xref: /freebsd/sys/kern/vfs_vnops.c (revision 2008043f386721d58158e37e0d7e50df8095942d)
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 #include "opt_hwpmc_hooks.h"
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/disk.h>
51 #include <sys/fail.h>
52 #include <sys/fcntl.h>
53 #include <sys/file.h>
54 #include <sys/kdb.h>
55 #include <sys/ktr.h>
56 #include <sys/stat.h>
57 #include <sys/priv.h>
58 #include <sys/proc.h>
59 #include <sys/limits.h>
60 #include <sys/lock.h>
61 #include <sys/mman.h>
62 #include <sys/mount.h>
63 #include <sys/mutex.h>
64 #include <sys/namei.h>
65 #include <sys/vnode.h>
66 #include <sys/dirent.h>
67 #include <sys/bio.h>
68 #include <sys/buf.h>
69 #include <sys/filio.h>
70 #include <sys/resourcevar.h>
71 #include <sys/rwlock.h>
72 #include <sys/prng.h>
73 #include <sys/sx.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/sysctl.h>
76 #include <sys/ttycom.h>
77 #include <sys/conf.h>
78 #include <sys/syslog.h>
79 #include <sys/unistd.h>
80 #include <sys/user.h>
81 #include <sys/ktrace.h>
82 
83 #include <security/audit/audit.h>
84 #include <security/mac/mac_framework.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_extern.h>
88 #include <vm/pmap.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pager.h>
93 
94 #ifdef HWPMC_HOOKS
95 #include <sys/pmckern.h>
96 #endif
97 
98 static fo_rdwr_t	vn_read;
99 static fo_rdwr_t	vn_write;
100 static fo_rdwr_t	vn_io_fault;
101 static fo_truncate_t	vn_truncate;
102 static fo_ioctl_t	vn_ioctl;
103 static fo_poll_t	vn_poll;
104 static fo_kqfilter_t	vn_kqfilter;
105 static fo_close_t	vn_closefile;
106 static fo_mmap_t	vn_mmap;
107 static fo_fallocate_t	vn_fallocate;
108 static fo_fspacectl_t	vn_fspacectl;
109 
110 struct 	fileops vnops = {
111 	.fo_read = vn_io_fault,
112 	.fo_write = vn_io_fault,
113 	.fo_truncate = vn_truncate,
114 	.fo_ioctl = vn_ioctl,
115 	.fo_poll = vn_poll,
116 	.fo_kqfilter = vn_kqfilter,
117 	.fo_stat = vn_statfile,
118 	.fo_close = vn_closefile,
119 	.fo_chmod = vn_chmod,
120 	.fo_chown = vn_chown,
121 	.fo_sendfile = vn_sendfile,
122 	.fo_seek = vn_seek,
123 	.fo_fill_kinfo = vn_fill_kinfo,
124 	.fo_mmap = vn_mmap,
125 	.fo_fallocate = vn_fallocate,
126 	.fo_fspacectl = vn_fspacectl,
127 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
128 };
129 
130 const u_int io_hold_cnt = 16;
131 static int vn_io_fault_enable = 1;
132 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
133     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
134 static int vn_io_fault_prefault = 0;
135 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
136     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
137 static int vn_io_pgcache_read_enable = 1;
138 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
139     &vn_io_pgcache_read_enable, 0,
140     "Enable copying from page cache for reads, avoiding fs");
141 static u_long vn_io_faults_cnt;
142 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
143     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
144 
145 static int vfs_allow_read_dir = 0;
146 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
147     &vfs_allow_read_dir, 0,
148     "Enable read(2) of directory by root for filesystems that support it");
149 
150 /*
151  * Returns true if vn_io_fault mode of handling the i/o request should
152  * be used.
153  */
154 static bool
155 do_vn_io_fault(struct vnode *vp, struct uio *uio)
156 {
157 	struct mount *mp;
158 
159 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
160 	    (mp = vp->v_mount) != NULL &&
161 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
162 }
163 
164 /*
165  * Structure used to pass arguments to vn_io_fault1(), to do either
166  * file- or vnode-based I/O calls.
