *off;
2619 			goto out;
2620 		}
2621 	}
2622 	if (noff > size)
2623 		noff = size;
2624 	/* noff == size. There is an implicit hole at the end of file. */
2625 	if (cmd == FIOSEEKDATA)
2626 		error = ENXIO;
2627 out:
2628 	if (error == 0)
2629 		*off = noff;
2630 	return (error);
2631 }
2632 
2633 int
2634 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2635 {
2636 	int error;
2637 
2638 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2639 	    ("%s: Wrong command %lu", __func__, cmd));
2640 
2641 	if (vn_lock(vp, LK_EXCLUSIVE) != 0)
2642 		return (EBADF);
2643 	error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2644 	VOP_UNLOCK(vp);
2645 	return (error);
2646 }
2647 
2648 int
2649 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2650 {
2651 	struct ucred *cred;
2652 	struct vnode *vp;
2653 	off_t foffset, fsize, size;
2654 	int error, noneg;
2655 
2656 	cred = td->td_ucred;
2657 	vp = fp->f_vnode;
2658 	noneg = (vp->v_type != VCHR);
2659 	/*
2660 	 * Try to dodge locking for common case of querying the offset.
2661 	 */
2662 	if (whence == L_INCR && offset == 0) {
2663 		foffset = foffset_read(fp);
2664 		if (__predict_false(foffset < 0 && noneg)) {
2665 			return (EOVERFLOW);
2666 		}
2667 		td->td_uretoff.tdu_off = foffset;
2668 		return (0);
2669 	}
2670 	foffset = foffset_lock(fp, 0);
2671 	error = 0;
2672 	switch (whence) {
2673 	case L_INCR:
2674 		if (noneg &&
2675 		    (foffset < 0 ||
2676 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2677 			error = EOVERFLOW;
2678 			break;
2679 		}
2680 		offset += foffset;
2681 		break;
2682 	case L_XTND:
2683 		error = vn_getsize(vp, &fsize, cred);
2684 		if (error != 0)
2685 			break;
2686 
2687 		/*
2688 		 * If the file references a disk device, then fetch
2689 		 * the media size and use that to determine the ending
2690 		 * offset.
2691 		 */
2692 		if (fsize == 0 && vp->v_type == VCHR &&
2693 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2694 			fsize = size;
2695 		if (noneg && offset > 0 && fsize > OFF_MAX - offset) {
2696 			error = EOVERFLOW;
2697 			break;
2698 		}
2699 		offset += fsize;
2700 		break;
2701 	case L_SET:
2702 		break;
2703 	case SEEK_DATA:
2704 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2705 		if (error == ENOTTY)
2706 			error = EINVAL;
2707 		break;
2708 	case SEEK_HOLE:
2709 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2710 		if (error == ENOTTY)
2711 			error = EINVAL;
2712 		break;
2713 	default:
2714 		error = EINVAL;
2715 	}
2716 	if (error == 0 && noneg && offset < 0)
2717 		error = EINVAL;
2718 	if (error != 0)
2719 		goto drop;
2720 	VFS_KNOTE_UNLOCKED(vp, 0);
2721 	td->td_uretoff.tdu_off = offset;
2722 drop:
2723 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2724 	return (error);
2725 }
2726 
2727 int
2728 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2729     struct thread *td)
2730 {
2731 	int error;
2732 
2733 	/*
2734 	 * Grant permission if the caller is the owner of the file, or
2735 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2736 	 * on the file.  If the time pointer is null, then write
2737 	 * permission on the file is also sufficient.
2738 	 *
2739 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2740 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2741 	 * will be allowed to set the times [..] to the current
2742 	 * server time.
2743 	 */
2744 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2745 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2746 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2747 	return (error);
2748 }
2749 
2750 int
2751 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2752 {
2753 	struct vnode *vp;
2754 	int error;
2755 
2756 	if (fp->f_type == DTYPE_FIFO)
2757 		kif->kf_type = KF_TYPE_FIFO;
2758 	else
2759 		kif->kf_type = KF_TYPE_VNODE;
2760 	vp = fp->f_vnode;
2761 	vref(vp);
2762 	FILEDESC_SUNLOCK(fdp);
2763 	error = vn_fill_kinfo_vnode(vp, kif);
2764 	vrele(vp);
2765 	FILEDESC_SLOCK(fdp);
2766 	return (error);
2767 }
2768 
2769 static inline void
2770 vn_fill_junk(struct kinfo_file *kif)
2771 {
2772 	size_t len, olen;
2773 
2774 	/*
2775 	 * Simulate vn_fullpath returning changing values for a given
2776 	 * vp during e.g. coredump.
2777 	 */
2778 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2779 	olen = strlen(kif->kf_path);
2780 	if (len < olen)
2781 		strcpy(&kif->kf_path[len - 1], "$");
2782 	else
2783 		for (; olen < len; olen++)
2784 			strcpy(&kif->kf_path[olen], "A");
2785 }
2786 
2787 int
2788 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2789 {
2790 	struct vattr va;
2791 	char *fullpath, *freepath;
2792 	int error;
2793 
2794 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2795 	freepath = NULL;
2796 	fullpath = "-";
2797 	error = vn_fullpath(vp, &fullpath, &freepath);
2798 	if (error == 0) {
2799 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2800 	}
2801 	if (freepath != NULL)
2802 		free(freepath, M_TEMP);
2803 
2804 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2805 		vn_fill_junk(kif);
2806 	);
2807 
2808 	/*
2809 	 * Retrieve vnode attributes.
2810 	 */
2811 	va.va_fsid = VNOVAL;
2812 	va.va_rdev = NODEV;
2813 	vn_lock(vp, LK_SHARED | LK_RETRY);
2814 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2815 	VOP_UNLOCK(vp);
2816 	if (error != 0)
2817 		return (error);
2818 	if (va.va_fsid != VNOVAL)
2819 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2820 	else
2821 		kif->kf_un.kf_file.kf_file_fsid =
2822 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2823 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2824 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2825 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2826 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2827 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2828 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2829 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2830 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2831 	kif->kf_un.kf_file.kf_file_nlink = va.va_nlink;
2832 	return (0);
2833 }
2834 
2835 int
2836 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2837     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2838     struct thread *td)
2839 {
2840 #ifdef HWPMC_HOOKS
2841 	struct pmckern_map_in pkm;
2842 #endif
2843 	struct mount *mp;
2844 	struct vnode *vp;
2845 	vm_object_t object;
2846 	vm_prot_t maxprot;
2847 	boolean_t writecounted;
2848 	int error;
2849 
2850 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2851     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2852 	/*
2853 	 * POSIX shared-memory objects are defined to have
2854 	 * kernel persistence, and are not defined to support
2855 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2856 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2857 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2858 	 * flag to request this behavior.
