xref: /freebsd/sys/kern/vfs_vnops.c (revision d9a42747950146bf03cda7f6e25d219253f8a57a)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/prng.h>
74 #include <sys/sx.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
78 #include <sys/conf.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
81 #include <sys/user.h>
82 #include <sys/ktrace.h>
83 
84 #include <security/audit/audit.h>
85 #include <security/mac/mac_framework.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_extern.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pager.h>
94 
95 #ifdef HWPMC_HOOKS
96 #include <sys/pmckern.h>
97 #endif
98 
99 static fo_rdwr_t	vn_read;
100 static fo_rdwr_t	vn_write;
101 static fo_rdwr_t	vn_io_fault;
102 static fo_truncate_t	vn_truncate;
103 static fo_ioctl_t	vn_ioctl;
104 static fo_poll_t	vn_poll;
105 static fo_kqfilter_t	vn_kqfilter;
106 static fo_close_t	vn_closefile;
107 static fo_mmap_t	vn_mmap;
108 static fo_fallocate_t	vn_fallocate;
109 static fo_fspacectl_t	vn_fspacectl;
110 
111 struct 	fileops vnops = {
112 	.fo_read = vn_io_fault,
113 	.fo_write = vn_io_fault,
114 	.fo_truncate = vn_truncate,
115 	.fo_ioctl = vn_ioctl,
116 	.fo_poll = vn_poll,
117 	.fo_kqfilter = vn_kqfilter,
118 	.fo_stat = vn_statfile,
119 	.fo_close = vn_closefile,
120 	.fo_chmod = vn_chmod,
121 	.fo_chown = vn_chown,
122 	.fo_sendfile = vn_sendfile,
123 	.fo_seek = vn_seek,
124 	.fo_fill_kinfo = vn_fill_kinfo,
125 	.fo_mmap = vn_mmap,
126 	.fo_fallocate = vn_fallocate,
127 	.fo_fspacectl = vn_fspacectl,
128 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
129 };
130 
131 const u_int io_hold_cnt = 16;
132 static int vn_io_fault_enable = 1;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
134     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
135 static int vn_io_fault_prefault = 0;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
137     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
138 static int vn_io_pgcache_read_enable = 1;
139 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
140     &vn_io_pgcache_read_enable, 0,
141     "Enable copying from page cache for reads, avoiding fs");
142 static u_long vn_io_faults_cnt;
143 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
144     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
145 
146 static int vfs_allow_read_dir = 0;
147 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
148     &vfs_allow_read_dir, 0,
149     "Enable read(2) of directory by root for filesystems that support it");
150 
151 /*
152  * Returns true if vn_io_fault mode of handling the i/o request should
153  * be used.
154  */
155 static bool
156 do_vn_io_fault(struct vnode *vp, struct uio *uio)
157 {
158 	struct mount *mp;
159 
160 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
161 	    (mp = vp->v_mount) != NULL &&
162 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
163 }
164 
165 /*
166  * Structure used to pass arguments to vn_io_fault1(), to do either
167  * file- or vnode-based I/O calls.
168  */
169 struct vn_io_fault_args {
170 	enum {
171 		VN_IO_FAULT_FOP,
172 		VN_IO_FAULT_VOP
173 	} kind;
174 	struct ucred *cred;
175 	int flags;
176 	union {
177 		struct fop_args_tag {
178 			struct file *fp;
179 			fo_rdwr_t *doio;
180 		} fop_args;
181 		struct vop_args_tag {
182 			struct vnode *vp;
183 		} vop_args;
184 	} args;
185 };
186 
187 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
188     struct vn_io_fault_args *args, struct thread *td);
189 
190 int
191 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
192 {
193 	struct thread *td = curthread;
194 
195 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
196 }
197 
198 static uint64_t
199 open2nameif(int fmode, u_int vn_open_flags)
200 {
201 	uint64_t res;
202 
203 	res = ISOPEN | LOCKLEAF;
204 	if ((fmode & O_RESOLVE_BENEATH) != 0)
205 		res |= RBENEATH;
206 	if ((fmode & O_EMPTY_PATH) != 0)
207 		res |= EMPTYPATH;
208 	if ((fmode & FREAD) != 0)
209 		res |= OPENREAD;
210 	if ((fmode & FWRITE) != 0)
211 		res |= OPENWRITE;
212 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
213 		res |= AUDITVNODE1;
214 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
215 		res |= NOCAPCHECK;
216 	if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
217 		res |= WANTIOCTLCAPS;
218 	return (res);
219 }
220 
221 /*
222  * Common code for vnode open operations via a name lookup.
223  * Lookup the vnode and invoke VOP_CREATE if needed.
224  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
225  *
226  * Note that this does NOT free nameidata for the successful case,
227  * due to the NDINIT being done elsewhere.
228  */
229 int
230 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
231     struct ucred *cred, struct file *fp)
232 {
233 	struct vnode *vp;
234 	struct mount *mp;
235 	struct vattr vat;
236 	struct vattr *vap = &vat;
237 	int fmode, error;
238 	bool first_open;
239 
240 restart:
241 	first_open = false;
242 	fmode = *flagp;
243 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
244 	    O_EXCL | O_DIRECTORY) ||
245 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
246 		return (EINVAL);
247 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
248 		ndp->ni_cnd.cn_nameiop = CREATE;
249 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
250 		/*
251 		 * Set NOCACHE to avoid flushing the cache when
252 		 * rolling in many files at once.
253 		 *
254 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
255 		 * exist despite NOCACHE.
256 		 */
257 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
258 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
259 			ndp->ni_cnd.cn_flags |= FOLLOW;
260 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
261 			bwillwrite();
262 		if ((error = namei(ndp)) != 0)
263 			return (error);
264 		if (ndp->ni_vp == NULL) {
265 			VATTR_NULL(vap);
266 			vap->va_type = VREG;
267 			vap->va_mode = cmode;
268 			if (fmode & O_EXCL)
269 				vap->va_vaflags |= VA_EXCLUSIVE;
270 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
271 				NDFREE_PNBUF(ndp);
272 				vput(ndp->ni_dvp);
273 				if ((error = vn_start_write(NULL, &mp,
274 				    V_XSLEEP | V_PCATCH)) != 0)
275 					return (error);
276 				NDREINIT(ndp);
277 				goto restart;
278 			}
279 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
280 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
281 #ifdef MAC
282 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
283 			    &ndp->ni_cnd, vap);
284 			if (error == 0)
285 #endif
286 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
287 				    &ndp->ni_cnd, vap);
288 			vp = ndp->ni_vp;
289 			if (error == 0 && (fmode & O_EXCL) != 0 &&
290 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
291 				VI_LOCK(vp);
292 				vp->v_iflag |= VI_FOPENING;
293 				VI_UNLOCK(vp);
294 				first_open = true;
295 			}
296 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
297 			    false);
298 			vn_finished_write(mp);
299 			if (error) {
300 				NDFREE_PNBUF(ndp);
301 				if (error == ERELOOKUP) {
302 					NDREINIT(ndp);
303 					goto restart;
304 				}
305 				return (error);
306 			}
307 			fmode &= ~O_TRUNC;
308 		} else {
309 			if (ndp->ni_dvp == ndp->ni_vp)
310 				vrele(ndp->ni_dvp);
311 			else
312 				vput(ndp->ni_dvp);
313 			ndp->ni_dvp = NULL;
314 			vp = ndp->ni_vp;
315 			if (fmode & O_EXCL) {
316 				error = EEXIST;
317 				goto bad;
318 			}
319 			if (vp->v_type == VDIR) {
320 				error = EISDIR;
321 				goto bad;
322 			}
323 			fmode &= ~O_CREAT;
324 		}
325 	} else {
326 		ndp->ni_cnd.cn_nameiop = LOOKUP;
327 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
328 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
329 		    FOLLOW;
330 		if ((fmode & FWRITE) == 0)
331 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
332 		if ((error = namei(ndp)) != 0)
333 			return (error);
334 		vp = ndp->ni_vp;
335 	}
336 	error = vn_open_vnode(vp, fmode, cred, curthread, fp);
337 	if (first_open) {
338 		VI_LOCK(vp);
339 		vp->v_iflag &= ~VI_FOPENING;
340 		wakeup(vp);
341 		VI_UNLOCK(vp);
342 	}
343 	if (error)
344 		goto bad;
345 	*flagp = fmode;
346 	return (0);
347 bad:
348 	NDFREE_PNBUF(ndp);
349 	vput(vp);
350 	*flagp = fmode;
351 	ndp->ni_vp = NULL;
352 	return (error);
353 }
354 
355 static int
356 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
357 {
358 	struct flock lf;
359 	int error, lock_flags, type;
360 
361 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
362 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
363 		return (0);
364 	KASSERT(fp != NULL, ("open with flock requires fp"));
365 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
366 		return (EOPNOTSUPP);
367 
368 	lock_flags = VOP_ISLOCKED(vp);
369 	VOP_UNLOCK(vp);
370 
371 	lf.l_whence = SEEK_SET;
372 	lf.l_start = 0;
373 	lf.l_len = 0;
374 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
375 	type = F_FLOCK;
376 	if ((fmode & FNONBLOCK) == 0)
377 		type |= F_WAIT;
378 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
379 		type |= F_FIRSTOPEN;
380 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
381 	if (error == 0)
382 		fp->f_flag |= FHASLOCK;
383 
384 	vn_lock(vp, lock_flags | LK_RETRY);
385 	return (error);
386 }
387 
388 /*
389  * Common code for vnode open operations once a vnode is located.
390  * Check permissions, and call the VOP_OPEN routine.
391  */
392 int
393 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
394     struct thread *td, struct file *fp)
395 {
396 	accmode_t accmode;
397 	int error;
398 
399 	if (vp->v_type == VLNK) {
400 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
401 			return (EMLINK);
402 	}
403 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
404 		return (ENOTDIR);
405 
406 	accmode = 0;
407 	if ((fmode & O_PATH) == 0) {
408 		if (vp->v_type == VSOCK)
409 			return (EOPNOTSUPP);
410 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
411 			if (vp->v_type == VDIR)
412 				return (EISDIR);
413 			accmode |= VWRITE;
414 		}
415 		if ((fmode & FREAD) != 0)
416 			accmode |= VREAD;
417 		if ((fmode & O_APPEND) && (fmode & FWRITE))
418 			accmode |= VAPPEND;
419 #ifdef MAC
420 		if ((fmode & O_CREAT) != 0)
421 			accmode |= VCREAT;
422 #endif
423 	}
424 	if ((fmode & FEXEC) != 0)
425 		accmode |= VEXEC;
426 #ifdef MAC
427 	if ((fmode & O_VERIFY) != 0)
428 		accmode |= VVERIFY;
429 	error = mac_vnode_check_open(cred, vp, accmode);
430 	if (error != 0)
431 		return (error);
432 
433 	accmode &= ~(VCREAT | VVERIFY);
434 #endif
435 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
436 		error = VOP_ACCESS(vp, accmode, cred, td);
437 		if (error != 0)
438 			return (error);
439 	}
440 	if ((fmode & O_PATH) != 0) {
441 		if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
442 		    VOP_ACCESS(vp, VREAD, cred, td) == 0)
443 			fp->f_flag |= FKQALLOWED;
444 		return (0);
445 	}
446 
447 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
448 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
449 	error = VOP_OPEN(vp, fmode, cred, td, fp);
450 	if (error != 0)
451 		return (error);
452 
453 	error = vn_open_vnode_advlock(vp, fmode, fp);
454 	if (error == 0 && (fmode & FWRITE) != 0) {
455 		error = VOP_ADD_WRITECOUNT(vp, 1);
456 		if (error == 0) {
457 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
458 			     __func__, vp, vp->v_writecount);
459 		}
460 	}
461 
462 	/*
463 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
464 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
465 	 */
466 	if (error != 0) {
467 		if (fp != NULL) {
468 			/*
469 			 * Arrange the call by having fdrop() to use
470 			 * vn_closefile().  This is to satisfy
471 			 * filesystems like devfs or tmpfs, which
472 			 * override fo_close().
473 			 */
474 			fp->f_flag |= FOPENFAILED;
475 			fp->f_vnode = vp;
476 			if (fp->f_ops == &badfileops) {
477 				fp->f_type = DTYPE_VNODE;
478 				fp->f_ops = &vnops;
479 			}
480 			vref(vp);
481 		} else {
482 			/*
483 			 * If there is no fp, due to kernel-mode open,
484 			 * we can call VOP_CLOSE() now.
485 			 */
486 			if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
487 			    !MNT_EXTENDED_SHARED(vp->v_mount) &&
488 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
489 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
490 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
491 			    cred, td);
492 		}
493 	}
494 
495 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
496 	return (error);
497 
498 }
499 
500 /*
501  * Check for write permissions on the specified vnode.
502  * Prototype text segments cannot be written.
503  * It is racy.
504  */
505 int
506 vn_writechk(struct vnode *vp)
507 {
508 
509 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
510 	/*
511 	 * If there's shared text associated with
512 	 * the vnode, try to free it up once.  If
513 	 * we fail, we can't allow writing.
514 	 */
515 	if (VOP_IS_TEXT(vp))
516 		return (ETXTBSY);
517 
518 	return (0);
519 }
520 
521 /*
522  * Vnode close call
523  */
524 static int
525 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
526     struct thread *td, bool keep_ref)
527 {
528 	struct mount *mp;
529 	int error, lock_flags;
530 
531 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
532 	    MNT_EXTENDED_SHARED(vp->v_mount))
533 		lock_flags = LK_SHARED;
534 	else
535 		lock_flags = LK_EXCLUSIVE;
536 
537 	vn_start_write(vp, &mp, V_WAIT);
538 	vn_lock(vp, lock_flags | LK_RETRY);
539 	AUDIT_ARG_VNODE1(vp);
540 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
541 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
542 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
543 		    __func__, vp, vp->v_writecount);
544 	}
545 	error = VOP_CLOSE(vp, flags, file_cred, td);
546 	if (keep_ref)
547 		VOP_UNLOCK(vp);
548 	else
549 		vput(vp);
550 	vn_finished_write(mp);
551 	return (error);
552 }
553 
554 int
555 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
556     struct thread *td)
557 {
558 
559 	return (vn_close1(vp, flags, file_cred, td, false));
560 }
561 
562 /*
563  * Heuristic to detect sequential operation.
564  */
565 static int
566 sequential_heuristic(struct uio *uio, struct file *fp)
567 {
568 	enum uio_rw rw;
569 
570 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
571 
572 	rw = uio->uio_rw;
573 	if (fp->f_flag & FRDAHEAD)
574 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
575 
576 	/*
577 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
578 	 * that the first I/O is normally considered to be slightly
579 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
580 	 * unless previous seeks have reduced f_seqcount to 0, in which
581 	 * case offset 0 is not special.