167  */
168 struct vn_io_fault_args {
169 	enum {
170 		VN_IO_FAULT_FOP,
171 		VN_IO_FAULT_VOP
172 	} kind;
173 	struct ucred *cred;
174 	int flags;
175 	union {
176 		struct fop_args_tag {
177 			struct file *fp;
178 			fo_rdwr_t *doio;
179 		} fop_args;
180 		struct vop_args_tag {
181 			struct vnode *vp;
182 		} vop_args;
183 	} args;
184 };
185 
186 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
187     struct vn_io_fault_args *args, struct thread *td);
188 
189 int
190 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
191 {
192 	struct thread *td = curthread;
193 
194 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
195 }
196 
197 static uint64_t
198 open2nameif(int fmode, u_int vn_open_flags)
199 {
200 	uint64_t res;
201 
202 	res = ISOPEN | LOCKLEAF;
203 	if ((fmode & O_RESOLVE_BENEATH) != 0)
204 		res |= RBENEATH;
205 	if ((fmode & O_EMPTY_PATH) != 0)
206 		res |= EMPTYPATH;
207 	if ((fmode & FREAD) != 0)
208 		res |= OPENREAD;
209 	if ((fmode & FWRITE) != 0)
210 		res |= OPENWRITE;
211 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
212 		res |= AUDITVNODE1;
213 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
214 		res |= NOCAPCHECK;
215 	if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0)
216 		res |= WANTIOCTLCAPS;
217 	return (res);
218 }
219 
220 /*
221  * Common code for vnode open operations via a name lookup.
222  * Lookup the vnode and invoke VOP_CREATE if needed.
223  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
224  *
225  * Note that this does NOT free nameidata for the successful case,
226  * due to the NDINIT being done elsewhere.
227  */
228 int
229 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
230     struct ucred *cred, struct file *fp)
231 {
232 	struct vnode *vp;
233 	struct mount *mp;
234 	struct vattr vat;
235 	struct vattr *vap = &vat;
236 	int fmode, error;
237 	bool first_open;
238 
239 restart:
240 	first_open = false;
241 	fmode = *flagp;
242 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
243 	    O_EXCL | O_DIRECTORY) ||
244 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
245 		return (EINVAL);
246 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
247 		ndp->ni_cnd.cn_nameiop = CREATE;
248 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
249 		/*
250 		 * Set NOCACHE to avoid flushing the cache when
251 		 * rolling in many files at once.
252 		 *
253 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
254 		 * exist despite NOCACHE.
255 		 */
256 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
257 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
258 			ndp->ni_cnd.cn_flags |= FOLLOW;
259 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
260 			bwillwrite();
261 		if ((error = namei(ndp)) != 0)
262 			return (error);
263 		if (ndp->ni_vp == NULL) {
264 			VATTR_NULL(vap);
265 			vap->va_type = VREG;
266 			vap->va_mode = cmode;
267 			if (fmode & O_EXCL)
268 				vap->va_vaflags |= VA_EXCLUSIVE;
269 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
270 				NDFREE_PNBUF(ndp);
271 				vput(ndp->ni_dvp);
272 				if ((error = vn_start_write(NULL, &mp,
273 				    V_XSLEEP | V_PCATCH)) != 0)
274 					return (error);
275 				NDREINIT(ndp);
276 				goto restart;
277 			}
278 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
279 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
280 #ifdef MAC
281 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
282 			    &ndp->ni_cnd, vap);
283 			if (error == 0)
284 #endif
285 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
286 				    &ndp->ni_cnd, vap);
287 			vp = ndp->ni_vp;
288 			if (error == 0 && (fmode & O_EXCL) != 0 &&
289 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
290 				VI_LOCK(vp);
291 				vp->v_iflag |= VI_FOPENING;
292 				VI_UNLOCK(vp);
293 				first_open = true;
294 			}
295 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
296 			    false);
297 			vn_finished_write(mp);
298 			if (error) {
299 				NDFREE_PNBUF(ndp);
300 				if (error == ERELOOKUP) {
301 					NDREINIT(ndp);
302 					goto restart;
303 				}
304 				return (error);
305 			}
306 			fmode &= ~O_TRUNC;
307 		} else {
308 			if (ndp->ni_dvp == ndp->ni_vp)
309 				vrele(ndp->ni_dvp);
310 			else
311 				vput(ndp->ni_dvp);
312 			ndp->ni_dvp = NULL;
313 			vp = ndp->ni_vp;
314 			if (fmode & O_EXCL) {
315 				error = EEXIST;