2859 	 */
2860 	if ((fp->f_flag & FPOSIXSHM) != 0)
2861 		flags |= MAP_NOSYNC;
2862 #endif
2863 	vp = fp->f_vnode;
2864 
2865 	/*
2866 	 * Ensure that file and memory protections are
2867 	 * compatible.  Note that we only worry about
2868 	 * writability if mapping is shared; in this case,
2869 	 * current and max prot are dictated by the open file.
2870 	 * XXX use the vnode instead?  Problem is: what
2871 	 * credentials do we use for determination? What if
2872 	 * proc does a setuid?
2873 	 */
2874 	mp = vp->v_mount;
2875 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2876 		maxprot = VM_PROT_NONE;
2877 		if ((prot & VM_PROT_EXECUTE) != 0)
2878 			return (EACCES);
2879 	} else
2880 		maxprot = VM_PROT_EXECUTE;
2881 	if ((fp->f_flag & FREAD) != 0)
2882 		maxprot |= VM_PROT_READ;
2883 	else if ((prot & VM_PROT_READ) != 0)
2884 		return (EACCES);
2885 
2886 	/*
2887 	 * If we are sharing potential changes via MAP_SHARED and we
2888 	 * are trying to get write permission although we opened it
2889 	 * without asking for it, bail out.
2890 	 */
2891 	if ((flags & MAP_SHARED) != 0) {
2892 		if ((fp->f_flag & FWRITE) != 0)
2893 			maxprot |= VM_PROT_WRITE;
2894 		else if ((prot & VM_PROT_WRITE) != 0)
2895 			return (EACCES);
2896 	} else {
2897 		maxprot |= VM_PROT_WRITE;
2898 		cap_maxprot |= VM_PROT_WRITE;
2899 	}
2900 	maxprot &= cap_maxprot;
2901 
2902 	/*
2903 	 * For regular files and shared memory, POSIX requires that
2904 	 * the value of foff be a legitimate offset within the data
2905 	 * object.  In particular, negative offsets are invalid.
2906 	 * Blocking negative offsets and overflows here avoids
2907 	 * possible wraparound or user-level access into reserved
2908 	 * ranges of the data object later.  In contrast, POSIX does
2909 	 * not dictate how offsets are used by device drivers, so in
2910 	 * the case of a device mapping a negative offset is passed
2911 	 * on.
2912 	 */
2913 	if (
2914 #ifdef _LP64
2915 	    size > OFF_MAX ||
2916 #endif
2917 	    foff > OFF_MAX - size)
2918 		return (EINVAL);
2919 
2920 	writecounted = FALSE;
2921 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2922 	    &foff, &object, &writecounted);
2923 	if (error != 0)
2924 		return (error);
2925 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2926 	    foff, writecounted, td);
2927 	if (error != 0) {
2928 		/*
2929 		 * If this mapping was accounted for in the vnode's
2930 		 * writecount, then undo that now.
2931 		 */
2932 		if (writecounted)
2933 			vm_pager_release_writecount(object, 0, size);
2934 		vm_object_deallocate(object);
2935 	}
2936 #ifdef HWPMC_HOOKS
2937 	/* Inform hwpmc(4) if an executable is being mapped. */
2938 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2939 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2940 			pkm.pm_file = vp;
2941 			pkm.pm_address = (uintptr_t) *addr;
2942 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2943 		}
2944 	}
2945 #endif
2946 	return (error);
2947 }
2948 
2949 void
2950 vn_fsid(struct vnode *vp, struct vattr *va)
2951 {
2952 	fsid_t *f;
2953 
2954 	f = &vp->v_mount->mnt_stat.f_fsid;
2955 	va->va_fsid = (uint32_t)f->val[1];
2956 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2957 	va->va_fsid += (uint32_t)f->val[0];
2958 }
2959 
2960 int
2961 vn_fsync_buf(struct vnode *vp, int waitfor)
2962 {
2963 	struct buf *bp, *nbp;
2964 	struct bufobj *bo;
2965 	struct mount *mp;
2966 	int error, maxretry;
2967 
2968 	error = 0;
2969 	maxretry = 10000;     /* large, arbitrarily chosen */
2970 	mp = NULL;
2971 	if (vp->v_type == VCHR) {
2972 		VI_LOCK(vp);
2973 		mp = vp->v_rdev->si_mountpt;
2974 		VI_UNLOCK(vp);
2975 	}
2976 	bo = &vp->v_bufobj;
2977 	BO_LOCK(bo);
2978 loop1:
2979 	/*
2980 	 * MARK/SCAN initialization to avoid infinite loops.
2981 	 */
2982         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2983 		bp->b_vflags &= ~BV_SCANNED;
2984 		bp->b_error = 0;
2985 	}
2986 
2987 	/*
2988 	 * Flush all dirty buffers associated with a vnode.
2989 	 */
2990 loop2:
2991 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2992 		if ((bp->b_vflags & BV_SCANNED) != 0)
2993 			continue;
2994 		bp->b_vflags |= BV_SCANNED;
2995 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2996 			if (waitfor != MNT_WAIT)
2997 				continue;
2998 			if (BUF_LOCK(bp,
2999 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
3000 			    BO_LOCKPTR(bo)) != 0) {
3001 				BO_LOCK(bo);
3002 				goto loop1;
3003 			}
3004 			BO_LOCK(bo);
3005 		}
3006 		BO_UNLOCK(bo);
3007 		KASSERT(bp->b_bufobj == bo,
3008 		    ("bp %p wrong b_bufobj %p should be %p",
3009 		    bp, bp->b_bufobj, bo));
3010 		if ((bp->b_flags & B_DELWRI) == 0)
3011 			panic("fsync: not dirty");
3012 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
3013 			vfs_bio_awrite(bp);
3014 		} else {
3015 			bremfree(bp);
3016 			bawrite(bp);
3017 		}
3018 		if (maxretry < 1000)
3019 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
3020 		BO_LOCK(bo);
3021 		goto loop2;
3022 	}
3023 
3024 	/*
3025 	 * If synchronous the caller expects us to completely resolve all
3026 	 * dirty buffers in the system.  Wait for in-progress I/O to
3027 	 * complete (which could include background bitmap writes), then
3028 	 * retry if dirty blocks still exist.
3029 	 */
3030 	if (waitfor == MNT_WAIT) {
3031 		bufobj_wwait(bo, 0, 0);
3032 		if (bo->bo_dirty.bv_cnt > 0) {
3033 			/*
3034 			 * If we are unable to write any of these buffers
3035 			 * then we fail now rather than trying endlessly
3036 			 * to write them out.
3037 			 */
3038 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
3039 				if ((error = bp->b_error) != 0)
3040 					break;
3041 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
3042 			    (error == 0 && --maxretry >= 0))
3043 				goto loop1;
3044 			if (error == 0)
3045 				error = EAGAIN;
3046 		}
3047 	}
3048 	BO_UNLOCK(bo);
3049 	if (error != 0)
3050 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3051 
3052 	return (error);
3053 }
3054 
3055 /*
3056  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
3057  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3058  * to do the actual copy.