582 	 */
583 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
584 	    uio->uio_offset == fp->f_nextoff[rw]) {
585 		/*
586 		 * f_seqcount is in units of fixed-size blocks so that it
587 		 * depends mainly on the amount of sequential I/O and not
588 		 * much on the number of sequential I/O's.  The fixed size
589 		 * of 16384 is hard-coded here since it is (not quite) just
590 		 * a magic size that works well here.  This size is more
591 		 * closely related to the best I/O size for real disks than
592 		 * to any block size used by software.
593 		 */
594 		if (uio->uio_resid >= IO_SEQMAX * 16384)
595 			fp->f_seqcount[rw] = IO_SEQMAX;
596 		else {
597 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
598 			if (fp->f_seqcount[rw] > IO_SEQMAX)
599 				fp->f_seqcount[rw] = IO_SEQMAX;
600 		}
601 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
602 	}
603 
604 	/* Not sequential.  Quickly draw-down sequentiality. */
605 	if (fp->f_seqcount[rw] > 1)
606 		fp->f_seqcount[rw] = 1;
607 	else
608 		fp->f_seqcount[rw] = 0;
609 	return (0);
610 }
611 
612 /*
613  * Package up an I/O request on a vnode into a uio and do it.
614  */
615 int
616 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
617     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
618     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
619 {
620 	struct uio auio;
621 	struct iovec aiov;
622 	struct mount *mp;
623 	struct ucred *cred;
624 	void *rl_cookie;
625 	struct vn_io_fault_args args;
626 	int error, lock_flags;
627 
628 	if (offset < 0 && vp->v_type != VCHR)
629 		return (EINVAL);
630 	auio.uio_iov = &aiov;
631 	auio.uio_iovcnt = 1;
632 	aiov.iov_base = base;
633 	aiov.iov_len = len;
634 	auio.uio_resid = len;
635 	auio.uio_offset = offset;
636 	auio.uio_segflg = segflg;
637 	auio.uio_rw = rw;
638 	auio.uio_td = td;
639 	error = 0;
640 
641 	if ((ioflg & IO_NODELOCKED) == 0) {
642 		if ((ioflg & IO_RANGELOCKED) == 0) {
643 			if (rw == UIO_READ) {
644 				rl_cookie = vn_rangelock_rlock(vp, offset,
645 				    offset + len);
646 			} else if ((ioflg & IO_APPEND) != 0) {
647 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
648 			} else {
649 				rl_cookie = vn_rangelock_wlock(vp, offset,
650 				    offset + len);
651 			}
652 		} else
653 			rl_cookie = NULL;
654 		mp = NULL;
655 		if (rw == UIO_WRITE) {
656 			if (vp->v_type != VCHR &&
657 			    (error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH))
658 			    != 0)
659 				goto out;
660 			lock_flags = vn_lktype_write(mp, vp);
661 		} else
662 			lock_flags = LK_SHARED;
663 		vn_lock(vp, lock_flags | LK_RETRY);
664 	} else
665 		rl_cookie = NULL;
666 
667 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
668 #ifdef MAC
669 	if ((ioflg & IO_NOMACCHECK) == 0) {
670 		if (rw == UIO_READ)
671 			error = mac_vnode_check_read(active_cred, file_cred,
672 			    vp);
673 		else
674 			error = mac_vnode_check_write(active_cred, file_cred,
675 			    vp);
676 	}
677 #endif
678 	if (error == 0) {
679 		if (file_cred != NULL)
680 			cred = file_cred;
681 		else
682 			cred = active_cred;
683 		if (do_vn_io_fault(vp, &auio)) {
684 			args.kind = VN_IO_FAULT_VOP;
685 			args.cred = cred;
686 			args.flags = ioflg;
687 			args.args.vop_args.vp = vp;
688 			error = vn_io_fault1(vp, &auio, &args, td);
689 		} else if (rw == UIO_READ) {
690 			error = VOP_READ(vp, &auio, ioflg, cred);
691 		} else /* if (rw == UIO_WRITE) */ {
692 			error = VOP_WRITE(vp, &auio, ioflg, cred);
693 		}
694 	}
695 	if (aresid)
696 		*aresid = auio.uio_resid;
697 	else
698 		if (auio.uio_resid && error == 0)
699 			error = EIO;
700 	if ((ioflg & IO_NODELOCKED) == 0) {
701 		VOP_UNLOCK(vp);
702 		if (mp != NULL)
703 			vn_finished_write(mp);
704 	}
705  out:
706 	if (rl_cookie != NULL)
707 		vn_rangelock_unlock(vp, rl_cookie);
708 	return (error);
709 }
710 
711 /*
712  * Package up an I/O request on a vnode into a uio and do it.  The I/O
713  * request is split up into smaller chunks and we try to avoid saturating
714  * the buffer cache while potentially holding a vnode locked, so we
715  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
716  * to give other processes a chance to lock the vnode (either other processes
717  * core'ing the same binary, or unrelated processes scanning the directory).
718  */
719 int
720 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
721     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
722     struct ucred *file_cred, size_t *aresid, struct thread *td)
723 {
724 	int error = 0;
725 	ssize_t iaresid;
726 
727 	do {
728 		int chunk;
729 
730 		/*
731 		 * Force `offset' to a multiple of MAXBSIZE except possibly
732 		 * for the first chunk, so that filesystems only need to
733 		 * write full blocks except possibly for the first and last
734 		 * chunks.
735 		 */
736 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
737 
738 		if (chunk > len)
739 			chunk = len;
740 		if (rw != UIO_READ && vp->v_type == VREG)
741 			bwillwrite();
742 		iaresid = 0;
743 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
744 		    ioflg, active_cred, file_cred, &iaresid, td);
745 		len -= chunk;	/* aresid calc already includes length */
746 		if (error)
747 			break;
748 		offset += chunk;
749 		base = (char *)base + chunk;
750 		kern_yield(PRI_USER);
751 	} while (len);
752 	if (aresid)
753 		*aresid = len + iaresid;
754 	return (error);
755 }
756 
757 #if OFF_MAX <= LONG_MAX
758 off_t
759 foffset_lock(struct file *fp, int flags)
760 {
761 	volatile short *flagsp;
762 	off_t res;
763 	short state;
764 
765 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
766 
767 	if ((flags & FOF_NOLOCK) != 0)
768 		return (atomic_load_long(&fp->f_offset));
769 
770 	/*
771 	 * According to McKusick the vn lock was protecting f_offset here.
772 	 * It is now protected by the FOFFSET_LOCKED flag.
773 	 */
774 	flagsp = &fp->f_vnread_flags;
775 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
776 		return (atomic_load_long(&fp->f_offset));
777 
778 	sleepq_lock(&fp->f_vnread_flags);
779 	state = atomic_load_16(flagsp);
780 	for (;;) {
781 		if ((state & FOFFSET_LOCKED) == 0) {
782 			if (!atomic_fcmpset_acq_16(flagsp, &state,
783 			    FOFFSET_LOCKED))
784 				continue;
785 			break;
786 		}
787 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
788 			if (!atomic_fcmpset_acq_16(flagsp, &state,
789 			    state | FOFFSET_LOCK_WAITING))
790 				continue;
791 		}
792 		DROP_GIANT();
793 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
794 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
795 		PICKUP_GIANT();
796 		sleepq_lock(&fp->f_vnread_flags);
797 		state = atomic_load_16(flagsp);
798 	}
799 	res = atomic_load_long(&fp->f_offset);
800 	sleepq_release(&fp->f_vnread_flags);
801 	return (res);
802 }
803 
804 void
805 foffset_unlock(struct file *fp, off_t val, int flags)
806 {
807 	volatile short *flagsp;
808 	short state;
809 
810 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
811 
812 	if ((flags & FOF_NOUPDATE) == 0)
813 		atomic_store_long(&fp->f_offset, val);
814 	if ((flags & FOF_NEXTOFF_R) != 0)
815 		fp->f_nextoff[UIO_READ] = val;
816 	if ((flags & FOF_NEXTOFF_W) != 0)
817 		fp->f_nextoff[UIO_WRITE] = val;
818 
819 	if ((flags & FOF_NOLOCK) != 0)
820 		return;
821 
822 	flagsp = &fp->f_vnread_flags;
823 	state = atomic_load_16(flagsp);
824 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
825 	    atomic_cmpset_rel_16(flagsp, state, 0))
826 		return;
827 
828 	sleepq_lock(&fp->f_vnread_flags);
829 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
830 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
831 	fp->f_vnread_flags = 0;
832 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
833 	sleepq_release(&fp->f_vnread_flags);
834 }
835 #else
836 off_t
837 foffset_lock(struct file *fp, int flags)
838 {
839 	struct mtx *mtxp;
840 	off_t res;
841 
842 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
843 
844 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
845 	mtx_lock(mtxp);
846 	if ((flags & FOF_NOLOCK) == 0) {
847 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
848 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
849 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
850 			    "vofflock", 0);
851 		}
852 		fp->f_vnread_flags |= FOFFSET_LOCKED;
853 	}
854 	res = fp->f_offset;
855 	mtx_unlock(mtxp);
856 	return (res);
857 }
858 
859 void
860 foffset_unlock(struct file *fp, off_t val, int flags)
861 {
862 	struct mtx *mtxp;
863 
864 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
865 
866 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
867 	mtx_lock(mtxp);
868 	if ((flags & FOF_NOUPDATE) == 0)
869 		fp->f_offset = val;
870 	if ((flags & FOF_NEXTOFF_R) != 0)
871 		fp->f_nextoff[UIO_READ] = val;
872 	if ((flags & FOF_NEXTOFF_W) != 0)
873 		fp->f_nextoff[UIO_WRITE] = val;
874 	if ((flags & FOF_NOLOCK) == 0) {
875 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
876 		    ("Lost FOFFSET_LOCKED"));
877 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
878 			wakeup(&fp->f_vnread_flags);
879 		fp->f_vnread_flags = 0;
880 	}
881 	mtx_unlock(mtxp);
882 }
883 #endif
884 
885 void
886 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
887 {
888 
889 	if ((flags & FOF_OFFSET) == 0)
890 		uio->uio_offset = foffset_lock(fp, flags);
891 }
892 
893 void
894 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
895 {
896 
897 	if ((flags & FOF_OFFSET) == 0)
898 		foffset_unlock(fp, uio->uio_offset, flags);
899 }
900 
901 static int
902 get_advice(struct file *fp, struct uio *uio)
903 {
904 	struct mtx *mtxp;
905 	int ret;
906 
907 	ret = POSIX_FADV_NORMAL;
908 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
909 		return (ret);
910 
911 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
912 	mtx_lock(mtxp);
913 	if (fp->f_advice != NULL &&
914 	    uio->uio_offset >= fp->f_advice->fa_start &&
915 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
916 		ret = fp->f_advice->fa_advice;
917 	mtx_unlock(mtxp);
918 	return (ret);
919 }
920 
921 static int
922 get_write_ioflag(struct file *fp)
923 {
924 	int ioflag;
925 	struct mount *mp;
926 	struct vnode *vp;
927 
928 	ioflag = 0;
929 	vp = fp->f_vnode;
930 	mp = atomic_load_ptr(&vp->v_mount);
931 
932 	if ((fp->f_flag & O_DIRECT) != 0)
933 		ioflag |= IO_DIRECT;
934 
935 	if ((fp->f_flag & O_FSYNC) != 0 ||
936 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
937 		ioflag |= IO_SYNC;
938 
939 	/*
940 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
941 	 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
942 	 * fall back to full O_SYNC behavior.
943 	 */
944 	if ((fp->f_flag & O_DSYNC) != 0)
945 		ioflag |= IO_SYNC | IO_DATASYNC;
946 
947 	return (ioflag);
948 }
949 
950 int
951 vn_read_from_obj(struct vnode *vp, struct uio *uio)
952 {
953 	vm_object_t obj;
954 	vm_page_t ma[io_hold_cnt + 2];
955 	off_t off, vsz;
956 	ssize_t resid;
957 	int error, i, j;
958 
959 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
960 	obj = atomic_load_ptr(&vp->v_object);
961 	if (obj == NULL)
962 		return (EJUSTRETURN);
963 
964 	/*
965 	 * Depends on type stability of vm_objects.
966 	 */
967 	vm_object_pip_add(obj, 1);
968 	if ((obj->flags & OBJ_DEAD) != 0) {
969 		/*
970 		 * Note that object might be already reused from the
971 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
972 		 * we recheck for DOOMED vnode state after all pages
973 		 * are busied, and retract then.
974 		 *
975 		 * But we check for OBJ_DEAD to ensure that we do not
976 		 * busy pages while vm_object_terminate_pages()
977 		 * processes the queue.
978 		 */
979 		error = EJUSTRETURN;
980 		goto out_pip;
981 	}
982 
983 	resid = uio->uio_resid;
984 	off = uio->uio_offset;
985 	for (i = 0; resid > 0; i++) {
986 		MPASS(i < io_hold_cnt + 2);
987 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
988 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
989 		    VM_ALLOC_NOWAIT);
990 		if (ma[i] == NULL)
991 			break;
992 
993 		/*
994 		 * Skip invalid pages.  Valid mask can be partial only
995 		 * at EOF, and we clip later.
996 		 */
997 		if (vm_page_none_valid(ma[i])) {
998 			vm_page_sunbusy(ma[i]);
999 			break;
1000 		}
1001 
1002 		resid -= PAGE_SIZE;
1003 		off += PAGE_SIZE;
1004 	}
1005 	if (i == 0) {
1006 		error = EJUSTRETURN;
1007 		goto out_pip;
1008 	}
1009 
1010 	/*
1011 	 * Check VIRF_DOOMED after we busied our pages.  Since
1012 	 * vgonel() terminates the vnode' vm_object, it cannot
1013 	 * process past pages busied by us.
1014 	 */
1015 	if (VN_IS_DOOMED(vp)) {
1016 		error = EJUSTRETURN;
1017 		goto out;
1018 	}
1019 
1020 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1021 	if (resid > uio->uio_resid)
1022 		resid = uio->uio_resid;
1023 
1024 	/*
1025 	 * Unlocked read of vnp_size is safe because truncation cannot
1026 	 * pass busied page.  But we load vnp_size into a local
1027 	 * variable so that possible concurrent extension does not
1028 	 * break calculation.