316 				goto bad;
317 			}
318 			if (vp->v_type == VDIR) {
319 				error = EISDIR;
320 				goto bad;
321 			}
322 			fmode &= ~O_CREAT;
323 		}
324 	} else {
325 		ndp->ni_cnd.cn_nameiop = LOOKUP;
326 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
327 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
328 		    FOLLOW;
329 		if ((fmode & FWRITE) == 0)
330 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
331 		if ((error = namei(ndp)) != 0)
332 			return (error);
333 		vp = ndp->ni_vp;
334 	}
335 	error = vn_open_vnode(vp, fmode, cred, curthread, fp);
336 	if (first_open) {
337 		VI_LOCK(vp);
338 		vp->v_iflag &= ~VI_FOPENING;
339 		wakeup(vp);
340 		VI_UNLOCK(vp);
341 	}
342 	if (error)
343 		goto bad;
344 	*flagp = fmode;
345 	return (0);
346 bad:
347 	NDFREE_PNBUF(ndp);
348 	vput(vp);
349 	*flagp = fmode;
350 	ndp->ni_vp = NULL;
351 	return (error);
352 }
353 
354 static int
355 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
356 {
357 	struct flock lf;
358 	int error, lock_flags, type;
359 
360 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
361 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
362 		return (0);
363 	KASSERT(fp != NULL, ("open with flock requires fp"));
364 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
365 		return (EOPNOTSUPP);
366 
367 	lock_flags = VOP_ISLOCKED(vp);
368 	VOP_UNLOCK(vp);
369 
370 	lf.l_whence = SEEK_SET;
371 	lf.l_start = 0;
372 	lf.l_len = 0;
373 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
374 	type = F_FLOCK;
375 	if ((fmode & FNONBLOCK) == 0)
376 		type |= F_WAIT;
377 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
378 		type |= F_FIRSTOPEN;
379 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
380 	if (error == 0)
381 		fp->f_flag |= FHASLOCK;
382 
383 	vn_lock(vp, lock_flags | LK_RETRY);
384 	return (error);
385 }
386 
387 /*
388  * Common code for vnode open operations once a vnode is located.
389  * Check permissions, and call the VOP_OPEN routine.
390  */
391 int
392 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
393     struct thread *td, struct file *fp)
394 {
395 	accmode_t accmode;
396 	int error;
397 
398 	if (vp->v_type == VLNK) {
399 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
400 			return (EMLINK);
401 	}
402 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
403 		return (ENOTDIR);
404 
405 	accmode = 0;
406 	if ((fmode & O_PATH) == 0) {
407 		if (vp->v_type == VSOCK)
408 			return (EOPNOTSUPP);
409 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
410 			if (vp->v_type == VDIR)
411 				return (EISDIR);
412 			accmode |= VWRITE;
413 		}
414 		if ((fmode & FREAD) != 0)
415 			accmode |= VREAD;
416 		if ((fmode & O_APPEND) && (fmode & FWRITE))
417 			accmode |= VAPPEND;
418 #ifdef MAC
419 		if ((fmode & O_CREAT) != 0)
420 			accmode |= VCREAT;
421 #endif
422 	}
423 	if ((fmode & FEXEC) != 0)
424 		accmode |= VEXEC;
425 #ifdef MAC
426 	if ((fmode & O_VERIFY) != 0)
427 		accmode |= VVERIFY;
428 	error = mac_vnode_check_open(cred, vp, accmode);
429 	if (error != 0)
430 		return (error);
431 
432 	accmode &= ~(VCREAT | VVERIFY);
433 #endif
434 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
435 		error = VOP_ACCESS(vp, accmode, cred, td);
436 		if (error != 0)
437 			return (error);
438 	}
439 	if ((fmode & O_PATH) != 0) {
440 		if (vp->v_type != VFIFO && vp->v_type != VSOCK &&
441 		    VOP_ACCESS(vp, VREAD, cred, td) == 0)
442 			fp->f_flag |= FKQALLOWED;
443 		return (0);
444 	}
445 
446 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
447 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
448 	error = VOP_OPEN(vp, fmode, cred, td, fp);
449 	if (error != 0)
450 		return (error);
451 
452 	error = vn_open_vnode_advlock(vp, fmode, fp);
453 	if (error == 0 && (fmode & FWRITE) != 0) {
454 		error = VOP_ADD_WRITECOUNT(vp, 1);
455 		if (error == 0) {
456 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
457 			     __func__, vp, vp->v_writecount);
458 		}
459 	}
460 
461 	/*
462 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
463 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
464 	 */
465 	if (error != 0) {
466 		if (fp != NULL) {
467 			/*
468 			 * Arrange the call by having fdrop() to use
469 			 * vn_closefile().  This is to satisfy
470 			 * filesystems like devfs or tmpfs, which
471 			 * override fo_close().