3059  * vn_generic_copy_file_range() is factored out, so it can be called
3060  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3061  * different file systems.
3062  */
3063 int
3064 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3065     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3066     struct ucred *outcred, struct thread *fsize_td)
3067 {
3068 	struct mount *inmp, *outmp;
3069 	struct vnode *invpl, *outvpl;
3070 	int error;
3071 	size_t len;
3072 	uint64_t uval;
3073 
3074 	invpl = outvpl = NULL;
3075 	len = *lenp;
3076 	*lenp = 0;		/* For error returns. */
3077 	error = 0;
3078 
3079 	/* Do some sanity checks on the arguments. */
3080 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
3081 		error = EISDIR;
3082 	else if (*inoffp < 0 || *outoffp < 0 ||
3083 	    invp->v_type != VREG || outvp->v_type != VREG)
3084 		error = EINVAL;
3085 	if (error != 0)
3086 		goto out;
3087 
3088 	/* Ensure offset + len does not wrap around. */
3089 	uval = *inoffp;
3090 	uval += len;
3091 	if (uval > INT64_MAX)
3092 		len = INT64_MAX - *inoffp;
3093 	uval = *outoffp;
3094 	uval += len;
3095 	if (uval > INT64_MAX)
3096 		len = INT64_MAX - *outoffp;
3097 	if (len == 0)
3098 		goto out;
3099 
3100 	error = VOP_GETLOWVNODE(invp, &invpl, FREAD);
3101 	if (error != 0)
3102 		goto out;
3103 	error = VOP_GETLOWVNODE(outvp, &outvpl, FWRITE);
3104 	if (error != 0)
3105 		goto out1;
3106 
3107 	inmp = invpl->v_mount;
3108 	outmp = outvpl->v_mount;
3109 	if (inmp == NULL || outmp == NULL)
3110 		goto out2;
3111 
3112 	for (;;) {
3113 		error = vfs_busy(inmp, 0);
3114 		if (error != 0)
3115 			goto out2;
3116 		if (inmp == outmp)
3117 			break;
3118 		error = vfs_busy(outmp, MBF_NOWAIT);
3119 		if (error != 0) {
3120 			vfs_unbusy(inmp);
3121 			error = vfs_busy(outmp, 0);
3122 			if (error == 0) {
3123 				vfs_unbusy(outmp);
3124 				continue;
3125 			}
3126 			goto out2;
3127 		}
3128 		break;
3129 	}
3130 
3131 	/*
3132 	 * If the two vnodes are for the same file system type, call
3133 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3134 	 * which can handle copies across multiple file system types.
3135 	 */
3136 	*lenp = len;
3137 	if (inmp == outmp || inmp->mnt_vfc == outmp->mnt_vfc)
3138 		error = VOP_COPY_FILE_RANGE(invpl, inoffp, outvpl, outoffp,
3139 		    lenp, flags, incred, outcred, fsize_td);
3140 	else
3141 		error = ENOSYS;
3142 	if (error == ENOSYS)
3143 		error = vn_generic_copy_file_range(invpl, inoffp, outvpl,
3144 		    outoffp, lenp, flags, incred, outcred, fsize_td);
3145 	vfs_unbusy(outmp);
3146 	if (inmp != outmp)
3147 		vfs_unbusy(inmp);
3148 out2:
3149 	if (outvpl != NULL)
3150 		vrele(outvpl);
3151 out1:
3152 	if (invpl != NULL)
3153 		vrele(invpl);
3154 out:
3155 	return (error);
3156 }
3157 
3158 /*
3159  * Test len bytes of data starting at dat for all bytes == 0.
3160  * Return true if all bytes are zero, false otherwise.
3161  * Expects dat to be well aligned.
3162  */
3163 static bool
3164 mem_iszero(void *dat, int len)
3165 {
3166 	int i;
3167 	const u_int *p;
3168 	const char *cp;
3169 
3170 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
3171 		if (len >= sizeof(*p)) {
3172 			if (*p != 0)
3173 				return (false);
3174 		} else {
3175 			cp = (const char *)p;
3176 			for (i = 0; i < len; i++, cp++)
3177 				if (*cp != '\0')
3178 					return (false);
3179 		}
3180 	}
3181 	return (true);
3182 }
3183 
3184 /*
3185  * Look for a hole in the output file and, if found, adjust *outoffp
3186  * and *xferp to skip past the hole.
3187  * *xferp is the entire hole length to be written and xfer2 is how many bytes
3188  * to be written as 0's upon return.
3189  */
3190 static off_t
3191 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3192     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3193 {
3194 	int error;
3195 	off_t delta;
3196 
3197 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3198 		*dataoffp = *outoffp;
3199 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3200 		    curthread);
3201 		if (error == 0) {
3202 			*holeoffp = *dataoffp;
3203 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3204 			    curthread);
3205 		}
3206 		if (error != 0 || *holeoffp == *dataoffp) {
3207 			/*
3208 			 * Since outvp is unlocked, it may be possible for
3209 			 * another thread to do a truncate(), lseek(), write()
3210 			 * creating a hole at startoff between the above
3211 			 * VOP_IOCTL() calls, if the other thread does not do
3212 			 * rangelocking.
3213 			 * If that happens, *holeoffp == *dataoffp and finding
3214 			 * the hole has failed, so disable vn_skip_hole().
3215 			 */
3216 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3217 			return (xfer2);
3218 		}
3219 		KASSERT(*dataoffp >= *outoffp,
3220 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3221 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3222 		KASSERT(*holeoffp > *dataoffp,
3223 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3224 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3225 	}
3226 
3227 	/*
3228 	 * If there is a hole before the data starts, advance *outoffp and
3229 	 * *xferp past the hole.
3230 	 */
3231 	if (*dataoffp > *outoffp) {
3232 		delta = *dataoffp - *outoffp;
3233 		if (delta >= *xferp) {
3234 			/* Entire *xferp is a hole. */
3235 			*outoffp += *xferp;
3236 			*xferp = 0;
3237 			return (0);
3238 		}
3239 		*xferp -= delta;
3240 		*outoffp += delta;
3241 		xfer2 = MIN(xfer2, *xferp);
3242 	}
3243 
3244 	/*
3245 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3246 	 * that the write ends at the start of the hole.
3247 	 * *holeoffp should always be greater than *outoffp, but for the
3248 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3249 	 * value.
3250 	 */
3251 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3252 		xfer2 = *holeoffp - *outoffp;
3253 	return (xfer2);
3254 }
3255 
3256 /*
3257  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3258  * dat is a maximum of blksize in length and can be written repeatedly in
3259  * the chunk.
3260  * If growfile == true, just grow the file via vn_truncate_locked() instead
3261  * of doing actual writes.
3262  * If checkhole == true, a hole is being punched, so skip over any hole
3263  * already in the output file.
3264  */
3265 static int
3266 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3267     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3268 {
3269 	struct mount *mp;
3270 	off_t dataoff, holeoff, xfer2;
3271 	int error;
3272 
3273 	/*
3274 	 * Loop around doing writes of blksize until write has been completed.