1029 	 */
1030 #if defined(__powerpc__) && !defined(__powerpc64__)
1031 	vsz = obj->un_pager.vnp.vnp_size;
1032 #else
1033 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1034 #endif
1035 	if (uio->uio_offset >= vsz) {
1036 		error = EJUSTRETURN;
1037 		goto out;
1038 	}
1039 	if (uio->uio_offset + resid > vsz)
1040 		resid = vsz - uio->uio_offset;
1041 
1042 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1043 
1044 out:
1045 	for (j = 0; j < i; j++) {
1046 		if (error == 0)
1047 			vm_page_reference(ma[j]);
1048 		vm_page_sunbusy(ma[j]);
1049 	}
1050 out_pip:
1051 	vm_object_pip_wakeup(obj);
1052 	if (error != 0)
1053 		return (error);
1054 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1055 }
1056 
1057 /*
1058  * File table vnode read routine.
1059  */
1060 static int
1061 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1062     struct thread *td)
1063 {
1064 	struct vnode *vp;
1065 	off_t orig_offset;
1066 	int error, ioflag;
1067 	int advice;
1068 
1069 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1070 	    uio->uio_td, td));
1071 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1072 	vp = fp->f_vnode;
1073 	ioflag = 0;
1074 	if (fp->f_flag & FNONBLOCK)
1075 		ioflag |= IO_NDELAY;
1076 	if (fp->f_flag & O_DIRECT)
1077 		ioflag |= IO_DIRECT;
1078 
1079 	/*
1080 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1081 	 * allows us to avoid unneeded work outright.
1082 	 */
1083 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1084 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1085 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1086 		if (error == 0) {
1087 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1088 			return (0);
1089 		}
1090 		if (error != EJUSTRETURN)
1091 			return (error);
1092 	}
1093 
1094 	advice = get_advice(fp, uio);
1095 	vn_lock(vp, LK_SHARED | LK_RETRY);
1096 
1097 	switch (advice) {
1098 	case POSIX_FADV_NORMAL:
1099 	case POSIX_FADV_SEQUENTIAL:
1100 	case POSIX_FADV_NOREUSE:
1101 		ioflag |= sequential_heuristic(uio, fp);
1102 		break;
1103 	case POSIX_FADV_RANDOM:
1104 		/* Disable read-ahead for random I/O. */
1105 		break;
1106 	}
1107 	orig_offset = uio->uio_offset;
1108 
1109 #ifdef MAC
1110 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1111 	if (error == 0)
1112 #endif
1113 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1114 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1115 	VOP_UNLOCK(vp);
1116 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1117 	    orig_offset != uio->uio_offset)
1118 		/*
1119 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1120 		 * for the backing file after a POSIX_FADV_NOREUSE
1121 		 * read(2).
1122 		 */
1123 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1124 		    POSIX_FADV_DONTNEED);
1125 	return (error);
1126 }
1127 
1128 /*
1129  * File table vnode write routine.
1130  */
1131 static int
1132 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1133     struct thread *td)
1134 {
1135 	struct vnode *vp;
1136 	struct mount *mp;
1137 	off_t orig_offset;
1138 	int error, ioflag;
1139 	int advice;
1140 	bool need_finished_write;
1141 
1142 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1143 	    uio->uio_td, td));
1144 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1145 	vp = fp->f_vnode;
1146 	if (vp->v_type == VREG)
1147 		bwillwrite();
1148 	ioflag = IO_UNIT;
1149 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1150 		ioflag |= IO_APPEND;
1151 	if ((fp->f_flag & FNONBLOCK) != 0)
1152 		ioflag |= IO_NDELAY;
1153 	ioflag |= get_write_ioflag(fp);
1154 
1155 	mp = NULL;
1156 	need_finished_write = false;
1157 	if (vp->v_type != VCHR) {
1158 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1159 		if (error != 0)
1160 			goto unlock;
1161 		need_finished_write = true;
1162 	}
1163 
1164 	advice = get_advice(fp, uio);
1165 
1166 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1167 	switch (advice) {
1168 	case POSIX_FADV_NORMAL:
1169 	case POSIX_FADV_SEQUENTIAL:
1170 	case POSIX_FADV_NOREUSE:
1171 		ioflag |= sequential_heuristic(uio, fp);
1172 		break;
1173 	case POSIX_FADV_RANDOM:
1174 		/* XXX: Is this correct? */
1175 		break;
1176 	}
1177 	orig_offset = uio->uio_offset;
1178 
1179 #ifdef MAC
1180 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1181 	if (error == 0)
1182 #endif
1183 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1184 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1185 	VOP_UNLOCK(vp);
1186 	if (need_finished_write)
1187 		vn_finished_write(mp);
1188 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1189 	    orig_offset != uio->uio_offset)
1190 		/*
1191 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1192 		 * for the backing file after a POSIX_FADV_NOREUSE
1193 		 * write(2).
1194 		 */
1195 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1196 		    POSIX_FADV_DONTNEED);
1197 unlock:
1198 	return (error);
1199 }
1200 
1201 /*
1202  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1203  * prevent the following deadlock:
1204  *
1205  * Assume that the thread A reads from the vnode vp1 into userspace
1206  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1207  * currently not resident, then system ends up with the call chain
1208  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1209  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1210  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1211  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1212  * backed by the pages of vnode vp1, and some page in buf2 is not
1213  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1214  *
1215  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1216  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1217  * Instead, it first tries to do the whole range i/o with pagefaults
1218  * disabled. If all pages in the i/o buffer are resident and mapped,
1219  * VOP will succeed (ignoring the genuine filesystem errors).
1220  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1221  * i/o in chunks, with all pages in the chunk prefaulted and held
1222  * using vm_fault_quick_hold_pages().
1223  *
1224  * Filesystems using this deadlock avoidance scheme should use the
1225  * array of the held pages from uio, saved in the curthread->td_ma,
1226  * instead of doing uiomove().  A helper function
1227  * vn_io_fault_uiomove() converts uiomove request into
1228  * uiomove_fromphys() over td_ma array.
1229  *
1230  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1231  * make the current i/o request atomic with respect to other i/os and
1232  * truncations.
1233  */
1234 
1235 /*
1236  * Decode vn_io_fault_args and perform the corresponding i/o.
1237  */
1238 static int
1239 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1240     struct thread *td)
1241 {
1242 	int error, save;
1243 
1244 	error = 0;
1245 	save = vm_fault_disable_pagefaults();
1246 	switch (args->kind) {
1247 	case VN_IO_FAULT_FOP:
1248 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1249 		    uio, args->cred, args->flags, td);
1250 		break;
1251 	case VN_IO_FAULT_VOP:
1252 		if (uio->uio_rw == UIO_READ) {
1253 			error = VOP_READ(args->args.vop_args.vp, uio,
1254 			    args->flags, args->cred);
1255 		} else if (uio->uio_rw == UIO_WRITE) {
1256 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1257 			    args->flags, args->cred);
1258 		}
1259 		break;
1260 	default:
1261 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1262 		    args->kind, uio->uio_rw);
1263 	}
1264 	vm_fault_enable_pagefaults(save);
1265 	return (error);
1266 }
1267 
1268 static int
1269 vn_io_fault_touch(char *base, const struct uio *uio)
1270 {
1271 	int r;
1272 
1273 	r = fubyte(base);
1274 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1275 		return (EFAULT);
1276 	return (0);
1277 }
1278 
1279 static int
1280 vn_io_fault_prefault_user(const struct uio *uio)
1281 {
1282 	char *base;
1283 	const struct iovec *iov;
1284 	size_t len;
1285 	ssize_t resid;
1286 	int error, i;
1287 
1288 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1289 	    ("vn_io_fault_prefault userspace"));
1290 
1291 	error = i = 0;
1292 	iov = uio->uio_iov;
1293 	resid = uio->uio_resid;
1294 	base = iov->iov_base;
1295 	len = iov->iov_len;
1296 	while (resid > 0) {
1297 		error = vn_io_fault_touch(base, uio);
1298 		if (error != 0)
1299 			break;
1300 		if (len < PAGE_SIZE) {
1301 			if (len != 0) {
1302 				error = vn_io_fault_touch(base + len - 1, uio);
1303 				if (error != 0)
1304 					break;
1305 				resid -= len;
1306 			}
1307 			if (++i >= uio->uio_iovcnt)
1308 				break;
1309 			iov = uio->uio_iov + i;
1310 			base = iov->iov_base;
1311 			len = iov->iov_len;
1312 		} else {
1313 			len -= PAGE_SIZE;
1314 			base += PAGE_SIZE;
1315 			resid -= PAGE_SIZE;
1316 		}
1317 	}
1318 	return (error);
1319 }
1320 
1321 /*
1322  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1323  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1324  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1325  * into args and call vn_io_fault1() to handle faults during the user
1326  * mode buffer accesses.
1327  */
1328 static int
1329 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1330     struct thread *td)
1331 {
1332 	vm_page_t ma[io_hold_cnt + 2];
1333 	struct uio *uio_clone, short_uio;
1334 	struct iovec short_iovec[1];
1335 	vm_page_t *prev_td_ma;
1336 	vm_prot_t prot;
1337 	vm_offset_t addr, end;
1338 	size_t len, resid;
1339 	ssize_t adv;
1340 	int error, cnt, saveheld, prev_td_ma_cnt;
1341 
1342 	if (vn_io_fault_prefault) {
1343 		error = vn_io_fault_prefault_user(uio);
1344 		if (error != 0)
1345 			return (error); /* Or ignore ? */
1346 	}
1347 
1348 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1349 
1350 	/*
1351 	 * The UFS follows IO_UNIT directive and replays back both
1352 	 * uio_offset and uio_resid if an error is encountered during the
1353 	 * operation.  But, since the iovec may be already advanced,
1354 	 * uio is still in an inconsistent state.
1355 	 *
1356 	 * Cache a copy of the original uio, which is advanced to the redo
1357 	 * point using UIO_NOCOPY below.
1358 	 */
1359 	uio_clone = cloneuio(uio);
1360 	resid = uio->uio_resid;
1361 
1362 	short_uio.uio_segflg = UIO_USERSPACE;
1363 	short_uio.uio_rw = uio->uio_rw;
1364 	short_uio.uio_td = uio->uio_td;
1365 
1366 	error = vn_io_fault_doio(args, uio, td);
1367 	if (error != EFAULT)
1368 		goto out;
1369 
1370 	atomic_add_long(&vn_io_faults_cnt, 1);
1371 	uio_clone->uio_segflg = UIO_NOCOPY;
1372 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1373 	uio_clone->uio_segflg = uio->uio_segflg;
1374 
1375 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1376 	prev_td_ma = td->td_ma;
1377 	prev_td_ma_cnt = td->td_ma_cnt;
1378 
1379 	while (uio_clone->uio_resid != 0) {
1380 		len = uio_clone->uio_iov->iov_len;
1381 		if (len == 0) {
1382 			KASSERT(uio_clone->uio_iovcnt >= 1,
1383 			    ("iovcnt underflow"));
1384 			uio_clone->uio_iov++;
1385 			uio_clone->uio_iovcnt--;
1386 			continue;
1387 		}
1388 		if (len > ptoa(io_hold_cnt))
1389 			len = ptoa(io_hold_cnt);
1390 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1391 		end = round_page(addr + len);
1392 		if (end < addr) {
1393 			error = EFAULT;
1394 			break;
1395 		}
1396 		/*
1397 		 * A perfectly misaligned address and length could cause
1398 		 * both the start and the end of the chunk to use partial
1399 		 * page.  +2 accounts for such a situation.
1400 		 */
1401 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1402 		    addr, len, prot, ma, io_hold_cnt + 2);
1403 		if (cnt == -1) {
1404 			error = EFAULT;
1405 			break;
1406 		}
1407 		short_uio.uio_iov = &short_iovec[0];
1408 		short_iovec[0].iov_base = (void *)addr;
1409 		short_uio.uio_iovcnt = 1;
1410 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1411 		short_uio.uio_offset = uio_clone->uio_offset;
1412 		td->td_ma = ma;
1413 		td->td_ma_cnt = cnt;
1414 
1415 		error = vn_io_fault_doio(args, &short_uio, td);
1416 		vm_page_unhold_pages(ma, cnt);
1417 		adv = len - short_uio.uio_resid;
1418 
1419 		uio_clone->uio_iov->iov_base =
1420 		    (char *)uio_clone->uio_iov->iov_base + adv;
1421 		uio_clone->uio_iov->iov_len -= adv;
1422 		uio_clone->uio_resid -= adv;
1423 		uio_clone->uio_offset += adv;
1424 
1425 		uio->uio_resid -= adv;
1426 		uio->uio_offset += adv;
1427 
1428 		if (error != 0 || adv == 0)
1429 			break;
1430 	}
1431 	td->td_ma = prev_td_ma;
1432 	td->td_ma_cnt = prev_td_ma_cnt;
1433 	curthread_pflags_restore(saveheld);
1434 out:
1435 	free(uio_clone, M_IOV);
1436 	return (error);
1437 }
1438 
1439 static int
1440 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1441     int flags, struct thread *td)
1442 {
1443 	fo_rdwr_t *doio;
1444 	struct vnode *vp;
1445 	void *rl_cookie;
1446 	struct vn_io_fault_args args;
1447 	int error;
1448 
1449 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1450 	vp = fp->f_vnode;
1451 
1452 	/*
1453 	 * The ability to read(2) on a directory has historically been
1454 	 * allowed for all users, but this can and has been the source of
1455 	 * at least one security issue in the past.  As such, it is now hidden
1456 	 * away behind a sysctl for those that actually need it to use it, and
1457 	 * restricted to root when it's turned on to make it relatively safe to
1458 	 * leave on for longer sessions of need.
1459 	 */
1460 	if (vp->v_type == VDIR) {
1461 		KASSERT(uio->uio_rw == UIO_READ,
1462 		    ("illegal write attempted on a directory"));
1463 		if (!vfs_allow_read_dir)
1464 			return (EISDIR);
1465 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1466 			return (EISDIR);
1467 	}
1468 
1469 	foffset_lock_uio(fp, uio, flags);
1470 	if (do_vn_io_fault(vp, uio)) {
1471 		args.kind = VN_IO_FAULT_FOP;
1472 		args.args.fop_args.fp = fp;
1473 		args.args.fop_args.doio = doio;
1474 		args.cred = active_cred;
1475 		args.flags = flags | FOF_OFFSET;
1476 		if (uio->uio_rw == UIO_READ) {
1477 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1478 			    uio->uio_offset + uio->uio_resid);
1479 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1480 		    (flags & FOF_OFFSET) == 0) {
1481 			/* For appenders, punt and lock the whole range. */
1482 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1483 		} else {
1484 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1485 			    uio->uio_offset + uio->uio_resid);
1486 		}
1487 		error = vn_io_fault1(vp, uio, &args, td);
1488 		vn_rangelock_unlock(vp, rl_cookie);
1489 	} else {
1490 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1491 	}
1492 	foffset_unlock_uio(fp, uio, flags);
1493 	return (error);
1494 }
1495 
1496 /*
1497  * Helper function to perform the requested uiomove operation using
1498  * the held pages for io->uio_iov[0].iov_base buffer instead of
1499  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1500  * instead of iov_base prevents page faults that could occur due to
1501  * pmap_collect() invalidating the mapping created by
1502  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1503  * object cleanup revoking the write access from page mappings.