472 			 */
473 			fp->f_flag |= FOPENFAILED;
474 			fp->f_vnode = vp;
475 			if (fp->f_ops == &badfileops) {
476 				fp->f_type = DTYPE_VNODE;
477 				fp->f_ops = &vnops;
478 			}
479 			vref(vp);
480 		} else {
481 			/*
482 			 * If there is no fp, due to kernel-mode open,
483 			 * we can call VOP_CLOSE() now.
484 			 */
485 			if ((vp->v_type == VFIFO ||
486 			    !MNT_EXTENDED_SHARED(vp->v_mount)) &&
487 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
488 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
489 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
490 			    cred, td);
491 		}
492 	}
493 
494 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
495 	return (error);
496 
497 }
498 
499 /*
500  * Check for write permissions on the specified vnode.
501  * Prototype text segments cannot be written.
502  * It is racy.
503  */
504 int
505 vn_writechk(struct vnode *vp)
506 {
507 
508 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
509 	/*
510 	 * If there's shared text associated with
511 	 * the vnode, try to free it up once.  If
512 	 * we fail, we can't allow writing.
513 	 */
514 	if (VOP_IS_TEXT(vp))
515 		return (ETXTBSY);
516 
517 	return (0);
518 }
519 
520 /*
521  * Vnode close call
522  */
523 static int
524 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
525     struct thread *td, bool keep_ref)
526 {
527 	struct mount *mp;
528 	int error, lock_flags;
529 
530 	lock_flags = vp->v_type != VFIFO && MNT_EXTENDED_SHARED(vp->v_mount) ?
531 	    LK_SHARED : LK_EXCLUSIVE;
532 
533 	vn_start_write(vp, &mp, V_WAIT);
534 	vn_lock(vp, lock_flags | LK_RETRY);
535 	AUDIT_ARG_VNODE1(vp);
536 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
537 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
538 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
539 		    __func__, vp, vp->v_writecount);
540 	}
541 	error = VOP_CLOSE(vp, flags, file_cred, td);
542 	if (keep_ref)
543 		VOP_UNLOCK(vp);
544 	else
545 		vput(vp);
546 	vn_finished_write(mp);
547 	return (error);
548 }
549 
550 int
551 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
552     struct thread *td)
553 {
554 
555 	return (vn_close1(vp, flags, file_cred, td, false));
556 }
557 
558 /*
559  * Heuristic to detect sequential operation.
560  */
561 static int
562 sequential_heuristic(struct uio *uio, struct file *fp)
563 {
564 	enum uio_rw rw;
565 
566 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
567 
568 	rw = uio->uio_rw;
569 	if (fp->f_flag & FRDAHEAD)
570 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
571 
572 	/*
573 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
574 	 * that the first I/O is normally considered to be slightly
575 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
576 	 * unless previous seeks have reduced f_seqcount to 0, in which
577 	 * case offset 0 is not special.
578 	 */
579 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
580 	    uio->uio_offset == fp->f_nextoff[rw]) {
581 		/*
582 		 * f_seqcount is in units of fixed-size blocks so that it
583 		 * depends mainly on the amount of sequential I/O and not
584 		 * much on the number of sequential I/O's.  The fixed size
585 		 * of 16384 is hard-coded here since it is (not quite) just
586 		 * a magic size that works well here.  This size is more
587 		 * closely related to the best I/O size for real disks than
588 		 * to any block size used by software.
589 		 */
590 		if (uio->uio_resid >= IO_SEQMAX * 16384)
591 			fp->f_seqcount[rw] = IO_SEQMAX;
592 		else {
593 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
594 			if (fp->f_seqcount[rw] > IO_SEQMAX)
595 				fp->f_seqcount[rw] = IO_SEQMAX;
596 		}
597 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
598 	}
599 
600 	/* Not sequential.  Quickly draw-down sequentiality. */
601 	if (fp->f_seqcount[rw] > 1)
602 		fp->f_seqcount[rw] = 1;
603 	else
604 		fp->f_seqcount[rw] = 0;
605 	return (0);
606 }
607 
608 /*
609  * Package up an I/O request on a vnode into a uio and do it.