3275 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3276 	 * done for each iteration, since the xfer argument can be very
3277 	 * large if there is a large hole to punch in the output file.
3278 	 */
3279 	error = 0;
3280 	holeoff = 0;
3281 	do {
3282 		xfer2 = MIN(xfer, blksize);
3283 		if (checkhole) {
3284 			/*
3285 			 * Punching a hole.  Skip writing if there is
3286 			 * already a hole in the output file.
3287 			 */
3288 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3289 			    &dataoff, &holeoff, cred);
3290 			if (xfer == 0)
3291 				break;
3292 			if (holeoff < 0)
3293 				checkhole = false;
3294 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3295 			    (intmax_t)xfer2));
3296 		}
3297 		bwillwrite();
3298 		mp = NULL;
3299 		error = vn_start_write(outvp, &mp, V_WAIT);
3300 		if (error != 0)
3301 			break;
3302 		if (growfile) {
3303 			error = vn_lock(outvp, LK_EXCLUSIVE);
3304 			if (error == 0) {
3305 				error = vn_truncate_locked(outvp, outoff + xfer,
3306 				    false, cred);
3307 				VOP_UNLOCK(outvp);
3308 			}
3309 		} else {
3310 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3311 			if (error == 0) {
3312 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3313 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3314 				    curthread->td_ucred, cred, NULL, curthread);
3315 				outoff += xfer2;
3316 				xfer -= xfer2;
3317 				VOP_UNLOCK(outvp);
3318 			}
3319 		}
3320 		if (mp != NULL)
3321 			vn_finished_write(mp);
3322 	} while (!growfile && xfer > 0 && error == 0);
3323 	return (error);
3324 }
3325 
3326 /*
3327  * Copy a byte range of one file to another.  This function can handle the
3328  * case where invp and outvp are on different file systems.
3329  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3330  * is no better file system specific way to do it.
3331  */
3332 int
3333 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3334     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3335     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3336 {
3337 	struct vattr inva;
3338 	struct mount *mp;
3339 	off_t startoff, endoff, xfer, xfer2;
3340 	u_long blksize;
3341 	int error, interrupted;
3342 	bool cantseek, readzeros, eof, lastblock, holetoeof, sparse;
3343 	ssize_t aresid, r = 0;
3344 	size_t copylen, len, savlen;
3345 	off_t outsize;
3346 	char *dat;
3347 	long holein, holeout;
3348 	struct timespec curts, endts;
3349 
3350 	holein = holeout = 0;
3351 	savlen = len = *lenp;
3352 	error = 0;
3353 	interrupted = 0;
3354 	dat = NULL;
3355 
3356 	error = vn_lock(invp, LK_SHARED);
3357 	if (error != 0)
3358 		goto out;
3359 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3360 		holein = 0;
3361 	error = VOP_GETATTR(invp, &inva, incred);
3362 	if (error == 0 && inva.va_size > OFF_MAX)
3363 		error = EFBIG;
3364 	VOP_UNLOCK(invp);
3365 	if (error != 0)
3366 		goto out;
3367 
3368 	/*
3369 	 * Use va_bytes >= va_size as a hint that the file does not have
3370 	 * sufficient holes to justify the overhead of doing FIOSEEKHOLE.
3371 	 * This hint does not work well for file systems doing compression
3372 	 * and may fail when allocations for extended attributes increases
3373 	 * the value of va_bytes to >= va_size.
3374 	 */
3375 	sparse = true;
3376 	if (holein != 0 && inva.va_bytes >= inva.va_size) {
3377 		holein = 0;
3378 		sparse = false;
3379 	}
3380 
3381 	mp = NULL;
3382 	error = vn_start_write(outvp, &mp, V_WAIT);
3383 	if (error == 0)
3384 		error = vn_lock(outvp, LK_EXCLUSIVE);
3385 	if (error == 0) {
3386 		/*
3387 		 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3388 		 * now that outvp is locked.
3389 		 */
3390 		if (fsize_td != NULL) {
3391 			struct uio io;
3392 
3393 			io.uio_offset = *outoffp;
3394 			io.uio_resid = len;
3395 			error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3396 			len = savlen = io.uio_resid;
3397 			/*
3398 			 * No need to call vn_rlimit_fsizex_res before return,
3399 			 * since the uio is local.
3400 			 */
3401 		}
3402 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3403 			holeout = 0;
3404 		/*
3405 		 * Holes that are past EOF do not need to be written as a block
3406 		 * of zero bytes.  So, truncate the output file as far as
3407 		 * possible and then use size to decide if writing 0
3408 		 * bytes is necessary in the loop below.
3409 		 */
3410 		if (error == 0)
3411 			error = vn_getsize_locked(outvp, &outsize, outcred);
3412 		if (error == 0 && outsize > *outoffp &&
3413 		    *outoffp <= OFF_MAX - len && outsize <= *outoffp + len &&
3414 		    *inoffp < inva.va_size &&
3415 		    *outoffp <= OFF_MAX - (inva.va_size - *inoffp) &&
3416 		    outsize <= *outoffp + (inva.va_size - *inoffp)) {
3417 #ifdef MAC
3418 			error = mac_vnode_check_write(curthread->td_ucred,
3419 			    outcred, outvp);
3420 			if (error == 0)
3421 #endif
3422 				error = vn_truncate_locked(outvp, *outoffp,
3423 				    false, outcred);
3424 			if (error == 0)
3425 				outsize = *outoffp;
3426 		}
3427 		VOP_UNLOCK(outvp);
3428 	}
3429 	if (mp != NULL)
3430 		vn_finished_write(mp);
3431 	if (error != 0)
3432 		goto out;
3433 
3434 	if (sparse && holein == 0 && holeout > 0) {
3435 		/*
3436 		 * For this special case, the input data will be scanned
3437 		 * for blocks of all 0 bytes.  For these blocks, the
3438 		 * write can be skipped for the output file to create
3439 		 * an unallocated region.
3440 		 * Therefore, use the appropriate size for the output file.
3441 		 */
3442 		blksize = holeout;
3443 		if (blksize <= 512) {
3444 			/*
3445 			 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3446 			 * of 512, although it actually only creates
3447 			 * unallocated regions for blocks >= f_iosize.
3448 			 */
3449 			blksize = outvp->v_mount->mnt_stat.f_iosize;
3450 		}
3451 	} else {
3452 		/*
3453 		 * Use the larger of the two f_iosize values.  If they are
3454 		 * not the same size, one will normally be an exact multiple of
3455 		 * the other, since they are both likely to be a power of 2.
3456 		 */
3457 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3458 		    outvp->v_mount->mnt_stat.f_iosize);
3459 	}
3460 
3461 	/* Clip to sane limits. */
3462 	if (blksize < 4096)
3463 		blksize = 4096;
3464 	else if (blksize > maxphys)
3465 		blksize = maxphys;
3466 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3467 
3468 	/*
3469 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3470 	 * to find holes.  Otherwise, just scan the read block for all 0s
3471 	 * in the inner loop where the data copying is done.