1504  *
1505  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1506  * instead of plain uiomove().
1507  */
1508 int
1509 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1510 {
1511 	struct uio transp_uio;
1512 	struct iovec transp_iov[1];
1513 	struct thread *td;
1514 	size_t adv;
1515 	int error, pgadv;
1516 
1517 	td = curthread;
1518 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1519 	    uio->uio_segflg != UIO_USERSPACE)
1520 		return (uiomove(data, xfersize, uio));
1521 
1522 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1523 	transp_iov[0].iov_base = data;
1524 	transp_uio.uio_iov = &transp_iov[0];
1525 	transp_uio.uio_iovcnt = 1;
1526 	if (xfersize > uio->uio_resid)
1527 		xfersize = uio->uio_resid;
1528 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1529 	transp_uio.uio_offset = 0;
1530 	transp_uio.uio_segflg = UIO_SYSSPACE;
1531 	/*
1532 	 * Since transp_iov points to data, and td_ma page array
1533 	 * corresponds to original uio->uio_iov, we need to invert the
1534 	 * direction of the i/o operation as passed to
1535 	 * uiomove_fromphys().
1536 	 */
1537 	switch (uio->uio_rw) {
1538 	case UIO_WRITE:
1539 		transp_uio.uio_rw = UIO_READ;
1540 		break;
1541 	case UIO_READ:
1542 		transp_uio.uio_rw = UIO_WRITE;
1543 		break;
1544 	}
1545 	transp_uio.uio_td = uio->uio_td;
1546 	error = uiomove_fromphys(td->td_ma,
1547 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1548 	    xfersize, &transp_uio);
1549 	adv = xfersize - transp_uio.uio_resid;
1550 	pgadv =
1551 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1552 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1553 	td->td_ma += pgadv;
1554 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1555 	    pgadv));
1556 	td->td_ma_cnt -= pgadv;
1557 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1558 	uio->uio_iov->iov_len -= adv;
1559 	uio->uio_resid -= adv;
1560 	uio->uio_offset += adv;
1561 	return (error);
1562 }
1563 
1564 int
1565 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1566     struct uio *uio)
1567 {
1568 	struct thread *td;
1569 	vm_offset_t iov_base;
1570 	int cnt, pgadv;
1571 
1572 	td = curthread;
1573 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1574 	    uio->uio_segflg != UIO_USERSPACE)
1575 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1576 
1577 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1578 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1579 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1580 	switch (uio->uio_rw) {
1581 	case UIO_WRITE:
1582 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1583 		    offset, cnt);
1584 		break;
1585 	case UIO_READ:
1586 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1587 		    cnt);
1588 		break;
1589 	}
1590 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1591 	td->td_ma += pgadv;
1592 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1593 	    pgadv));
1594 	td->td_ma_cnt -= pgadv;
1595 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1596 	uio->uio_iov->iov_len -= cnt;
1597 	uio->uio_resid -= cnt;
1598 	uio->uio_offset += cnt;
1599 	return (0);
1600 }
1601 
1602 /*
1603  * File table truncate routine.
1604  */
1605 static int
1606 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1607     struct thread *td)
1608 {
1609 	struct mount *mp;
1610 	struct vnode *vp;
1611 	void *rl_cookie;
1612 	int error;
1613 
1614 	vp = fp->f_vnode;
1615 
1616 retry:
1617 	/*
1618 	 * Lock the whole range for truncation.  Otherwise split i/o
1619 	 * might happen partly before and partly after the truncation.
1620 	 */
1621 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1622 	error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1623 	if (error)
1624 		goto out1;
1625 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1626 	AUDIT_ARG_VNODE1(vp);
1627 	if (vp->v_type == VDIR) {
1628 		error = EISDIR;
1629 		goto out;
1630 	}
1631 #ifdef MAC
1632 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1633 	if (error)
1634 		goto out;
1635 #endif
1636 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1637 	    fp->f_cred);
1638 out:
1639 	VOP_UNLOCK(vp);
1640 	vn_finished_write(mp);
1641 out1:
1642 	vn_rangelock_unlock(vp, rl_cookie);
1643 	if (error == ERELOOKUP)
1644 		goto retry;
1645 	return (error);
1646 }
1647 
1648 /*
1649  * Truncate a file that is already locked.
1650  */
1651 int
1652 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1653     struct ucred *cred)
1654 {
1655 	struct vattr vattr;
1656 	int error;
1657 
1658 	error = VOP_ADD_WRITECOUNT(vp, 1);
1659 	if (error == 0) {
1660 		VATTR_NULL(&vattr);
1661 		vattr.va_size = length;
1662 		if (sync)
1663 			vattr.va_vaflags |= VA_SYNC;
1664 		error = VOP_SETATTR(vp, &vattr, cred);
1665 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1666 	}
1667 	return (error);
1668 }
1669 
1670 /*
1671  * File table vnode stat routine.
1672  */
1673 int
1674 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred)
1675 {
1676 	struct vnode *vp = fp->f_vnode;
1677 	int error;
1678 
1679 	vn_lock(vp, LK_SHARED | LK_RETRY);
1680 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1681 	VOP_UNLOCK(vp);
1682 
1683 	return (error);
1684 }
1685 
1686 /*
1687  * File table vnode ioctl routine.
1688  */
1689 static int
1690 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1691     struct thread *td)
1692 {
1693 	struct vnode *vp;
1694 	struct fiobmap2_arg *bmarg;
1695 	off_t size;
1696 	int error;
1697 
1698 	vp = fp->f_vnode;
1699 	switch (vp->v_type) {
1700 	case VDIR:
1701 	case VREG:
1702 		switch (com) {
1703 		case FIONREAD:
1704 			error = vn_getsize(vp, &size, active_cred);
1705 			if (error == 0)
1706 				*(int *)data = size - fp->f_offset;
1707 			return (error);
1708 		case FIOBMAP2:
1709 			bmarg = (struct fiobmap2_arg *)data;
1710 			vn_lock(vp, LK_SHARED | LK_RETRY);
1711 #ifdef MAC
1712 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1713 			    vp);
1714 			if (error == 0)
1715 #endif
1716 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1717 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1718 			VOP_UNLOCK(vp);
1719 			return (error);
1720 		case FIONBIO:
1721 		case FIOASYNC:
1722 			return (0);
1723 		default:
1724 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1725 			    active_cred, td));
1726 		}
1727 		break;
1728 	case VCHR:
1729 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1730 		    active_cred, td));
1731 	default:
1732 		return (ENOTTY);
1733 	}
1734 }
1735 
1736 /*
1737  * File table vnode poll routine.
1738  */
1739 static int
1740 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1741     struct thread *td)
1742 {
1743 	struct vnode *vp;
1744 	int error;
1745 
1746 	vp = fp->f_vnode;
1747 #if defined(MAC) || defined(AUDIT)
1748 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1749 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1750 		AUDIT_ARG_VNODE1(vp);
1751 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1752 		VOP_UNLOCK(vp);
1753 		if (error != 0)
1754 			return (error);
1755 	}
1756 #endif
1757 	error = VOP_POLL(vp, events, fp->f_cred, td);
1758 	return (error);
1759 }
1760 
1761 /*
1762  * Acquire the requested lock and then check for validity.  LK_RETRY
1763  * permits vn_lock to return doomed vnodes.
1764  */
1765 static int __noinline
1766 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1767     int error)
1768 {
1769 
1770 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1771 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1772 
1773 	if (error == 0)
1774 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1775 
1776 	if ((flags & LK_RETRY) == 0) {
1777 		if (error == 0) {
1778 			VOP_UNLOCK(vp);
1779 			error = ENOENT;
1780 		}
1781 		return (error);
1782 	}
1783 
1784 	/*
1785 	 * LK_RETRY case.
1786 	 *
1787 	 * Nothing to do if we got the lock.
1788 	 */
1789 	if (error == 0)
1790 		return (0);
1791 
1792 	/*
1793 	 * Interlock was dropped by the call in _vn_lock.
1794 	 */
1795 	flags &= ~LK_INTERLOCK;
1796 	do {
1797 		error = VOP_LOCK1(vp, flags, file, line);
1798 	} while (error != 0);
1799 	return (0);
1800 }
1801 
1802 int
1803 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1804 {
1805 	int error;
1806 
1807 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1808 	    ("vn_lock: no locktype (%d passed)", flags));
1809 	VNPASS(vp->v_holdcnt > 0, vp);
1810 	error = VOP_LOCK1(vp, flags, file, line);
1811 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1812 		return (_vn_lock_fallback(vp, flags, file, line, error));
1813 	return (0);
1814 }
1815 
1816 /*
1817  * File table vnode close routine.
1818  */
1819 static int
1820 vn_closefile(struct file *fp, struct thread *td)
1821 {
1822 	struct vnode *vp;
1823 	struct flock lf;
1824 	int error;
1825 	bool ref;
1826 
1827 	vp = fp->f_vnode;
1828 	fp->f_ops = &badfileops;
1829 	ref = (fp->f_flag & FHASLOCK) != 0;
1830 
1831 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1832 
1833 	if (__predict_false(ref)) {
1834 		lf.l_whence = SEEK_SET;
1835 		lf.l_start = 0;
1836 		lf.l_len = 0;
1837 		lf.l_type = F_UNLCK;
1838 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1839 		vrele(vp);
1840 	}
1841 	return (error);
1842 }
1843 
1844 /*
1845  * Preparing to start a filesystem write operation. If the operation is
1846  * permitted, then we bump the count of operations in progress and
1847  * proceed. If a suspend request is in progress, we wait until the
1848  * suspension is over, and then proceed.
1849  */
1850 static int
1851 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1852 {
1853 	struct mount_pcpu *mpcpu;
1854 	int error, mflags;
1855 
1856 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1857 	    vfs_op_thread_enter(mp, mpcpu)) {
1858 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1859 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1860 		vfs_op_thread_exit(mp, mpcpu);
1861 		return (0);
1862 	}
1863 
1864 	if (mplocked)
1865 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1866 	else
1867 		MNT_ILOCK(mp);
1868 
1869 	error = 0;
1870 
1871 	/*
1872 	 * Check on status of suspension.
1873 	 */
1874 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1875 	    mp->mnt_susp_owner != curthread) {
1876 		mflags = 0;
1877 		if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
1878 			if (flags & V_PCATCH)
1879 				mflags |= PCATCH;
1880 		}
1881 		mflags |= (PUSER - 1);
1882 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1883 			if ((flags & V_NOWAIT) != 0) {
1884 				error = EWOULDBLOCK;
1885 				goto unlock;
1886 			}
1887 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1888 			    "suspfs", 0);
1889 			if (error != 0)
1890 				goto unlock;
1891 		}
1892 	}
1893 	if ((flags & V_XSLEEP) != 0)
1894 		goto unlock;
1895 	mp->mnt_writeopcount++;
1896 unlock:
1897 	if (error != 0 || (flags & V_XSLEEP) != 0)
1898 		MNT_REL(mp);
1899 	MNT_IUNLOCK(mp);
1900 	return (error);
1901 }
1902 
1903 int
1904 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1905 {
1906 	struct mount *mp;
1907 	int error;
1908 
1909 	KASSERT((flags & ~V_VALID_FLAGS) == 0,
1910 	    ("%s: invalid flags passed %d\n", __func__, flags));
1911 
1912 	error = 0;
1913 	/*
1914 	 * If a vnode is provided, get and return the mount point that
1915 	 * to which it will write.
1916 	 */
1917 	if (vp != NULL) {
1918 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1919 			*mpp = NULL;
1920 			if (error != EOPNOTSUPP)
1921 				return (error);
1922 			return (0);
1923 		}
1924 	}
1925 	if ((mp = *mpp) == NULL)
1926 		return (0);
1927 
1928 	/*
1929 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1930 	 * a vfs_ref().
1931 	 * As long as a vnode is not provided we need to acquire a
1932 	 * refcount for the provided mountpoint too, in order to
1933 	 * emulate a vfs_ref().
1934 	 */
1935 	if (vp == NULL)
1936 		vfs_ref(mp);
1937 
1938 	error = vn_start_write_refed(mp, flags, false);
1939 	if (error != 0 && (flags & V_NOWAIT) == 0)
1940 		*mpp = NULL;
1941 	return (error);
1942 }
1943 
1944 /*
1945  * Secondary suspension. Used by operations such as vop_inactive
1946  * routines that are needed by the higher level functions. These
1947  * are allowed to proceed until all the higher level functions have
1948  * completed (indicated by mnt_writeopcount dropping to zero). At that
1949  * time, these operations are halted until the suspension is over.
1950  */
1951 int
1952 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1953 {
1954 	struct mount *mp;
1955 	int error, mflags;
1956 
1957 	KASSERT((flags & (~V_VALID_FLAGS | V_XSLEEP)) == 0,
1958 	    ("%s: invalid flags passed %d\n", __func__, flags));
1959 
1960  retry:
1961 	if (vp != NULL) {
1962 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1963 			*mpp = NULL;
1964 			if (error != EOPNOTSUPP)
1965 				return (error);
1966 			return (0);
1967 		}
1968 	}
1969 	/*
1970 	 * If we are not suspended or have not yet reached suspended
1971 	 * mode, then let the operation proceed.
1972 	 */
1973 	if ((mp = *mpp) == NULL)
1974 		return (0);
1975 
1976 	/*
1977 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1978 	 * a vfs_ref().
1979 	 * As long as a vnode is not provided we need to acquire a
1980 	 * refcount for the provided mountpoint too, in order to
1981 	 * emulate a vfs_ref().
1982 	 */
1983 	MNT_ILOCK(mp);
1984 	if (vp == NULL)
1985 		MNT_REF(mp);
1986 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1987 		mp->mnt_secondary_writes++;
1988 		mp->mnt_secondary_accwrites++;
1989 		MNT_IUNLOCK(mp);
1990 		return (0);
1991 	}
1992 	if ((flags & V_NOWAIT) != 0) {
1993 		MNT_REL(mp);
1994 		MNT_IUNLOCK(mp);
1995 		*mpp = NULL;
1996 		return (EWOULDBLOCK);
1997 	}
1998 	/*
1999 	 * Wait for the suspension to finish.