610  */
611 int
612 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
613     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
614     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
615 {
616 	struct uio auio;
617 	struct iovec aiov;
618 	struct mount *mp;
619 	struct ucred *cred;
620 	void *rl_cookie;
621 	struct vn_io_fault_args args;
622 	int error, lock_flags;
623 
624 	if (offset < 0 && vp->v_type != VCHR)
625 		return (EINVAL);
626 	auio.uio_iov = &aiov;
627 	auio.uio_iovcnt = 1;
628 	aiov.iov_base = base;
629 	aiov.iov_len = len;
630 	auio.uio_resid = len;
631 	auio.uio_offset = offset;
632 	auio.uio_segflg = segflg;
633 	auio.uio_rw = rw;
634 	auio.uio_td = td;
635 	error = 0;
636 
637 	if ((ioflg & IO_NODELOCKED) == 0) {
638 		if ((ioflg & IO_RANGELOCKED) == 0) {
639 			if (rw == UIO_READ) {
640 				rl_cookie = vn_rangelock_rlock(vp, offset,
641 				    offset + len);
642 			} else if ((ioflg & IO_APPEND) != 0) {
643 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
644 			} else {
645 				rl_cookie = vn_rangelock_wlock(vp, offset,
646 				    offset + len);
647 			}
648 		} else
649 			rl_cookie = NULL;
650 		mp = NULL;
651 		if (rw == UIO_WRITE) {
652 			if (vp->v_type != VCHR &&
653 			    (error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH))
654 			    != 0)
655 				goto out;
656 			lock_flags = vn_lktype_write(mp, vp);
657 		} else
658 			lock_flags = LK_SHARED;
659 		vn_lock(vp, lock_flags | LK_RETRY);
660 	} else
661 		rl_cookie = NULL;
662 
663 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
664 #ifdef MAC
665 	if ((ioflg & IO_NOMACCHECK) == 0) {
666 		if (rw == UIO_READ)
667 			error = mac_vnode_check_read(active_cred, file_cred,
668 			    vp);
669 		else
670 			error = mac_vnode_check_write(active_cred, file_cred,
671 			    vp);
672 	}
673 #endif
674 	if (error == 0) {
675 		if (file_cred != NULL)
676 			cred = file_cred;
677 		else
678 			cred = active_cred;
679 		if (do_vn_io_fault(vp, &auio)) {
680 			args.kind = VN_IO_FAULT_VOP;
681 			args.cred = cred;
682 			args.flags = ioflg;
683 			args.args.vop_args.vp = vp;
684 			error = vn_io_fault1(vp, &auio, &args, td);
685 		} else if (rw == UIO_READ) {
686 			error = VOP_READ(vp, &auio, ioflg, cred);
687 		} else /* if (rw == UIO_WRITE) */ {
688 			error = VOP_WRITE(vp, &auio, ioflg, cred);
689 		}
690 	}
691 	if (aresid)
692 		*aresid = auio.uio_resid;
693 	else
694 		if (auio.uio_resid && error == 0)
695 			error = EIO;
696 	if ((ioflg & IO_NODELOCKED) == 0) {
697 		VOP_UNLOCK(vp);
698 		if (mp != NULL)
699 			vn_finished_write(mp);
700 	}
701  out:
702 	if (rl_cookie != NULL)
703 		vn_rangelock_unlock(vp, rl_cookie);
704 	return (error);
705 }
706 
707 /*
708  * Package up an I/O request on a vnode into a uio and do it.  The I/O
709  * request is split up into smaller chunks and we try to avoid saturating
710  * the buffer cache while potentially holding a vnode locked, so we
711  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
712  * to give other processes a chance to lock the vnode (either other processes
713  * core'ing the same binary, or unrelated processes scanning the directory).
714  */
715 int
716 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
717     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
718     struct ucred *file_cred, size_t *aresid, struct thread *td)
719 {
720 	int error = 0;
721 	ssize_t iaresid;
722 
723 	do {
724 		int chunk;
725 
726 		/*
727 		 * Force `offset' to a multiple of MAXBSIZE except possibly
728 		 * for the first chunk, so that filesystems only need to
729 		 * write full blocks except possibly for the first and last
730 		 * chunks.
731 		 */
732 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
733 
734 		if (chunk > len)
735 			chunk = len;
736 		if (rw != UIO_READ && vp->v_type == VREG)
737 			bwillwrite();
738 		iaresid = 0;
739 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
740 		    ioflg, active_cred, file_cred, &iaresid, td);
741 		len -= chunk;	/* aresid calc already includes length */
742 		if (error)
743 			break;
744 		offset += chunk;
745 		base = (char *)base + chunk;
746 		kern_yield(PRI_USER);
747 	} while (len);
748 	if (aresid)
749 		*aresid = len + iaresid;
750 	return (error);
751 }
752 
753 #if OFF_MAX <= LONG_MAX
754 off_t
755 foffset_lock(struct file *fp, int flags)
756 {
757 	volatile short *flagsp;
758 	off_t res;
759 	short state;
760 
761 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
762 
763 	if ((flags & FOF_NOLOCK) != 0)
764 		return (atomic_load_long(&fp->f_offset));
765 
766 	/*
767 	 * According to McKusick the vn lock was protecting f_offset here.