3472 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3473 	 * support holes on the server, but do not support FIOSEEKHOLE.
3474 	 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3475 	 * that this function should return after 1second with a partial
3476 	 * completion.
3477 	 */
3478 	if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3479 		getnanouptime(&endts);
3480 		endts.tv_sec++;
3481 	} else
3482 		timespecclear(&endts);
3483 	holetoeof = eof = false;
3484 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3485 		endoff = 0;			/* To shut up compilers. */
3486 		cantseek = true;
3487 		startoff = *inoffp;
3488 		copylen = len;
3489 
3490 		/*
3491 		 * Find the next data area.  If there is just a hole to EOF,
3492 		 * FIOSEEKDATA should fail with ENXIO.
3493 		 * (I do not know if any file system will report a hole to
3494 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3495 		 *  will fail for those file systems.)
3496 		 *
3497 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3498 		 * the code just falls through to the inner copy loop.
3499 		 */
3500 		error = EINVAL;
3501 		if (holein > 0) {
3502 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3503 			    incred, curthread);
3504 			if (error == ENXIO) {
3505 				startoff = endoff = inva.va_size;
3506 				eof = holetoeof = true;
3507 				error = 0;
3508 			}
3509 		}
3510 		if (error == 0 && !holetoeof) {
3511 			endoff = startoff;
3512 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3513 			    incred, curthread);
3514 			/*
3515 			 * Since invp is unlocked, it may be possible for
3516 			 * another thread to do a truncate(), lseek(), write()
3517 			 * creating a hole at startoff between the above
3518 			 * VOP_IOCTL() calls, if the other thread does not do
3519 			 * rangelocking.
3520 			 * If that happens, startoff == endoff and finding
3521 			 * the hole has failed, so set an error.
3522 			 */
3523 			if (error == 0 && startoff == endoff)
3524 				error = EINVAL; /* Any error. Reset to 0. */
3525 		}
3526 		if (error == 0) {
3527 			if (startoff > *inoffp) {
3528 				/* Found hole before data block. */
3529 				xfer = MIN(startoff - *inoffp, len);
3530 				if (*outoffp < outsize) {
3531 					/* Must write 0s to punch hole. */
3532 					xfer2 = MIN(outsize - *outoffp,
3533 					    xfer);
3534 					memset(dat, 0, MIN(xfer2, blksize));
3535 					error = vn_write_outvp(outvp, dat,
3536 					    *outoffp, xfer2, blksize, false,
3537 					    holeout > 0, outcred);
3538 				}
3539 
3540 				if (error == 0 && *outoffp + xfer >
3541 				    outsize && (xfer == len || holetoeof)) {
3542 					/* Grow output file (hole at end). */
3543 					error = vn_write_outvp(outvp, dat,
3544 					    *outoffp, xfer, blksize, true,
3545 					    false, outcred);
3546 				}
3547 				if (error == 0) {
3548 					*inoffp += xfer;
3549 					*outoffp += xfer;
3550 					len -= xfer;
3551 					if (len < savlen) {
3552 						interrupted = sig_intr();
3553 						if (timespecisset(&endts) &&
3554 						    interrupted == 0) {
3555 							getnanouptime(&curts);
3556 							if (timespeccmp(&curts,
3557 							    &endts, >=))
3558 								interrupted =
3559 								    EINTR;
3560 						}
3561 					}
3562 				}
3563 			}
3564 			copylen = MIN(len, endoff - startoff);
3565 			cantseek = false;
3566 		} else {
3567 			cantseek = true;
3568 			if (!sparse)
3569 				cantseek = false;
3570 			startoff = *inoffp;
3571 			copylen = len;
3572 			error = 0;
3573 		}
3574 
3575 		xfer = blksize;
3576 		if (cantseek) {
3577 			/*
3578 			 * Set first xfer to end at a block boundary, so that
3579 			 * holes are more likely detected in the loop below via
3580 			 * the for all bytes 0 method.
3581 			 */
3582 			xfer -= (*inoffp % blksize);
3583 		}
3584 		/* Loop copying the data block. */
3585 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3586 			if (copylen < xfer)
3587 				xfer = copylen;
3588 			error = vn_lock(invp, LK_SHARED);
3589 			if (error != 0)
3590 				goto out;
3591 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3592 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3593 			    curthread->td_ucred, incred, &aresid,
3594 			    curthread);
3595 			VOP_UNLOCK(invp);
3596 			lastblock = false;
3597 			if (error == 0 && aresid > 0) {
3598 				/* Stop the copy at EOF on the input file. */
3599 				xfer -= aresid;
3600 				eof = true;
3601 				lastblock = true;
3602 			}
3603 			if (error == 0) {
3604 				/*
3605 				 * Skip the write for holes past the initial EOF
3606 				 * of the output file, unless this is the last
3607 				 * write of the output file at EOF.
3608 				 */
3609 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3610 				    false;
3611 				if (xfer == len)
3612 					lastblock = true;
3613 				if (!cantseek || *outoffp < outsize ||
3614 				    lastblock || !readzeros)
3615 					error = vn_write_outvp(outvp, dat,
3616 					    *outoffp, xfer, blksize,
3617 					    readzeros && lastblock &&
3618 					    *outoffp >= outsize, false,
3619 					    outcred);
3620 				if (error == 0) {
3621 					*inoffp += xfer;
3622 					startoff += xfer;
3623 					*outoffp += xfer;
3624 					copylen -= xfer;
3625 					len -= xfer;
3626 					if (len < savlen) {
3627 						interrupted = sig_intr();
3628 						if (timespecisset(&endts) &&
3629 						    interrupted == 0) {
3630 							getnanouptime(&curts);
3631 							if (timespeccmp(&curts,
3632 							    &endts, >=))
3633 								interrupted =
3634 								    EINTR;
3635 						}
3636 					}
3637 				}
3638 			}
3639 			xfer = blksize;
3640 		}
3641 	}
3642 out:
3643 	*lenp = savlen - len;
3644 	free(dat, M_TEMP);
3645 	return (error);
3646 }
3647 
3648 static int
3649 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3650 {
3651 	struct mount *mp;
3652 	struct vnode *vp;
3653 	off_t olen, ooffset;
3654 	int error;
3655 #ifdef AUDIT
3656 	int audited_vnode1 = 0;
3657 #endif
3658 
3659 	vp = fp->f_vnode;
3660 	if (vp->v_type != VREG)
3661 		return (ENODEV);
3662 
3663 	/* Allocating blocks may take a long time, so iterate. */
3664 	for (;;) {
3665 		olen = len;
3666 		ooffset = offset;
3667 
3668 		bwillwrite();
3669 		mp = NULL;
3670 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
3671 		if (error != 0)
3672 			break;
3673 		error = vn_lock(vp, LK_EXCLUSIVE);
3674 		if (error != 0) {
3675 			vn_finished_write(mp);
3676 			break;
3677 		}
3678 #ifdef AUDIT
3679 		if (!audited_vnode1) {
3680 			AUDIT_ARG_VNODE1(vp);
3681 			audited_vnode1 = 1;
3682 		}
3683 #endif
3684 #ifdef MAC
3685 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3686 		if (error == 0)
3687 #endif
3688 			error = VOP_ALLOCATE(vp, &offset, &len, 0,
3689 			    td->td_ucred);
3690 		VOP_UNLOCK(vp);
3691 		vn_finished_write(mp);
3692 
3693 		if (olen + ooffset != offset + len) {
3694 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3695 			    ooffset, olen, offset, len);
3696 		}
3697 		if (error != 0 || len == 0)
3698 			break;
3699 		KASSERT(olen > len, ("Iteration did not make progress?"));
3700 		maybe_yield();
3701 	}
3702 
3703 	return (error);
3704 }
3705 
3706 static int
3707 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3708     int ioflag, struct ucred *cred, struct ucred *active_cred,
3709     struct ucred *file_cred)
3710 {
3711 	struct mount *mp;
3712 	void *rl_cookie;
3713 	off_t off, len;
3714 	int error;
3715 #ifdef AUDIT
3716 	bool audited_vnode1 = false;
3717 #endif
3718 
3719 	rl_cookie = NULL;
3720 	error = 0;
3721 	mp = NULL;
3722 	off = *offset;
3723 	len = *length;
3724 
3725 	if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3726 		rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3727 	while (len > 0 && error == 0) {
3728 		/*
3729 		 * Try to deallocate the longest range in one pass.