2000 	 */
2001 	mflags = 0;
2002 	if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
2003 		if ((flags & V_PCATCH) != 0)
2004 			mflags |= PCATCH;
2005 	}
2006 	mflags |= (PUSER - 1) | PDROP;
2007 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0);
2008 	vfs_rel(mp);
2009 	if (error == 0)
2010 		goto retry;
2011 	*mpp = NULL;
2012 	return (error);
2013 }
2014 
2015 /*
2016  * Filesystem write operation has completed. If we are suspending and this
2017  * operation is the last one, notify the suspender that the suspension is
2018  * now in effect.
2019  */
2020 void
2021 vn_finished_write(struct mount *mp)
2022 {
2023 	struct mount_pcpu *mpcpu;
2024 	int c;
2025 
2026 	if (mp == NULL)
2027 		return;
2028 
2029 	if (vfs_op_thread_enter(mp, mpcpu)) {
2030 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2031 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2032 		vfs_op_thread_exit(mp, mpcpu);
2033 		return;
2034 	}
2035 
2036 	MNT_ILOCK(mp);
2037 	vfs_assert_mount_counters(mp);
2038 	MNT_REL(mp);
2039 	c = --mp->mnt_writeopcount;
2040 	if (mp->mnt_vfs_ops == 0) {
2041 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2042 		MNT_IUNLOCK(mp);
2043 		return;
2044 	}
2045 	if (c < 0)
2046 		vfs_dump_mount_counters(mp);
2047 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2048 		wakeup(&mp->mnt_writeopcount);
2049 	MNT_IUNLOCK(mp);
2050 }
2051 
2052 /*
2053  * Filesystem secondary write operation has completed. If we are
2054  * suspending and this operation is the last one, notify the suspender
2055  * that the suspension is now in effect.
2056  */
2057 void
2058 vn_finished_secondary_write(struct mount *mp)
2059 {
2060 	if (mp == NULL)
2061 		return;
2062 	MNT_ILOCK(mp);
2063 	MNT_REL(mp);
2064 	mp->mnt_secondary_writes--;
2065 	if (mp->mnt_secondary_writes < 0)
2066 		panic("vn_finished_secondary_write: neg cnt");
2067 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2068 	    mp->mnt_secondary_writes <= 0)
2069 		wakeup(&mp->mnt_secondary_writes);
2070 	MNT_IUNLOCK(mp);
2071 }
2072 
2073 /*
2074  * Request a filesystem to suspend write operations.
2075  */
2076 int
2077 vfs_write_suspend(struct mount *mp, int flags)
2078 {
2079 	int error;
2080 
2081 	vfs_op_enter(mp);
2082 
2083 	MNT_ILOCK(mp);
2084 	vfs_assert_mount_counters(mp);
2085 	if (mp->mnt_susp_owner == curthread) {
2086 		vfs_op_exit_locked(mp);
2087 		MNT_IUNLOCK(mp);
2088 		return (EALREADY);
2089 	}
2090 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2091 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2092 
2093 	/*
2094 	 * Unmount holds a write reference on the mount point.  If we
2095 	 * own busy reference and drain for writers, we deadlock with
2096 	 * the reference draining in the unmount path.  Callers of
2097 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2098 	 * vfs_busy() reference is owned and caller is not in the
2099 	 * unmount context.
2100 	 */
2101 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2102 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2103 		vfs_op_exit_locked(mp);
2104 		MNT_IUNLOCK(mp);
2105 		return (EBUSY);
2106 	}
2107 
2108 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2109 	mp->mnt_susp_owner = curthread;
2110 	if (mp->mnt_writeopcount > 0)
2111 		(void) msleep(&mp->mnt_writeopcount,
2112 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2113 	else
2114 		MNT_IUNLOCK(mp);
2115 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2116 		vfs_write_resume(mp, 0);
2117 		/* vfs_write_resume does vfs_op_exit() for us */
2118 	}
2119 	return (error);
2120 }
2121 
2122 /*
2123  * Request a filesystem to resume write operations.
2124  */
2125 void
2126 vfs_write_resume(struct mount *mp, int flags)
2127 {
2128 
2129 	MNT_ILOCK(mp);
2130 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2131 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2132 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2133 				       MNTK_SUSPENDED);
2134 		mp->mnt_susp_owner = NULL;
2135 		wakeup(&mp->mnt_writeopcount);
2136 		wakeup(&mp->mnt_flag);
2137 		curthread->td_pflags &= ~TDP_IGNSUSP;
2138 		if ((flags & VR_START_WRITE) != 0) {
2139 			MNT_REF(mp);
2140 			mp->mnt_writeopcount++;
2141 		}
2142 		MNT_IUNLOCK(mp);
2143 		if ((flags & VR_NO_SUSPCLR) == 0)
2144 			VFS_SUSP_CLEAN(mp);
2145 		vfs_op_exit(mp);
2146 	} else if ((flags & VR_START_WRITE) != 0) {
2147 		MNT_REF(mp);
2148 		vn_start_write_refed(mp, 0, true);
2149 	} else {
2150 		MNT_IUNLOCK(mp);
2151 	}
2152 }
2153 
2154 /*
2155  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2156  * methods.
2157  */
2158 int
2159 vfs_write_suspend_umnt(struct mount *mp)
2160 {
2161 	int error;
2162 
2163 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2164 	    ("vfs_write_suspend_umnt: recursed"));
2165 
2166 	/* dounmount() already called vn_start_write(). */
2167 	for (;;) {
2168 		vn_finished_write(mp);
2169 		error = vfs_write_suspend(mp, 0);
2170 		if (error != 0) {
2171 			vn_start_write(NULL, &mp, V_WAIT);
2172 			return (error);
2173 		}
2174 		MNT_ILOCK(mp);
2175 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2176 			break;
2177 		MNT_IUNLOCK(mp);
2178 		vn_start_write(NULL, &mp, V_WAIT);
2179 	}
2180 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2181 	wakeup(&mp->mnt_flag);
2182 	MNT_IUNLOCK(mp);
2183 	curthread->td_pflags |= TDP_IGNSUSP;
2184 	return (0);
2185 }
2186 
2187 /*
2188  * Implement kqueues for files by translating it to vnode operation.
2189  */
2190 static int
2191 vn_kqfilter(struct file *fp, struct knote *kn)
2192 {
2193 
2194 	return (VOP_KQFILTER(fp->f_vnode, kn));
2195 }
2196 
2197 int
2198 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2199 {
2200 	if ((fp->f_flag & FKQALLOWED) == 0)
2201 		return (EBADF);
2202 	return (vn_kqfilter(fp, kn));
2203 }
2204 
2205 /*
2206  * Simplified in-kernel wrapper calls for extended attribute access.
2207  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2208  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2209  */
2210 int
2211 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2212     const char *attrname, int *buflen, char *buf, struct thread *td)
2213 {
2214 	struct uio	auio;
2215 	struct iovec	iov;
2216 	int	error;
2217 
2218 	iov.iov_len = *buflen;
2219 	iov.iov_base = buf;
2220 
2221 	auio.uio_iov = &iov;
2222 	auio.uio_iovcnt = 1;
2223 	auio.uio_rw = UIO_READ;
2224 	auio.uio_segflg = UIO_SYSSPACE;
2225 	auio.uio_td = td;
2226 	auio.uio_offset = 0;
2227 	auio.uio_resid = *buflen;
2228 
2229 	if ((ioflg & IO_NODELOCKED) == 0)
2230 		vn_lock(vp, LK_SHARED | LK_RETRY);
2231 
2232 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2233 
2234 	/* authorize attribute retrieval as kernel */
2235 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2236 	    td);
2237 
2238 	if ((ioflg & IO_NODELOCKED) == 0)
2239 		VOP_UNLOCK(vp);
2240 
2241 	if (error == 0) {
2242 		*buflen = *buflen - auio.uio_resid;
2243 	}
2244 
2245 	return (error);
2246 }
2247 
2248 /*
2249  * XXX failure mode if partially written?
2250  */
2251 int
2252 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2253     const char *attrname, int buflen, char *buf, struct thread *td)
2254 {
2255 	struct uio	auio;
2256 	struct iovec	iov;
2257 	struct mount	*mp;
2258 	int	error;
2259 
2260 	iov.iov_len = buflen;
2261 	iov.iov_base = buf;
2262 
2263 	auio.uio_iov = &iov;
2264 	auio.uio_iovcnt = 1;
2265 	auio.uio_rw = UIO_WRITE;
2266 	auio.uio_segflg = UIO_SYSSPACE;
2267 	auio.uio_td = td;
2268 	auio.uio_offset = 0;
2269 	auio.uio_resid = buflen;
2270 
2271 	if ((ioflg & IO_NODELOCKED) == 0) {
2272 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2273 			return (error);
2274 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2275 	}
2276 
2277 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2278 
2279 	/* authorize attribute setting as kernel */
2280 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2281 
2282 	if ((ioflg & IO_NODELOCKED) == 0) {
2283 		vn_finished_write(mp);
2284 		VOP_UNLOCK(vp);
2285 	}
2286 
2287 	return (error);
2288 }
2289 
2290 int
2291 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2292     const char *attrname, struct thread *td)
2293 {
2294 	struct mount	*mp;
2295 	int	error;
2296 
2297 	if ((ioflg & IO_NODELOCKED) == 0) {
2298 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2299 			return (error);
2300 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2301 	}
2302 
2303 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2304 
2305 	/* authorize attribute removal as kernel */
2306 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2307 	if (error == EOPNOTSUPP)
2308 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2309 		    NULL, td);
2310 
2311 	if ((ioflg & IO_NODELOCKED) == 0) {
2312 		vn_finished_write(mp);
2313 		VOP_UNLOCK(vp);
2314 	}
2315 
2316 	return (error);
2317 }
2318 
2319 static int
2320 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2321     struct vnode **rvp)
2322 {
2323 
2324 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2325 }
2326 
2327 int
2328 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2329 {
2330 
2331 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2332 	    lkflags, rvp));
2333 }
2334 
2335 int
2336 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2337     int lkflags, struct vnode **rvp)
2338 {
2339 	struct mount *mp;
2340 	int ltype, error;
2341 
2342 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2343 	mp = vp->v_mount;
2344 	ltype = VOP_ISLOCKED(vp);
2345 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2346 	    ("vn_vget_ino: vp not locked"));
2347 	error = vfs_busy(mp, MBF_NOWAIT);
2348 	if (error != 0) {
2349 		vfs_ref(mp);
2350 		VOP_UNLOCK(vp);
2351 		error = vfs_busy(mp, 0);
2352 		vn_lock(vp, ltype | LK_RETRY);
2353 		vfs_rel(mp);
2354 		if (error != 0)
2355 			return (ENOENT);
2356 		if (VN_IS_DOOMED(vp)) {
2357 			vfs_unbusy(mp);
2358 			return (ENOENT);
2359 		}
2360 	}
2361 	VOP_UNLOCK(vp);
2362 	error = alloc(mp, alloc_arg, lkflags, rvp);
2363 	vfs_unbusy(mp);
2364 	if (error != 0 || *rvp != vp)
2365 		vn_lock(vp, ltype | LK_RETRY);
2366 	if (VN_IS_DOOMED(vp)) {
2367 		if (error == 0) {
2368 			if (*rvp == vp)
2369 				vunref(vp);
2370 			else
2371 				vput(*rvp);
2372 		}
2373 		error = ENOENT;
2374 	}
2375 	return (error);
2376 }
2377 
2378 static void
2379 vn_send_sigxfsz(struct proc *p)
2380 {
2381 	PROC_LOCK(p);
2382 	kern_psignal(p, SIGXFSZ);
2383 	PROC_UNLOCK(p);
2384 }
2385 
2386 int
2387 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2388 {
2389 	if (size <= lim_cur(td, RLIMIT_FSIZE))
2390 		return (0);
2391 	vn_send_sigxfsz(td->td_proc);
2392 	return (EFBIG);
2393 }
2394 
2395 static int
2396 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2397     bool adj, struct thread *td)
2398 {
2399 	off_t lim;
2400 	bool ktr_write;
2401 
2402 	if (vp->v_type != VREG)
2403 		return (0);
2404 
2405 	/*
2406 	 * Handle file system maximum file size.
2407 	 */
2408 	if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2409 		if (!adj || uio->uio_offset >= maxfsz)
2410 			return (EFBIG);
2411 		uio->uio_resid = maxfsz - uio->uio_offset;
2412 	}
2413 
2414 	/*
2415 	 * This is kernel write (e.g. vnode_pager) or accounting
2416 	 * write, ignore limit.
2417 	 */
2418 	if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2419 		return (0);
2420 
2421 	/*
2422 	 * Calculate file size limit.
2423 	 */
2424 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2425 	lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2426 	    lim_cur(td, RLIMIT_FSIZE);
2427 
2428 	/*
2429 	 * Is the limit reached?
2430 	 */
2431 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2432 		return (0);
2433 
2434 	/*
2435 	 * Prepared filesystems can handle writes truncated to the
2436 	 * file size limit.
2437 	 */
2438 	if (adj && (uoff_t)uio->uio_offset < lim) {
2439 		uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2440 		return (0);
2441 	}
2442 
2443 	if (!ktr_write || ktr_filesize_limit_signal)
2444 		vn_send_sigxfsz(td->td_proc);
2445 	return (EFBIG);
2446 }
2447 
2448 /*
2449  * Helper for VOP_WRITE() implementations, the common code to
2450  * handle maximum supported file size on the filesystem, and
2451  * RLIMIT_FSIZE, except for special writes from accounting subsystem
2452  * and ktrace.
2453  *
2454  * For maximum file size (maxfsz argument):
2455  * - return EFBIG if uio_offset is beyond it
2456  * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2457  *
2458  * For RLIMIT_FSIZE:
2459  * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2460  * - otherwise, clamp uio_resid if write would extend file beyond limit.
2461  *
2462  * If clamping occured, the adjustment for uio_resid is stored in
2463  * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2464  * from the VOP.