768 	 * It is now protected by the FOFFSET_LOCKED flag.
769 	 */
770 	flagsp = &fp->f_vnread_flags;
771 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
772 		return (atomic_load_long(&fp->f_offset));
773 
774 	sleepq_lock(&fp->f_vnread_flags);
775 	state = atomic_load_16(flagsp);
776 	for (;;) {
777 		if ((state & FOFFSET_LOCKED) == 0) {
778 			if (!atomic_fcmpset_acq_16(flagsp, &state,
779 			    FOFFSET_LOCKED))
780 				continue;
781 			break;
782 		}
783 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
784 			if (!atomic_fcmpset_acq_16(flagsp, &state,
785 			    state | FOFFSET_LOCK_WAITING))
786 				continue;
787 		}
788 		DROP_GIANT();
789 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
790 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
791 		PICKUP_GIANT();
792 		sleepq_lock(&fp->f_vnread_flags);
793 		state = atomic_load_16(flagsp);
794 	}
795 	res = atomic_load_long(&fp->f_offset);
796 	sleepq_release(&fp->f_vnread_flags);
797 	return (res);
798 }
799 
800 void
801 foffset_unlock(struct file *fp, off_t val, int flags)
802 {
803 	volatile short *flagsp;
804 	short state;
805 
806 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
807 
808 	if ((flags & FOF_NOUPDATE) == 0)
809 		atomic_store_long(&fp->f_offset, val);
810 	if ((flags & FOF_NEXTOFF_R) != 0)
811 		fp->f_nextoff[UIO_READ] = val;
812 	if ((flags & FOF_NEXTOFF_W) != 0)
813 		fp->f_nextoff[UIO_WRITE] = val;
814 
815 	if ((flags & FOF_NOLOCK) != 0)
816 		return;
817 
818 	flagsp = &fp->f_vnread_flags;
819 	state = atomic_load_16(flagsp);
820 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
821 	    atomic_cmpset_rel_16(flagsp, state, 0))
822 		return;
823 
824 	sleepq_lock(&fp->f_vnread_flags);
825 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
826 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
827 	fp->f_vnread_flags = 0;
828 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
829 	sleepq_release(&fp->f_vnread_flags);
830 }
831 #else
832 off_t
833 foffset_lock(struct file *fp, int flags)
834 {
835 	struct mtx *mtxp;
836 	off_t res;
837 
838 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
839 
840 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
841 	mtx_lock(mtxp);
842 	if ((flags & FOF_NOLOCK) == 0) {
843 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
844 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
845 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
846 			    "vofflock", 0);
847 		}
848 		fp->f_vnread_flags |= FOFFSET_LOCKED;
849 	}
850 	res = fp->f_offset;
851 	mtx_unlock(mtxp);
852 	return (res);
853 }
854 
855 void
856 foffset_unlock(struct file *fp, off_t val, int flags)
857 {
858 	struct mtx *mtxp;
859 
860 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
861 
862 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
863 	mtx_lock(mtxp);
864 	if ((flags & FOF_NOUPDATE) == 0)
865 		fp->f_offset = val;
866 	if ((flags & FOF_NEXTOFF_R) != 0)
867 		fp->f_nextoff[UIO_READ] = val;
868 	if ((flags & FOF_NEXTOFF_W) != 0)
869 		fp->f_nextoff[UIO_WRITE] = val;
870 	if ((flags & FOF_NOLOCK) == 0) {
871 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
872 		    ("Lost FOFFSET_LOCKED"));
873 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
874 			wakeup(&fp->f_vnread_flags);
875 		fp->f_vnread_flags = 0;
876 	}
877 	mtx_unlock(mtxp);
878 }
879 #endif
880 
881 void
882 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
883 {
884 
885 	if ((flags & FOF_OFFSET) == 0)
886 		uio->uio_offset = foffset_lock(fp, flags);
887 }
888 
889 void
890 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
891 {
892 
893 	if ((flags & FOF_OFFSET) == 0)
894 		foffset_unlock(fp, uio->uio_offset, flags);
895 }
896 
897 static int
898 get_advice(struct file *fp, struct uio *uio)
899 {
900 	struct mtx *mtxp;
901 	int ret;
902 
903 	ret = POSIX_FADV_NORMAL;
904 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
905 		return (ret);
906 
907 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
908 	mtx_lock(mtxp);
909 	if (fp->f_advice != NULL &&
910 	    uio->uio_offset >= fp->f_advice->fa_start &&
911 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
912 		ret = fp->f_advice->fa_advice;
913 	mtx_unlock(mtxp);
914 	return (ret);
915 }
916 
917 static int
918 get_write_ioflag(struct file *fp)
919 {
920 	int ioflag;
921 	struct mount *mp;
922 	struct vnode *vp;
923 
924 	ioflag = 0;
925 	vp = fp->f_vnode;
926 	mp = atomic_load_ptr(&vp->v_mount);
927 
928 	if ((fp->f_flag & O_DIRECT) != 0)
929 		ioflag |= IO_DIRECT;
930 
931 	if ((fp->f_flag & O_FSYNC) != 0 ||
932 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0))
933 		ioflag |= IO_SYNC;
934 
935 	/*
936 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
937 	 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC
938 	 * fall back to full O_SYNC behavior.