3730 		 * In case a pass takes too long to be executed, it returns
3731 		 * partial result. The residue will be proceeded in the next
3732 		 * pass.
3733 		 */
3734 
3735 		if ((ioflag & IO_NODELOCKED) == 0) {
3736 			bwillwrite();
3737 			if ((error = vn_start_write(vp, &mp,
3738 			    V_WAIT | V_PCATCH)) != 0)
3739 				goto out;
3740 			vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3741 		}
3742 #ifdef AUDIT
3743 		if (!audited_vnode1) {
3744 			AUDIT_ARG_VNODE1(vp);
3745 			audited_vnode1 = true;
3746 		}
3747 #endif
3748 
3749 #ifdef MAC
3750 		if ((ioflag & IO_NOMACCHECK) == 0)
3751 			error = mac_vnode_check_write(active_cred, file_cred,
3752 			    vp);
3753 #endif
3754 		if (error == 0)
3755 			error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3756 			    cred);
3757 
3758 		if ((ioflag & IO_NODELOCKED) == 0) {
3759 			VOP_UNLOCK(vp);
3760 			if (mp != NULL) {
3761 				vn_finished_write(mp);
3762 				mp = NULL;
3763 			}
3764 		}
3765 		if (error == 0 && len != 0)
3766 			maybe_yield();
3767 	}
3768 out:
3769 	if (rl_cookie != NULL)
3770 		vn_rangelock_unlock(vp, rl_cookie);
3771 	*offset = off;
3772 	*length = len;
3773 	return (error);
3774 }
3775 
3776 /*
3777  * This function is supposed to be used in the situations where the deallocation
3778  * is not triggered by a user request.
3779  */
3780 int
3781 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3782     int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3783 {
3784 	struct ucred *cred;
3785 
3786 	if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3787 	    flags != 0)
3788 		return (EINVAL);
3789 	if (vp->v_type != VREG)
3790 		return (ENODEV);
3791 
3792 	cred = file_cred != NOCRED ? file_cred : active_cred;
3793 	return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3794 	    active_cred, file_cred));
3795 }
3796 
3797 static int
3798 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3799     struct ucred *active_cred, struct thread *td)
3800 {
3801 	int error;
3802 	struct vnode *vp;
3803 	int ioflag;
3804 
3805 	KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd"));
3806 	KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
3807 	    ("vn_fspacectl: non-zero flags"));
3808 	KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
3809 	    ("vn_fspacectl: offset/length overflow or underflow"));
3810 	vp = fp->f_vnode;
3811 
3812 	if (vp->v_type != VREG)
3813 		return (ENODEV);
3814 
3815 	ioflag = get_write_ioflag(fp);
3816 
3817 	switch (cmd) {
3818 	case SPACECTL_DEALLOC:
3819 		error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3820 		    active_cred, active_cred, fp->f_cred);
3821 		break;
3822 	default:
3823 		panic("vn_fspacectl: unknown cmd %d", cmd);
3824 	}
3825 
3826 	return (error);
3827 }
3828 
3829 /*
3830  * Keep this assert as long as sizeof(struct dirent) is used as the maximum
3831  * entry size.
3832  */
3833 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent),
3834     "'struct dirent' size must be a multiple of its alignment "
3835     "(see _GENERIC_DIRLEN())");
3836 
3837 /*
3838  * Returns successive directory entries through some caller's provided buffer.
3839  *
3840  * This function automatically refills the provided buffer with calls to
3841  * VOP_READDIR() (after MAC permission checks).
3842  *
3843  * 'td' is used for credentials and passed to uiomove().  'dirbuf' is the
3844  * caller's buffer to fill and 'dirbuflen' its allocated size.  'dirbuf' must
3845  * be properly aligned to access 'struct dirent' structures and 'dirbuflen'
3846  * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning
3847  * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always
3848  * be returned if this requirement is not verified).  '*dpp' points to the
3849  * current directory entry in the buffer and '*len' contains the remaining
3850  * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry).
3851  *
3852  * At first call (or when restarting the read), '*len' must have been set to 0,
3853  * '*off' to 0 (or any valid start offset) and '*eofflag' to 0.  There are no
3854  * more entries as soon as '*len' is 0 after a call that returned 0.  Calling
3855  * again this function after such a condition is considered an error and EINVAL
3856  * will be returned.  Other possible error codes are those of VOP_READDIR(),
3857  * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL
3858  * (bad call).  All errors are unrecoverable, i.e., the state ('*len', '*off'
3859  * and '*eofflag') must be re-initialized before a subsequent call.  On error
3860  * or at end of directory, '*dpp' is reset to NULL.
3861  *
3862  * '*len', '*off' and '*eofflag' are internal state the caller should not
3863  * tamper with except as explained above.  '*off' is the next directory offset
3864  * to read from to refill the buffer.  '*eofflag' is set to 0 or 1 by the last
3865  * internal call to VOP_READDIR() that returned without error, indicating
3866  * whether it reached the end of the directory, and to 2 by this function after
3867  * all entries have been read.