2465  */
2466 int
2467 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2468     ssize_t *resid_adj, struct thread *td)
2469 {
2470 	ssize_t resid_orig;
2471 	int error;
2472 	bool adj;
2473 
2474 	resid_orig = uio->uio_resid;
2475 	adj = resid_adj != NULL;
2476 	error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2477 	if (adj)
2478 		*resid_adj = resid_orig - uio->uio_resid;
2479 	return (error);
2480 }
2481 
2482 void
2483 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2484 {
2485 	uio->uio_resid += resid_adj;
2486 }
2487 
2488 int
2489 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2490     struct thread *td)
2491 {
2492 	return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2493 	    td));
2494 }
2495 
2496 int
2497 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2498     struct thread *td)
2499 {
2500 	struct vnode *vp;
2501 
2502 	vp = fp->f_vnode;
2503 #ifdef AUDIT
2504 	vn_lock(vp, LK_SHARED | LK_RETRY);
2505 	AUDIT_ARG_VNODE1(vp);
2506 	VOP_UNLOCK(vp);
2507 #endif
2508 	return (setfmode(td, active_cred, vp, mode));
2509 }
2510 
2511 int
2512 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2513     struct thread *td)
2514 {
2515 	struct vnode *vp;
2516 
2517 	vp = fp->f_vnode;
2518 #ifdef AUDIT
2519 	vn_lock(vp, LK_SHARED | LK_RETRY);
2520 	AUDIT_ARG_VNODE1(vp);
2521 	VOP_UNLOCK(vp);
2522 #endif
2523 	return (setfown(td, active_cred, vp, uid, gid));
2524 }
2525 
2526 /*
2527  * Remove pages in the range ["start", "end") from the vnode's VM object.  If
2528  * "end" is 0, then the range extends to the end of the object.
2529  */
2530 void
2531 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2532 {
2533 	vm_object_t object;
2534 
2535 	if ((object = vp->v_object) == NULL)
2536 		return;
2537 	VM_OBJECT_WLOCK(object);
2538 	vm_object_page_remove(object, start, end, 0);
2539 	VM_OBJECT_WUNLOCK(object);
2540 }
2541 
2542 /*
2543  * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2544  * mapped into any process' address space.  Filesystems may use this in
2545  * preference to vn_pages_remove() to avoid blocking on pages busied in
2546  * preparation for a VOP_GETPAGES.
2547  */
2548 void
2549 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2550 {
2551 	vm_object_t object;
2552 
2553 	if ((object = vp->v_object) == NULL)
2554 		return;
2555 	VM_OBJECT_WLOCK(object);
2556 	vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2557 	VM_OBJECT_WUNLOCK(object);
2558 }
2559 
2560 int
2561 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2562     struct ucred *cred)
2563 {
2564 	vm_object_t obj;
2565 	off_t size;
2566 	daddr_t bn, bnp;
2567 	uint64_t bsize;
2568 	off_t noff;
2569 	int error;
2570 
2571 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2572 	    ("%s: Wrong command %lu", __func__, cmd));
2573 	ASSERT_VOP_ELOCKED(vp, "vn_bmap_seekhole_locked");
2574 
2575 	if (vp->v_type != VREG) {
2576 		error = ENOTTY;
2577 		goto out;
2578 	}
2579 	error = vn_getsize_locked(vp, &size, cred);
2580 	if (error != 0)
2581 		goto out;
2582 	noff = *off;
2583 	if (noff < 0 || noff >= size) {
2584 		error = ENXIO;
2585 		goto out;
2586 	}
2587 
2588 	/* See the comment in ufs_bmap_seekdata(). */
2589 	obj = vp->v_object;
2590 	if (obj != NULL) {
2591 		VM_OBJECT_WLOCK(obj);
2592 		vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
2593 		VM_OBJECT_WUNLOCK(obj);
2594 	}
2595 
2596 	bsize = vp->v_mount->mnt_stat.f_iosize;
2597 	for (bn = noff / bsize; noff < size; bn++, noff += bsize -
2598 	    noff % bsize) {
2599 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2600 		if (error == EOPNOTSUPP) {
2601 			error = ENOTTY;
2602 			goto out;
2603 		}
2604 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2605 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2606 			noff = bn * bsize;
2607 			if (noff < *off)
2608 				noff = *off;
2609 			goto out;
2610 		}
2611 	}
2612 	if (noff > size)
2613 		noff = size;
2614 	/* noff == size. There is an implicit hole at the end of file. */
2615 	if (cmd == FIOSEEKDATA)
2616 		error = ENXIO;
2617 out:
2618 	if (error == 0)
2619 		*off = noff;
2620 	return (error);
2621 }
2622 
2623 int
2624 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2625 {
2626 	int error;
2627 
2628 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2629 	    ("%s: Wrong command %lu", __func__, cmd));
2630 
2631 	if (vn_lock(vp, LK_EXCLUSIVE) != 0)
2632 		return (EBADF);
2633 	error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2634 	VOP_UNLOCK(vp);
2635 	return (error);
2636 }
2637 
2638 int
2639 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2640 {
2641 	struct ucred *cred;
2642 	struct vnode *vp;
2643 	off_t foffset, fsize, size;
2644 	int error, noneg;
2645 
2646 	cred = td->td_ucred;
2647 	vp = fp->f_vnode;
2648 	foffset = foffset_lock(fp, 0);
2649 	noneg = (vp->v_type != VCHR);
2650 	error = 0;
2651 	switch (whence) {
2652 	case L_INCR:
2653 		if (noneg &&
2654 		    (foffset < 0 ||
2655 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2656 			error = EOVERFLOW;
2657 			break;
2658 		}
2659 		offset += foffset;
2660 		break;
2661 	case L_XTND:
2662 		error = vn_getsize(vp, &fsize, cred);
2663 		if (error != 0)
2664 			break;
2665 
2666 		/*
2667 		 * If the file references a disk device, then fetch
2668 		 * the media size and use that to determine the ending
2669 		 * offset.
2670 		 */
2671 		if (fsize == 0 && vp->v_type == VCHR &&
2672 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2673 			fsize = size;
2674 		if (noneg && offset > 0 && fsize > OFF_MAX - offset) {
2675 			error = EOVERFLOW;
2676 			break;
2677 		}
2678 		offset += fsize;
2679 		break;
2680 	case L_SET:
2681 		break;
2682 	case SEEK_DATA:
2683 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2684 		if (error == ENOTTY)
2685 			error = EINVAL;
2686 		break;
2687 	case SEEK_HOLE:
2688 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2689 		if (error == ENOTTY)
2690 			error = EINVAL;
2691 		break;
2692 	default:
2693 		error = EINVAL;
2694 	}
2695 	if (error == 0 && noneg && offset < 0)
2696 		error = EINVAL;
2697 	if (error != 0)
2698 		goto drop;
2699 	VFS_KNOTE_UNLOCKED(vp, 0);
2700 	td->td_uretoff.tdu_off = offset;
2701 drop:
2702 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2703 	return (error);
2704 }
2705 
2706 int
2707 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2708     struct thread *td)
2709 {
2710 	int error;
2711 
2712 	/*
2713 	 * Grant permission if the caller is the owner of the file, or
2714 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2715 	 * on the file.  If the time pointer is null, then write
2716 	 * permission on the file is also sufficient.
2717 	 *
2718 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2719 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2720 	 * will be allowed to set the times [..] to the current
2721 	 * server time.
2722 	 */
2723 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2724 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2725 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2726 	return (error);
2727 }
2728 
2729 int
2730 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2731 {
2732 	struct vnode *vp;
2733 	int error;
2734 
2735 	if (fp->f_type == DTYPE_FIFO)
2736 		kif->kf_type = KF_TYPE_FIFO;
2737 	else
2738 		kif->kf_type = KF_TYPE_VNODE;
2739 	vp = fp->f_vnode;
2740 	vref(vp);
2741 	FILEDESC_SUNLOCK(fdp);
2742 	error = vn_fill_kinfo_vnode(vp, kif);
2743 	vrele(vp);
2744 	FILEDESC_SLOCK(fdp);
2745 	return (error);
2746 }
2747 
2748 static inline void
2749 vn_fill_junk(struct kinfo_file *kif)
2750 {
2751 	size_t len, olen;
2752 
2753 	/*
2754 	 * Simulate vn_fullpath returning changing values for a given
2755 	 * vp during e.g. coredump.
2756 	 */
2757 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2758 	olen = strlen(kif->kf_path);
2759 	if (len < olen)
2760 		strcpy(&kif->kf_path[len - 1], "$");
2761 	else
2762 		for (; olen < len; olen++)
2763 			strcpy(&kif->kf_path[olen], "A");
2764 }
2765 
2766 int
2767 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2768 {
2769 	struct vattr va;
2770 	char *fullpath, *freepath;
2771 	int error;
2772 
2773 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2774 	freepath = NULL;
2775 	fullpath = "-";
2776 	error = vn_fullpath(vp, &fullpath, &freepath);
2777 	if (error == 0) {
2778 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2779 	}
2780 	if (freepath != NULL)
2781 		free(freepath, M_TEMP);
2782 
2783 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2784 		vn_fill_junk(kif);
2785 	);
2786 
2787 	/*
2788 	 * Retrieve vnode attributes.
2789 	 */
2790 	va.va_fsid = VNOVAL;
2791 	va.va_rdev = NODEV;
2792 	vn_lock(vp, LK_SHARED | LK_RETRY);
2793 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2794 	VOP_UNLOCK(vp);
2795 	if (error != 0)
2796 		return (error);
2797 	if (va.va_fsid != VNOVAL)
2798 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2799 	else
2800 		kif->kf_un.kf_file.kf_file_fsid =
2801 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2802 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2803 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2804 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2805 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2806 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2807 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2808 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2809 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2810 	kif->kf_un.kf_file.kf_file_nlink = va.va_nlink;
2811 	return (0);
2812 }
2813 
2814 int
2815 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2816     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2817     struct thread *td)
2818 {
2819 #ifdef HWPMC_HOOKS
2820 	struct pmckern_map_in pkm;
2821 #endif
2822 	struct mount *mp;
2823 	struct vnode *vp;
2824 	vm_object_t object;
2825 	vm_prot_t maxprot;
2826 	boolean_t writecounted;
2827 	int error;
2828 
2829 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2830     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2831 	/*
2832 	 * POSIX shared-memory objects are defined to have
2833 	 * kernel persistence, and are not defined to support
2834 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2835 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2836 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2837 	 * flag to request this behavior.
2838 	 */
2839 	if ((fp->f_flag & FPOSIXSHM) != 0)
2840 		flags |= MAP_NOSYNC;
2841 #endif
2842 	vp = fp->f_vnode;
2843 
2844 	/*
2845 	 * Ensure that file and memory protections are
2846 	 * compatible.  Note that we only worry about
2847 	 * writability if mapping is shared; in this case,
2848 	 * current and max prot are dictated by the open file.
2849 	 * XXX use the vnode instead?  Problem is: what
2850 	 * credentials do we use for determination? What if
2851 	 * proc does a setuid?
2852 	 */
2853 	mp = vp->v_mount;
2854 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2855 		maxprot = VM_PROT_NONE;
2856 		if ((prot & VM_PROT_EXECUTE) != 0)
2857 			return (EACCES);
2858 	} else
2859 		maxprot = VM_PROT_EXECUTE;
2860 	if ((fp->f_flag & FREAD) != 0)
2861 		maxprot |= VM_PROT_READ;
2862 	else if ((prot & VM_PROT_READ) != 0)
2863 		return (EACCES);
2864 
2865 	/*
2866 	 * If we are sharing potential changes via MAP_SHARED and we
2867 	 * are trying to get write permission although we opened it
2868 	 * without asking for it, bail out.
2869 	 */
2870 	if ((flags & MAP_SHARED) != 0) {
2871 		if ((fp->f_flag & FWRITE) != 0)
2872 			maxprot |= VM_PROT_WRITE;
2873 		else if ((prot & VM_PROT_WRITE) != 0)
2874 			return (EACCES);
2875 	} else {
2876 		maxprot |= VM_PROT_WRITE;
2877 		cap_maxprot |= VM_PROT_WRITE;
2878 	}
2879 	maxprot &= cap_maxprot;
2880 
2881 	/*
2882 	 * For regular files and shared memory, POSIX requires that
2883 	 * the value of foff be a legitimate offset within the data
2884 	 * object.  In particular, negative offsets are invalid.
2885 	 * Blocking negative offsets and overflows here avoids
2886 	 * possible wraparound or user-level access into reserved
2887 	 * ranges of the data object later.  In contrast, POSIX does
2888 	 * not dictate how offsets are used by device drivers, so in
2889 	 * the case of a device mapping a negative offset is passed
2890 	 * on.
2891 	 */
2892 	if (
2893 #ifdef _LP64
2894 	    size > OFF_MAX ||
2895 #endif
2896 	    foff > OFF_MAX - size)
2897 		return (EINVAL);
2898 
2899 	writecounted = FALSE;
2900 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2901 	    &foff, &object, &writecounted);
2902 	if (error != 0)
2903 		return (error);
2904 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2905 	    foff, writecounted, td);
2906 	if (error != 0) {
2907 		/*
2908 		 * If this mapping was accounted for in the vnode's
2909 		 * writecount, then undo that now.
2910 		 */
2911 		if (writecounted)
2912 			vm_pager_release_writecount(object, 0, size);
2913 		vm_object_deallocate(object);
2914 	}
2915 #ifdef HWPMC_HOOKS
2916 	/* Inform hwpmc(4) if an executable is being mapped. */
2917 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2918 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2919 			pkm.pm_file = vp;
2920 			pkm.pm_address = (uintptr_t) *addr;
2921 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2922 		}
2923 	}
2924 #endif
2925 	return (error);
2926 }
2927 
2928 void
2929 vn_fsid(struct vnode *vp, struct vattr *va)
2930 {
2931 	fsid_t *f;
2932 
2933 	f = &vp->v_mount->mnt_stat.f_fsid;
2934 	va->va_fsid = (uint32_t)f->val[1];
2935 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2936 	va->va_fsid += (uint32_t)f->val[0];
2937 }
2938 
2939 int
2940 vn_fsync_buf(struct vnode *vp, int waitfor)
2941 {
2942 	struct buf *bp, *nbp;
2943 	struct bufobj *bo;
2944 	struct mount *mp;
2945 	int error, maxretry;
2946 
2947 	error = 0;
2948 	maxretry = 10000;     /* large, arbitrarily chosen */
2949 	mp = NULL;
2950 	if (vp->v_type == VCHR) {
2951 		VI_LOCK(vp);
2952 		mp = vp->v_rdev->si_mountpt;
2953 		VI_UNLOCK(vp);
2954 	}
2955 	bo = &vp->v_bufobj;
2956 	BO_LOCK(bo);
2957 loop1:
2958 	/*
2959 	 * MARK/SCAN initialization to avoid infinite loops.
2960 	 */
2961         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2962 		bp->b_vflags &= ~BV_SCANNED;
2963 		bp->b_error = 0;
2964 	}
2965 
2966 	/*
2967 	 * Flush all dirty buffers associated with a vnode.