939 	 */
940 	if ((fp->f_flag & O_DSYNC) != 0)
941 		ioflag |= IO_SYNC | IO_DATASYNC;
942 
943 	return (ioflag);
944 }
945 
946 int
947 vn_read_from_obj(struct vnode *vp, struct uio *uio)
948 {
949 	vm_object_t obj;
950 	vm_page_t ma[io_hold_cnt + 2];
951 	off_t off, vsz;
952 	ssize_t resid;
953 	int error, i, j;
954 
955 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
956 	obj = atomic_load_ptr(&vp->v_object);
957 	if (obj == NULL)
958 		return (EJUSTRETURN);
959 
960 	/*
961 	 * Depends on type stability of vm_objects.
962 	 */
963 	vm_object_pip_add(obj, 1);
964 	if ((obj->flags & OBJ_DEAD) != 0) {
965 		/*
966 		 * Note that object might be already reused from the
967 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
968 		 * we recheck for DOOMED vnode state after all pages
969 		 * are busied, and retract then.
970 		 *
971 		 * But we check for OBJ_DEAD to ensure that we do not
972 		 * busy pages while vm_object_terminate_pages()
973 		 * processes the queue.
974 		 */
975 		error = EJUSTRETURN;
976 		goto out_pip;
977 	}
978 
979 	resid = uio->uio_resid;
980 	off = uio->uio_offset;
981 	for (i = 0; resid > 0; i++) {
982 		MPASS(i < io_hold_cnt + 2);
983 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
984 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
985 		    VM_ALLOC_NOWAIT);
986 		if (ma[i] == NULL)
987 			break;
988 
989 		/*
990 		 * Skip invalid pages.  Valid mask can be partial only
991 		 * at EOF, and we clip later.
992 		 */
993 		if (vm_page_none_valid(ma[i])) {
994 			vm_page_sunbusy(ma[i]);
995 			break;
996 		}
997 
998 		resid -= PAGE_SIZE;
999 		off += PAGE_SIZE;
1000 	}
1001 	if (i == 0) {
1002 		error = EJUSTRETURN;
1003 		goto out_pip;
1004 	}
1005 
1006 	/*
1007 	 * Check VIRF_DOOMED after we busied our pages.  Since
1008 	 * vgonel() terminates the vnode' vm_object, it cannot
1009 	 * process past pages busied by us.
1010 	 */
1011 	if (VN_IS_DOOMED(vp)) {
1012 		error = EJUSTRETURN;
1013 		goto out;
1014 	}
1015 
1016 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
1017 	if (resid > uio->uio_resid)
1018 		resid = uio->uio_resid;
1019 
1020 	/*
1021 	 * Unlocked read of vnp_size is safe because truncation cannot
1022 	 * pass busied page.  But we load vnp_size into a local
1023 	 * variable so that possible concurrent extension does not
1024 	 * break calculation.
1025 	 */
1026 #if defined(__powerpc__) && !defined(__powerpc64__)
1027 	vsz = obj->un_pager.vnp.vnp_size;
1028 #else
1029 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1030 #endif
1031 	if (uio->uio_offset >= vsz) {
1032 		error = EJUSTRETURN;
1033 		goto out;
1034 	}
1035 	if (uio->uio_offset + resid > vsz)
1036 		resid = vsz - uio->uio_offset;
1037 
1038 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1039 
1040 out:
1041 	for (j = 0; j < i; j++) {
1042 		if (error == 0)
1043 			vm_page_reference(ma[j]);
1044 		vm_page_sunbusy(ma[j]);
1045 	}
1046 out_pip:
1047 	vm_object_pip_wakeup(obj);
1048 	if (error != 0)
1049 		return (error);
1050 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1051 }
1052 
1053 /*
1054  * File table vnode read routine.