3868  */
3869 int
3870 vn_dir_next_dirent(struct vnode *vp, struct thread *td,
3871     char *dirbuf, size_t dirbuflen,
3872     struct dirent **dpp, size_t *len, off_t *off, int *eofflag)
3873 {
3874 	struct dirent *dp = NULL;
3875 	int reclen;
3876 	int error;
3877 	struct uio uio;
3878 	struct iovec iov;
3879 
3880 	ASSERT_VOP_LOCKED(vp, "vnode not locked");
3881 	VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory"));
3882 	MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen,
3883 	    "Address space overflow");
3884 
3885 	if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) {
3886 		/* Don't take any chances in this case */
3887 		error = EINVAL;
3888 		goto out;
3889 	}
3890 
3891 	if (*len != 0) {
3892 		dp = *dpp;
3893 
3894 		/*
3895 		 * The caller continued to call us after an error (we set dp to
3896 		 * NULL in a previous iteration).  Bail out right now.
3897 		 */
3898 		if (__predict_false(dp == NULL))
3899 			return (EINVAL);
3900 
3901 		MPASS(*len <= dirbuflen);
3902 		MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp &&
3903 		    (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen,
3904 		    "Filled range not inside buffer");
3905 
3906 		reclen = dp->d_reclen;
3907 		if (reclen >= *len) {
3908 			/* End of buffer reached */
3909 			*len = 0;
3910 		} else {
3911 			dp = (struct dirent *)((char *)dp + reclen);
3912 			*len -= reclen;
3913 		}
3914 	}
3915 
3916 	if (*len == 0) {
3917 		dp = NULL;
3918 
3919 		/* Have to refill. */
3920 		switch (*eofflag) {
3921 		case 0:
3922 			break;
3923 
3924 		case 1:
3925 			/* Nothing more to read. */
3926 			*eofflag = 2; /* Remember the caller reached EOF. */
3927 			goto success;
3928 
3929 		default:
3930 			/* The caller didn't test for EOF. */
3931 			error = EINVAL;
3932 			goto out;
3933 		}
3934 
3935 		iov.iov_base = dirbuf;
3936 		iov.iov_len = dirbuflen;
3937 
3938 		uio.uio_iov = &iov;
3939 		uio.uio_iovcnt = 1;
3940 		uio.uio_offset = *off;
3941 		uio.uio_resid = dirbuflen;
3942 		uio.uio_segflg = UIO_SYSSPACE;
3943 		uio.uio_rw = UIO_READ;
3944 		uio.uio_td = td;
3945 
3946 #ifdef MAC
3947 		error = mac_vnode_check_readdir(td->td_ucred, vp);
3948 		if (error == 0)
3949 #endif
3950 			error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag,
3951 			    NULL, NULL);
3952 		if (error != 0)
3953 			goto out;
3954 
3955 		*len = dirbuflen - uio.uio_resid;
3956 		*off = uio.uio_offset;
3957 
3958 		if (*len == 0) {
3959 			/* Sanity check on INVARIANTS. */
3960 			MPASS(*eofflag != 0);
3961 			*eofflag = 1;
3962 			goto success;
3963 		}
3964 
3965 		/*
3966 		 * Normalize the flag returned by VOP_READDIR(), since we use 2
3967 		 * as a sentinel value.
3968 		 */
3969 		if (*eofflag != 0)
3970 			*eofflag = 1;
3971 
3972 		dp = (struct dirent *)dirbuf;
3973 	}
3974 
3975 	if (__predict_false(*len < GENERIC_MINDIRSIZ ||
3976 	    dp->d_reclen < GENERIC_MINDIRSIZ)) {
3977 		error = EINTEGRITY;
3978 		dp = NULL;
3979 		goto out;
3980 	}
3981 
3982 success:
3983 	error = 0;
3984 out:
3985 	*dpp = dp;
3986 	return (error);
3987 }
3988 
3989 /*
3990  * Checks whether a directory is empty or not.
3991  *
3992  * If the directory is empty, returns 0, and if it is not, ENOTEMPTY.  Other
3993  * values are genuine errors preventing the check.
3994  */
3995 int
3996 vn_dir_check_empty(struct vnode *vp)
3997 {
3998 	struct thread *const td = curthread;
3999 	char *dirbuf;
4000 	size_t dirbuflen, len;
4001 	off_t off;
4002 	int eofflag, error;
4003 	struct dirent *dp;
4004 	struct vattr va;
4005 
4006 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
4007 	VNPASS(vp->v_type == VDIR, vp);
4008 
4009 	error = VOP_GETATTR(vp, &va, td->td_ucred);
4010 	if (error != 0)
4011 		return (error);
4012 
4013 	dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ);
4014 	if (dirbuflen < va.va_blocksize)
4015 		dirbuflen = va.va_blocksize;
4016 	dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK);
4017 
4018 	len = 0;
4019 	off = 0;
4020 	eofflag = 0;
4021 
4022 	for (;;) {
4023 		error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen,
4024 		    &dp, &len, &off, &eofflag);
4025 		if (error != 0)
4026 			goto end;
4027 
4028 		if (len == 0) {
4029 			/* EOF */
4030 			error = 0;
4031 			goto end;
4032 		}
4033 
4034 		/*
4035 		 * Skip whiteouts.  Unionfs operates on filesystems only and
4036 		 * not on hierarchies, so these whiteouts would be shadowed on
4037 		 * the system hierarchy but not for a union using the
4038 		 * filesystem of their directories as the upper layer.
4039 		 * Additionally, unionfs currently transparently exposes
4040 		 * union-specific metadata of its upper layer, meaning that
4041 		 * whiteouts can be seen through the union view in empty
4042 		 * directories.  Taking into account these whiteouts would then
4043 		 * prevent mounting another filesystem on such effectively
4044 		 * empty directories.
4045 		 */
4046 		if (dp->d_type == DT_WHT)
4047 			continue;
4048 
4049 		/*
4050 		 * Any file in the directory which is not '.' or '..' indicates
4051 		 * the directory is not empty.
4052 		 */
4053 		switch (dp->d_namlen) {
4054 		case 2:
4055 			if (dp->d_name[1] != '.') {
4056 				/* Can't be '..' (nor '.') */
4057 				error = ENOTEMPTY;
4058 				goto end;
4059 			}
4060 			/* FALLTHROUGH */
4061 		case 1:
4062 			if (dp->d_name[0] != '.') {
4063 				/* Can't be '..' nor '.' */
4064 				error = ENOTEMPTY;
4065 				goto end;
4066 			}
4067 			break;
4068 
4069 		default:
4070 			error = ENOTEMPTY;
4071 			goto end;
4072 		}
4073 	}
4074 
4075 end:
4076 	free(dirbuf, M_TEMP);
4077 	return (error);
4078 }
4079 
4080 
4081 static u_long vn_lock_pair_pause_cnt;
4082 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
4083     &vn_lock_pair_pause_cnt, 0,
4084     "Count of vn_lock_pair deadlocks");
4085 
4086 u_int vn_lock_pair_pause_max;
4087 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
4088     &vn_lock_pair_pause_max, 0,
4089     "Max ticks for vn_lock_pair deadlock avoidance sleep");
4090 
4091 static void
4092 vn_lock_pair_pause(const char *wmesg)
4093 {
4094 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
4095 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
4096 }
4097 
4098 /*
4099  * Lock pair of (possibly same) vnodes vp1, vp2, avoiding lock order
4100  * reversal.  vp1_locked indicates whether vp1 is locked; if not, vp1
4101  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
4102  * can be NULL.