2968 	 */
2969 loop2:
2970 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2971 		if ((bp->b_vflags & BV_SCANNED) != 0)
2972 			continue;
2973 		bp->b_vflags |= BV_SCANNED;
2974 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2975 			if (waitfor != MNT_WAIT)
2976 				continue;
2977 			if (BUF_LOCK(bp,
2978 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2979 			    BO_LOCKPTR(bo)) != 0) {
2980 				BO_LOCK(bo);
2981 				goto loop1;
2982 			}
2983 			BO_LOCK(bo);
2984 		}
2985 		BO_UNLOCK(bo);
2986 		KASSERT(bp->b_bufobj == bo,
2987 		    ("bp %p wrong b_bufobj %p should be %p",
2988 		    bp, bp->b_bufobj, bo));
2989 		if ((bp->b_flags & B_DELWRI) == 0)
2990 			panic("fsync: not dirty");
2991 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2992 			vfs_bio_awrite(bp);
2993 		} else {
2994 			bremfree(bp);
2995 			bawrite(bp);
2996 		}
2997 		if (maxretry < 1000)
2998 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2999 		BO_LOCK(bo);
3000 		goto loop2;
3001 	}
3002 
3003 	/*
3004 	 * If synchronous the caller expects us to completely resolve all
3005 	 * dirty buffers in the system.  Wait for in-progress I/O to
3006 	 * complete (which could include background bitmap writes), then
3007 	 * retry if dirty blocks still exist.
3008 	 */
3009 	if (waitfor == MNT_WAIT) {
3010 		bufobj_wwait(bo, 0, 0);
3011 		if (bo->bo_dirty.bv_cnt > 0) {
3012 			/*
3013 			 * If we are unable to write any of these buffers
3014 			 * then we fail now rather than trying endlessly
3015 			 * to write them out.
3016 			 */
3017 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
3018 				if ((error = bp->b_error) != 0)
3019 					break;
3020 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
3021 			    (error == 0 && --maxretry >= 0))
3022 				goto loop1;
3023 			if (error == 0)
3024 				error = EAGAIN;
3025 		}
3026 	}
3027 	BO_UNLOCK(bo);
3028 	if (error != 0)
3029 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3030 
3031 	return (error);
3032 }
3033 
3034 /*
3035  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
3036  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3037  * to do the actual copy.
3038  * vn_generic_copy_file_range() is factored out, so it can be called
3039  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3040  * different file systems.
3041  */
3042 int
3043 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3044     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3045     struct ucred *outcred, struct thread *fsize_td)
3046 {
3047 	int error;
3048 	size_t len;
3049 	uint64_t uval;
3050 
3051 	len = *lenp;
3052 	*lenp = 0;		/* For error returns. */
3053 	error = 0;
3054 
3055 	/* Do some sanity checks on the arguments. */
3056 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
3057 		error = EISDIR;
3058 	else if (*inoffp < 0 || *outoffp < 0 ||
3059 	    invp->v_type != VREG || outvp->v_type != VREG)
3060 		error = EINVAL;
3061 	if (error != 0)
3062 		goto out;
3063 
3064 	/* Ensure offset + len does not wrap around. */
3065 	uval = *inoffp;
3066 	uval += len;
3067 	if (uval > INT64_MAX)
3068 		len = INT64_MAX - *inoffp;
3069 	uval = *outoffp;
3070 	uval += len;
3071 	if (uval > INT64_MAX)
3072 		len = INT64_MAX - *outoffp;
3073 	if (len == 0)
3074 		goto out;
3075 
3076 	/*
3077 	 * If the two vnode are for the same file system, call
3078 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3079 	 * which can handle copies across multiple file systems.
3080 	 */
3081 	*lenp = len;
3082 	if (invp->v_mount == outvp->v_mount)
3083 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
3084 		    lenp, flags, incred, outcred, fsize_td);
3085 	else
3086 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
3087 		    outoffp, lenp, flags, incred, outcred, fsize_td);
3088 out:
3089 	return (error);
3090 }
3091 
3092 /*
3093  * Test len bytes of data starting at dat for all bytes == 0.
3094  * Return true if all bytes are zero, false otherwise.
3095  * Expects dat to be well aligned.
3096  */
3097 static bool
3098 mem_iszero(void *dat, int len)
3099 {
3100 	int i;
3101 	const u_int *p;
3102 	const char *cp;
3103 
3104 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
3105 		if (len >= sizeof(*p)) {
3106 			if (*p != 0)
3107 				return (false);
3108 		} else {
3109 			cp = (const char *)p;
3110 			for (i = 0; i < len; i++, cp++)
3111 				if (*cp != '\0')
3112 					return (false);
3113 		}
3114 	}
3115 	return (true);
3116 }
3117 
3118 /*
3119  * Look for a hole in the output file and, if found, adjust *outoffp
3120  * and *xferp to skip past the hole.
3121  * *xferp is the entire hole length to be written and xfer2 is how many bytes
3122  * to be written as 0's upon return.
3123  */
3124 static off_t
3125 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3126     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3127 {
3128 	int error;
3129 	off_t delta;
3130 
3131 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3132 		*dataoffp = *outoffp;
3133 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3134 		    curthread);
3135 		if (error == 0) {
3136 			*holeoffp = *dataoffp;
3137 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3138 			    curthread);
3139 		}
3140 		if (error != 0 || *holeoffp == *dataoffp) {
3141 			/*
3142 			 * Since outvp is unlocked, it may be possible for
3143 			 * another thread to do a truncate(), lseek(), write()
3144 			 * creating a hole at startoff between the above
3145 			 * VOP_IOCTL() calls, if the other thread does not do
3146 			 * rangelocking.
3147 			 * If that happens, *holeoffp == *dataoffp and finding
3148 			 * the hole has failed, so disable vn_skip_hole().
3149 			 */
3150 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3151 			return (xfer2);
3152 		}
3153 		KASSERT(*dataoffp >= *outoffp,
3154 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3155 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3156 		KASSERT(*holeoffp > *dataoffp,
3157 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3158 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3159 	}
3160 
3161 	/*
3162 	 * If there is a hole before the data starts, advance *outoffp and
3163 	 * *xferp past the hole.
3164 	 */
3165 	if (*dataoffp > *outoffp) {
3166 		delta = *dataoffp - *outoffp;
3167 		if (delta >= *xferp) {
3168 			/* Entire *xferp is a hole. */
3169 			*outoffp += *xferp;
3170 			*xferp = 0;
3171 			return (0);
3172 		}
3173 		*xferp -= delta;
3174 		*outoffp += delta;
3175 		xfer2 = MIN(xfer2, *xferp);
3176 	}
3177 
3178 	/*
3179 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3180 	 * that the write ends at the start of the hole.
3181 	 * *holeoffp should always be greater than *outoffp, but for the
3182 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3183 	 * value.
3184 	 */
3185 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3186 		xfer2 = *holeoffp - *outoffp;
3187 	return (xfer2);
3188 }
3189 
3190 /*
3191  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3192  * dat is a maximum of blksize in length and can be written repeatedly in
3193  * the chunk.
3194  * If growfile == true, just grow the file via vn_truncate_locked() instead
3195  * of doing actual writes.
3196  * If checkhole == true, a hole is being punched, so skip over any hole
3197  * already in the output file.
3198  */
3199 static int
3200 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3201     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3202 {
3203 	struct mount *mp;
3204 	off_t dataoff, holeoff, xfer2;
3205 	int error;
3206 
3207 	/*
3208 	 * Loop around doing writes of blksize until write has been completed.
3209 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3210 	 * done for each iteration, since the xfer argument can be very
3211 	 * large if there is a large hole to punch in the output file.
3212 	 */
3213 	error = 0;
3214 	holeoff = 0;
3215 	do {
3216 		xfer2 = MIN(xfer, blksize);
3217 		if (checkhole) {
3218 			/*
3219 			 * Punching a hole.  Skip writing if there is
3220 			 * already a hole in the output file.
3221 			 */
3222 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3223 			    &dataoff, &holeoff, cred);
3224 			if (xfer == 0)
3225 				break;
3226 			if (holeoff < 0)
3227 				checkhole = false;
3228 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3229 			    (intmax_t)xfer2));
3230 		}
3231 		bwillwrite();
3232 		mp = NULL;
3233 		error = vn_start_write(outvp, &mp, V_WAIT);
3234 		if (error != 0)
3235 			break;
3236 		if (growfile) {
3237 			error = vn_lock(outvp, LK_EXCLUSIVE);
3238 			if (error == 0) {
3239 				error = vn_truncate_locked(outvp, outoff + xfer,
3240 				    false, cred);
3241 				VOP_UNLOCK(outvp);
3242 			}
3243 		} else {
3244 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3245 			if (error == 0) {
3246 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3247 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3248 				    curthread->td_ucred, cred, NULL, curthread);
3249 				outoff += xfer2;
3250 				xfer -= xfer2;
3251 				VOP_UNLOCK(outvp);
3252 			}
3253 		}
3254 		if (mp != NULL)
3255 			vn_finished_write(mp);
3256 	} while (!growfile && xfer > 0 && error == 0);
3257 	return (error);
3258 }
3259 
3260 /*
3261  * Copy a byte range of one file to another.  This function can handle the
3262  * case where invp and outvp are on different file systems.
3263  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3264  * is no better file system specific way to do it.
3265  */
3266 int
3267 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3268     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3269     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3270 {
3271 	struct mount *mp;
3272 	off_t startoff, endoff, xfer, xfer2;
3273 	u_long blksize;
3274 	int error, interrupted;
3275 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3276 	ssize_t aresid, r = 0;
3277 	size_t copylen, len, savlen;
3278 	off_t insize, outsize;
3279 	char *dat;
3280 	long holein, holeout;
3281 	struct timespec curts, endts;
3282 
3283 	holein = holeout = 0;
3284 	savlen = len = *lenp;
3285 	error = 0;
3286 	interrupted = 0;
3287 	dat = NULL;
3288 
3289 	error = vn_lock(invp, LK_SHARED);
3290 	if (error != 0)
3291 		goto out;
3292 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3293 		holein = 0;
3294 	if (holein > 0)
3295 		error = vn_getsize_locked(invp, &insize, incred);
3296 	VOP_UNLOCK(invp);
3297 	if (error != 0)
3298 		goto out;
3299 
3300 	mp = NULL;
3301 	error = vn_start_write(outvp, &mp, V_WAIT);
3302 	if (error == 0)
3303 		error = vn_lock(outvp, LK_EXCLUSIVE);
3304 	if (error == 0) {
3305 		/*
3306 		 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3307 		 * now that outvp is locked.
3308 		 */
3309 		if (fsize_td != NULL) {
3310 			struct uio io;
3311 
3312 			io.uio_offset = *outoffp;
3313 			io.uio_resid = len;
3314 			error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3315 			len = savlen = io.uio_resid;
3316 			/*
3317 			 * No need to call vn_rlimit_fsizex_res before return,
3318 			 * since the uio is local.
3319 			 */
3320 		}
3321 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3322 			holeout = 0;
3323 		/*
3324 		 * Holes that are past EOF do not need to be written as a block
3325 		 * of zero bytes.  So, truncate the output file as far as
3326 		 * possible and then use size to decide if writing 0
3327 		 * bytes is necessary in the loop below.
3328 		 */
3329 		if (error == 0)
3330 			error = vn_getsize_locked(outvp, &outsize, outcred);
3331 		if (error == 0 && outsize > *outoffp && outsize <= *outoffp + len) {
3332 #ifdef MAC
3333 			error = mac_vnode_check_write(curthread->td_ucred,
3334 			    outcred, outvp);
3335 			if (error == 0)
3336 #endif
3337 				error = vn_truncate_locked(outvp, *outoffp,
3338 				    false, outcred);
3339 			if (error == 0)
3340 				outsize = *outoffp;
3341 		}
3342 		VOP_UNLOCK(outvp);
3343 	}
3344 	if (mp != NULL)
3345 		vn_finished_write(mp);
3346 	if (error != 0)
3347 		goto out;
3348 
3349 	if (holein == 0 && holeout > 0) {
3350 		/*
3351 		 * For this special case, the input data will be scanned
3352 		 * for blocks of all 0 bytes.  For these blocks, the
3353 		 * write can be skipped for the output file to create
3354 		 * an unallocated region.
3355 		 * Therefore, use the appropriate size for the output file.
3356 		 */
3357 		blksize = holeout;
3358 		if (blksize <= 512) {
3359 			/*
3360 			 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3361 			 * of 512, although it actually only creates
3362 			 * unallocated regions for blocks >= f_iosize.
3363 			 */
3364 			blksize = outvp->v_mount->mnt_stat.f_iosize;
3365 		}
3366 	} else {
3367 		/*
3368 		 * Use the larger of the two f_iosize values.  If they are
3369 		 * not the same size, one will normally be an exact multiple of
3370 		 * the other, since they are both likely to be a power of 2.
3371 		 */
3372 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3373 		    outvp->v_mount->mnt_stat.f_iosize);
3374 	}
3375 
3376 	/* Clip to sane limits. */
3377 	if (blksize < 4096)
3378 		blksize = 4096;
3379 	else if (blksize > maxphys)
3380 		blksize = maxphys;
3381 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3382 
3383 	/*
3384 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3385 	 * to find holes.  Otherwise, just scan the read block for all 0s
3386 	 * in the inner loop where the data copying is done.
3387 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3388 	 * support holes on the server, but do not support FIOSEEKHOLE.
3389 	 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3390 	 * that this function should return after 1second with a partial
3391 	 * completion.
3392 	 */
3393 	if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3394 		getnanouptime(&endts);
3395 		endts.tv_sec++;
3396 	} else
3397 		timespecclear(&endts);
3398 	holetoeof = eof = false;
3399 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3400 		endoff = 0;			/* To shut up compilers. */
3401 		cantseek = true;
3402 		startoff = *inoffp;
3403 		copylen = len;
3404 
3405 		/*
3406 		 * Find the next data area.  If there is just a hole to EOF,
3407 		 * FIOSEEKDATA should fail with ENXIO.
3408 		 * (I do not know if any file system will report a hole to
3409 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3410 		 *  will fail for those file systems.)
3411 		 *
3412 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3413 		 * the code just falls through to the inner copy loop.
3414 		 */
3415 		error = EINVAL;
3416 		if (holein > 0) {
3417 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3418 			    incred, curthread);
3419 			if (error == ENXIO) {
3420 				startoff = endoff = insize;
3421 				eof = holetoeof = true;
3422 				error = 0;
3423 			}
3424 		}
3425 		if (error == 0 && !holetoeof) {
3426 			endoff = startoff;
3427 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3428 			    incred, curthread);
3429 			/*
3430 			 * Since invp is unlocked, it may be possible for
3431 			 * another thread to do a truncate(), lseek(), write()
3432 			 * creating a hole at startoff between the above
3433 			 * VOP_IOCTL() calls, if the other thread does not do
3434 			 * rangelocking.
3435 			 * If that happens, startoff == endoff and finding
3436 			 * the hole has failed, so set an error.
3437 			 */
3438 			if (error == 0 && startoff == endoff)
3439 				error = EINVAL; /* Any error. Reset to 0. */
3440 		}
3441 		if (error == 0) {
3442 			if (startoff > *inoffp) {
3443 				/* Found hole before data block. */
3444 				xfer = MIN(startoff - *inoffp, len);
3445 				if (*outoffp < outsize) {
3446 					/* Must write 0s to punch hole. */
3447 					xfer2 = MIN(outsize - *outoffp,
3448 					    xfer);
3449 					memset(dat, 0, MIN(xfer2, blksize));
3450 					error = vn_write_outvp(outvp, dat,
3451 					    *outoffp, xfer2, blksize, false,
3452 					    holeout > 0, outcred);
3453 				}
3454 
3455 				if (error == 0 && *outoffp + xfer >
3456 				    outsize && (xfer == len || holetoeof)) {
3457 					/* Grow output file (hole at end). */
3458 					error = vn_write_outvp(outvp, dat,
3459 					    *outoffp, xfer, blksize, true,
3460 					    false, outcred);
3461 				}
3462 				if (error == 0) {
3463 					*inoffp += xfer;
3464 					*outoffp += xfer;
3465 					len -= xfer;
3466 					if (len < savlen) {
3467 						interrupted = sig_intr();
3468 						if (timespecisset(&endts) &&
3469 						    interrupted == 0) {
3470 							getnanouptime(&curts);
3471 							if (timespeccmp(&curts,
3472 							    &endts, >=))
3473 								interrupted =
3474 								    EINTR;
3475 						}
3476 					}
3477 				}
3478 			}
3479 			copylen = MIN(len, endoff - startoff);
3480 			cantseek = false;
3481 		} else {
3482 			cantseek = true;
3483 			startoff = *inoffp;
3484 			copylen = len;
3485 			error = 0;
3486 		}
3487 
3488 		xfer = blksize;
3489 		if (cantseek) {
3490 			/*
3491 			 * Set first xfer to end at a block boundary, so that
3492 			 * holes are more likely detected in the loop below via
3493 			 * the for all bytes 0 method.
3494 			 */
3495 			xfer -= (*inoffp % blksize);
3496 		}
3497 		/* Loop copying the data block. */
3498 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3499 			if (copylen < xfer)
3500 				xfer = copylen;
3501 			error = vn_lock(invp, LK_SHARED);
3502 			if (error != 0)
3503 				goto out;
3504 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3505 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3506 			    curthread->td_ucred, incred, &aresid,
3507 			    curthread);
3508 			VOP_UNLOCK(invp);
3509 			lastblock = false;
3510 			if (error == 0 && aresid > 0) {
3511 				/* Stop the copy at EOF on the input file. */
3512 				xfer -= aresid;
3513 				eof = true;
3514 				lastblock = true;
3515 			}
3516 			if (error == 0) {
3517 				/*
3518 				 * Skip the write for holes past the initial EOF
3519 				 * of the output file, unless this is the last
3520 				 * write of the output file at EOF.
3521 				 */
3522 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3523 				    false;
3524 				if (xfer == len)
3525 					lastblock = true;
3526 				if (!cantseek || *outoffp < outsize ||
3527 				    lastblock || !readzeros)
3528 					error = vn_write_outvp(outvp, dat,
3529 					    *outoffp, xfer, blksize,
3530 					    readzeros && lastblock &&
3531 					    *outoffp >= outsize, false,
3532 					    outcred);
3533 				if (error == 0) {
3534 					*inoffp += xfer;
3535 					startoff += xfer;
3536 					*outoffp += xfer;
3537 					copylen -= xfer;
3538 					len -= xfer;
3539 					if (len < savlen) {
3540 						interrupted = sig_intr();
3541 						if (timespecisset(&endts) &&
3542 						    interrupted == 0) {
3543 							getnanouptime(&curts);
3544 							if (timespeccmp(&curts,
3545 							    &endts, >=))
3546 								interrupted =
3547 								    EINTR;
3548 						}
3549 					}
3550 				}
3551 			}
3552 			xfer = blksize;
3553 		}
3554 	}
3555 out:
3556 	*lenp = savlen - len;
3557 	free(dat, M_TEMP);
3558 	return (error);
3559 }
3560 
3561 static int
3562 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3563 {
3564 	struct mount *mp;
3565 	struct vnode *vp;
3566 	off_t olen, ooffset;
3567 	int error;
3568 #ifdef AUDIT
3569 	int audited_vnode1 = 0;
3570 #endif
3571 
3572 	vp = fp->f_vnode;
3573 	if (vp->v_type != VREG)
3574 		return (ENODEV);
3575 
3576 	/* Allocating blocks may take a long time, so iterate. */
3577 	for (;;) {
3578 		olen = len;
3579 		ooffset = offset;
3580 
3581 		bwillwrite();
3582 		mp = NULL;
3583 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
3584 		if (error != 0)
3585 			break;
3586 		error = vn_lock(vp, LK_EXCLUSIVE);
3587 		if (error != 0) {
3588 			vn_finished_write(mp);
3589 			break;
3590 		}
3591 #ifdef AUDIT
3592 		if (!audited_vnode1) {
3593 			AUDIT_ARG_VNODE1(vp);
3594 			audited_vnode1 = 1;
3595 		}
3596 #endif
3597 #ifdef MAC
3598 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3599 		if (error == 0)
3600 #endif
3601 			error = VOP_ALLOCATE(vp, &offset, &len, 0,
3602 			    td->td_ucred);
3603 		VOP_UNLOCK(vp);
3604 		vn_finished_write(mp);
3605 
3606 		if (olen + ooffset != offset + len) {
3607 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3608 			    ooffset, olen, offset, len);
3609 		}
3610 		if (error != 0 || len == 0)
3611 			break;
3612 		KASSERT(olen > len, ("Iteration did not make progress?"));
3613 		maybe_yield();
3614 	}
3615 
3616 	return (error);
3617 }
3618 
3619 static int
3620 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3621     int ioflag, struct ucred *cred, struct ucred *active_cred,
3622     struct ucred *file_cred)
3623 {
3624 	struct mount *mp;
3625 	void *rl_cookie;
3626 	off_t off, len;
3627 	int error;
3628 #ifdef AUDIT
3629 	bool audited_vnode1 = false;
3630 #endif
3631 
3632 	rl_cookie = NULL;
3633 	error = 0;
3634 	mp = NULL;
3635 	off = *offset;
3636 	len = *length;
3637 
3638 	if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3639 		rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3640 	while (len > 0 && error == 0) {
3641 		/*
3642 		 * Try to deallocate the longest range in one pass.
3643 		 * In case a pass takes too long to be executed, it returns
3644 		 * partial result. The residue will be proceeded in the next
3645 		 * pass.
3646 		 */
3647 
3648 		if ((ioflag & IO_NODELOCKED) == 0) {
3649 			bwillwrite();
3650 			if ((error = vn_start_write(vp, &mp,
3651 			    V_WAIT | V_PCATCH)) != 0)
3652 				goto out;
3653 			vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3654 		}
3655 #ifdef AUDIT
3656 		if (!audited_vnode1) {
3657 			AUDIT_ARG_VNODE1(vp);
3658 			audited_vnode1 = true;
3659 		}
3660 #endif
3661 
3662 #ifdef MAC
3663 		if ((ioflag & IO_NOMACCHECK) == 0)
3664 			error = mac_vnode_check_write(active_cred, file_cred,
3665 			    vp);
3666 #endif
3667 		if (error == 0)
3668 			error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3669 			    cred);
3670 
3671 		if ((ioflag & IO_NODELOCKED) == 0) {
3672 			VOP_UNLOCK(vp);
3673 			if (mp != NULL) {
3674 				vn_finished_write(mp);
3675 				mp = NULL;
3676 			}
3677 		}
3678 		if (error == 0 && len != 0)
3679 			maybe_yield();
3680 	}
3681 out:
3682 	if (rl_cookie != NULL)
3683 		vn_rangelock_unlock(vp, rl_cookie);
3684 	*offset = off;
3685 	*length = len;
3686 	return (error);
3687 }
3688 
3689 /*
3690  * This function is supposed to be used in the situations where the deallocation
3691  * is not triggered by a user request.
3692  */
3693 int
3694 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3695     int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3696 {
3697 	struct ucred *cred;
3698 
3699 	if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3700 	    flags != 0)
3701 		return (EINVAL);
3702 	if (vp->v_type != VREG)
3703 		return (ENODEV);
3704 
3705 	cred = file_cred != NOCRED ? file_cred : active_cred;
3706 	return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3707 	    active_cred, file_cred));
3708 }
3709 
3710 static int
3711 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3712     struct ucred *active_cred, struct thread *td)
3713 {
3714 	int error;
3715 	struct vnode *vp;
3716 	int ioflag;
3717 
3718 	KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd"));
3719 	KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
3720 	    ("vn_fspacectl: non-zero flags"));
3721 	KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
3722 	    ("vn_fspacectl: offset/length overflow or underflow"));
3723 	vp = fp->f_vnode;
3724 
3725 	if (vp->v_type != VREG)
3726 		return (ENODEV);
3727 
3728 	ioflag = get_write_ioflag(fp);
3729 
3730 	switch (cmd) {
3731 	case SPACECTL_DEALLOC:
3732 		error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3733 		    active_cred, active_cred, fp->f_cred);
3734 		break;
3735 	default:
3736 		panic("vn_fspacectl: unknown cmd %d", cmd);
3737 	}
3738 
3739 	return (error);
3740 }
3741 
3742 static u_long vn_lock_pair_pause_cnt;
3743 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3744     &vn_lock_pair_pause_cnt, 0,
3745     "Count of vn_lock_pair deadlocks");
3746 
3747 u_int vn_lock_pair_pause_max;
3748 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3749     &vn_lock_pair_pause_max, 0,
3750     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3751 
3752 static void
3753 vn_lock_pair_pause(const char *wmesg)
3754 {
3755 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3756 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3757 }
3758 
3759 /*
3760  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3761  * vp1_locked indicates whether vp1 is locked; if not, vp1 must be
3762  * unlocked.  Same for vp2 and vp2_locked.  One of the vnodes can be
3763  * NULL.
3764  *
3765  * The function returns with both vnodes exclusively or shared locked,
3766  * according to corresponding lkflags, and guarantees that it does not
3767  * create lock order reversal with other threads during its execution.
3768  * Both vnodes could be unlocked temporary (and reclaimed).
3769  *
3770  * If requesting shared locking, locked vnode lock must not be recursed.
3771  */
3772 void
3773 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1,
3774     struct vnode *vp2, bool vp2_locked, int lkflags2)
3775 {
3776 	int error;
3777 
3778 	MPASS(lkflags1 == LK_SHARED || lkflags1 == LK_EXCLUSIVE);
3779 	MPASS(lkflags2 == LK_SHARED || lkflags2 == LK_EXCLUSIVE);
3780 
3781 	if (vp1 == NULL && vp2 == NULL)
3782 		return;
3783 
3784 	if (vp1 != NULL) {
3785 		if (lkflags1 == LK_SHARED &&
3786 		    (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
3787 			lkflags1 = LK_EXCLUSIVE;
3788 		if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) {
3789 			ASSERT_VOP_LOCKED(vp1, "vp1");
3790 			if (lkflags1 == LK_EXCLUSIVE) {
3791 				VOP_UNLOCK(vp1);
3792 				ASSERT_VOP_UNLOCKED(vp1,
3793 				    "vp1 shared recursed");
3794 				vp1_locked = false;
3795 			}
3796 		} else if (!vp1_locked)
3797 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3798 	} else {
3799 		vp1_locked = true;
3800 	}
3801 
3802 	if (vp2 != NULL) {
3803 		if (lkflags2 == LK_SHARED &&
3804 		    (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
3805 			lkflags2 = LK_EXCLUSIVE;
3806 		if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) {
3807 			ASSERT_VOP_LOCKED(vp2, "vp2");
3808 			if (lkflags2 == LK_EXCLUSIVE) {
3809 				VOP_UNLOCK(vp2);
3810 				ASSERT_VOP_UNLOCKED(vp2,
3811 				    "vp2 shared recursed");
3812 				vp2_locked = false;
3813 			}
3814 		} else if (!vp2_locked)
3815 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3816 	} else {
3817 		vp2_locked = true;
3818 	}
3819 
3820 	if (!vp1_locked && !vp2_locked) {
3821 		vn_lock(vp1, lkflags1 | LK_RETRY);
3822 		vp1_locked = true;
3823 	}
3824 
3825 	while (!vp1_locked || !vp2_locked) {
3826 		if (vp1_locked && vp2 != NULL) {
3827 			if (vp1 != NULL) {
3828 				error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT,
3829 				    __FILE__, __LINE__);
3830 				if (error == 0)
3831 					break;
3832 				VOP_UNLOCK(vp1);
3833 				vp1_locked = false;
3834 				vn_lock_pair_pause("vlp1");
3835 			}
3836 			vn_lock(vp2, lkflags2 | LK_RETRY);
3837 			vp2_locked = true;
3838 		}
3839 		if (vp2_locked && vp1 != NULL) {
3840 			if (vp2 != NULL) {
3841 				error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT,
3842 				    __FILE__, __LINE__);
3843 				if (error == 0)
3844 					break;
3845 				VOP_UNLOCK(vp2);
3846 				vp2_locked = false;
3847 				vn_lock_pair_pause("vlp2");
3848 			}
3849 			vn_lock(vp1, lkflags1 | LK_RETRY);
3850 			vp1_locked = true;
3851 		}
3852 	}
3853 	if (vp1 != NULL) {
3854 		if (lkflags1 == LK_EXCLUSIVE)
3855 			ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3856 		else
3857 			ASSERT_VOP_LOCKED(vp1, "vp1 ret");
3858 	}
3859 	if (vp2 != NULL) {
3860 		if (lkflags2 == LK_EXCLUSIVE)
3861 			ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3862 		else
3863 			ASSERT_VOP_LOCKED(vp2, "vp2 ret");
3864 	}
3865 }
3866 
3867 int
3868 vn_lktype_write(struct mount *mp, struct vnode *vp)
3869 {
3870 	if (MNT_SHARED_WRITES(mp) ||
3871 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
3872 		return (LK_SHARED);
3873 	return (LK_EXCLUSIVE);
3874 }
3875