4103  *
4104  * The function returns with both vnodes exclusively or shared locked,
4105  * according to corresponding lkflags, and guarantees that it does not
4106  * create lock order reversal with other threads during its execution.
4107  * Both vnodes could be unlocked temporary (and reclaimed).
4108  *
4109  * If requesting shared locking, locked vnode lock must not be recursed.
4110  *
4111  * Only one of LK_SHARED and LK_EXCLUSIVE must be specified.
4112  * LK_NODDLKTREAT can be optionally passed.
4113  *
4114  * If vp1 == vp2, only one, most exclusive, lock is obtained on it.
4115  */
4116 void
4117 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1,
4118     struct vnode *vp2, bool vp2_locked, int lkflags2)
4119 {
4120 	int error, locked1;
4121 
4122 	MPASS((((lkflags1 & LK_SHARED) != 0) ^ ((lkflags1 & LK_EXCLUSIVE) != 0)) ||
4123 	    (vp1 == NULL && lkflags1 == 0));
4124 	MPASS((lkflags1 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
4125 	MPASS((((lkflags2 & LK_SHARED) != 0) ^ ((lkflags2 & LK_EXCLUSIVE) != 0)) ||
4126 	    (vp2 == NULL && lkflags2 == 0));
4127 	MPASS((lkflags2 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
4128 
4129 	if (vp1 == NULL && vp2 == NULL)
4130 		return;
4131 
4132 	if (vp1 == vp2) {
4133 		MPASS(vp1_locked == vp2_locked);
4134 
4135 		/* Select the most exclusive mode for lock. */
4136 		if ((lkflags1 & LK_TYPE_MASK) != (lkflags2 & LK_TYPE_MASK))
4137 			lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
4138 
4139 		if (vp1_locked) {
4140 			ASSERT_VOP_LOCKED(vp1, "vp1");
4141 
4142 			/* No need to relock if any lock is exclusive. */
4143 			if ((vp1->v_vnlock->lock_object.lo_flags &
4144 			    LK_NOSHARE) != 0)
4145 				return;
4146 
4147 			locked1 = VOP_ISLOCKED(vp1);
4148 			if (((lkflags1 & LK_SHARED) != 0 &&
4149 			    locked1 != LK_EXCLUSIVE) ||
4150 			    ((lkflags1 & LK_EXCLUSIVE) != 0 &&
4151 			    locked1 == LK_EXCLUSIVE))
4152 				return;
4153 			VOP_UNLOCK(vp1);
4154 		}
4155 
4156 		ASSERT_VOP_UNLOCKED(vp1, "vp1");
4157 		vn_lock(vp1, lkflags1 | LK_RETRY);
4158 		return;
4159 	}
4160 
4161 	if (vp1 != NULL) {
4162 		if ((lkflags1 & LK_SHARED) != 0 &&
4163 		    (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
4164 			lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
4165 		if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) {
4166 			ASSERT_VOP_LOCKED(vp1, "vp1");
4167 			if ((lkflags1 & LK_EXCLUSIVE) != 0) {
4168 				VOP_UNLOCK(vp1);
4169 				ASSERT_VOP_UNLOCKED(vp1,
4170 				    "vp1 shared recursed");
4171 				vp1_locked = false;
4172 			}
4173 		} else if (!vp1_locked)
4174 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
4175 	} else {
4176 		vp1_locked = true;
4177 	}
4178 
4179 	if (vp2 != NULL) {
4180 		if ((lkflags2 & LK_SHARED) != 0 &&
4181 		    (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
4182 			lkflags2 = (lkflags2 & ~LK_SHARED) | LK_EXCLUSIVE;
4183 		if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) {
4184 			ASSERT_VOP_LOCKED(vp2, "vp2");
4185 			if ((lkflags2 & LK_EXCLUSIVE) != 0) {
4186 				VOP_UNLOCK(vp2);
4187 				ASSERT_VOP_UNLOCKED(vp2,
4188 				    "vp2 shared recursed");
4189 				vp2_locked = false;
4190 			}
4191 		} else if (!vp2_locked)
4192 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
4193 	} else {
4194 		vp2_locked = true;
4195 	}
4196 
4197 	if (!vp1_locked && !vp2_locked) {
4198 		vn_lock(vp1, lkflags1 | LK_RETRY);
4199 		vp1_locked = true;
4200 	}
4201 
4202 	while (!vp1_locked || !vp2_locked) {
4203 		if (vp1_locked && vp2 != NULL) {
4204 			if (vp1 != NULL) {
4205 				error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT,
4206 				    __FILE__, __LINE__);
4207 				if (error == 0)
4208 					break;
4209 				VOP_UNLOCK(vp1);
4210 				vp1_locked = false;
4211 				vn_lock_pair_pause("vlp1");
4212 			}
4213 			vn_lock(vp2, lkflags2 | LK_RETRY);
4214 			vp2_locked = true;
4215 		}
4216 		if (vp2_locked && vp1 != NULL) {
4217 			if (vp2 != NULL) {
4218 				error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT,
4219 				    __FILE__, __LINE__);
4220 				if (error == 0)
4221 					break;
4222 				VOP_UNLOCK(vp2);
4223 				vp2_locked = false;
4224 				vn_lock_pair_pause("vlp2");
4225 			}
4226 			vn_lock(vp1, lkflags1 | LK_RETRY);
4227 			vp1_locked = true;
4228 		}
4229 	}
4230 	if (vp1 != NULL) {
4231 		if (lkflags1 == LK_EXCLUSIVE)
4232 			ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
4233 		else
4234 			ASSERT_VOP_LOCKED(vp1, "vp1 ret");
4235 	}
4236 	if (vp2 != NULL) {
4237 		if (lkflags2 == LK_EXCLUSIVE)
4238 			ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
4239 		else
4240 			ASSERT_VOP_LOCKED(vp2, "vp2 ret");
4241 	}
4242 }
4243 
4244 int
4245 vn_lktype_write(struct mount *mp, struct vnode *vp)
4246 {
4247 	if (MNT_SHARED_WRITES(mp) ||
4248 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
4249 		return (LK_SHARED);
4250 	return (LK_EXCLUSIVE);
4251 }
4252 
4253 int
4254 vn_cmp(struct file *fp1, struct file *fp2, struct thread *td)
4255 {
4256 	if (fp2->f_type != DTYPE_VNODE)
4257 		return (3);
4258 	return (kcmp_cmp((uintptr_t)fp1->f_vnode, (uintptr_t)fp2->f_vnode));
4259 }
4260