1055  */
1056 static int
1057 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1058     struct thread *td)
1059 {
1060 	struct vnode *vp;
1061 	off_t orig_offset;
1062 	int error, ioflag;
1063 	int advice;
1064 
1065 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1066 	    uio->uio_td, td));
1067 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1068 	vp = fp->f_vnode;
1069 	ioflag = 0;
1070 	if (fp->f_flag & FNONBLOCK)
1071 		ioflag |= IO_NDELAY;
1072 	if (fp->f_flag & O_DIRECT)
1073 		ioflag |= IO_DIRECT;
1074 
1075 	/*
1076 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1077 	 * allows us to avoid unneeded work outright.
1078 	 */
1079 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1080 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1081 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1082 		if (error == 0) {
1083 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1084 			return (0);
1085 		}
1086 		if (error != EJUSTRETURN)
1087 			return (error);
1088 	}
1089 
1090 	advice = get_advice(fp, uio);
1091 	vn_lock(vp, LK_SHARED | LK_RETRY);
1092 
1093 	switch (advice) {
1094 	case POSIX_FADV_NORMAL:
1095 	case POSIX_FADV_SEQUENTIAL:
1096 	case POSIX_FADV_NOREUSE:
1097 		ioflag |= sequential_heuristic(uio, fp);
1098 		break;
1099 	case POSIX_FADV_RANDOM:
1100 		/* Disable read-ahead for random I/O. */
1101 		break;
1102 	}
1103 	orig_offset = uio->uio_offset;
1104 
1105 #ifdef MAC
1106 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1107 	if (error == 0)
1108 #endif
1109 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1110 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1111 	VOP_UNLOCK(vp);
1112 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1113 	    orig_offset != uio->uio_offset)
1114 		/*
1115 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1116 		 * for the backing file after a POSIX_FADV_NOREUSE
1117 		 * read(2).
1118 		 */
1119 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1120 		    POSIX_FADV_DONTNEED);
1121 	return (error);
1122 }
1123 
1124 /*
1125  * File table vnode write routine.
1126  */
1127 static int
1128 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1129     struct thread *td)
1130 {
1131 	struct vnode *vp;
1132 	struct mount *mp;
1133 	off_t orig_offset;
1134 	int error, ioflag;
1135 	int advice;
1136 	bool need_finished_write;
1137 
1138 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1139 	    uio->uio_td, td));
1140 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1141 	vp = fp->f_vnode;
1142 	if (vp->v_type == VREG)
1143 		bwillwrite();
1144 	ioflag = IO_UNIT;
1145 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0)
1146 		ioflag |= IO_APPEND;
1147 	if ((fp->f_flag & FNONBLOCK) != 0)
1148 		ioflag |= IO_NDELAY;
1149 	ioflag |= get_write_ioflag(fp);
1150 
1151 	mp = NULL;
1152 	need_finished_write = false;
1153 	if (vp->v_type != VCHR) {
1154 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1155 		if (error != 0)
1156 			goto unlock;
1157 		need_finished_write = true;
1158 	}
1159 
1160 	advice = get_advice(fp, uio);
1161 
1162 	vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
1163 	switch (advice) {
1164 	case POSIX_FADV_NORMAL:
1165 	case POSIX_FADV_SEQUENTIAL:
1166 	case POSIX_FADV_NOREUSE:
1167 		ioflag |= sequential_heuristic(uio, fp);
1168 		break;
1169 	case POSIX_FADV_RANDOM:
1170 		/* XXX: Is this correct? */
1171 		break;
1172 	}
1173 	orig_offset = uio->uio_offset;
1174 
1175 #ifdef MAC
1176 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1177 	if (error == 0)
1178 #endif
1179 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1180 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1181 	VOP_UNLOCK(vp);
1182 	if (need_finished_write)
1183 		vn_finished_write(mp);
1184 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1185 	    orig_offset != uio->uio_offset)
1186 		/*
1187 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1188 		 * for the backing file after a POSIX_FADV_NOREUSE
1189 		 * write(2).
1190 		 */
1191 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1192 		    POSIX_FADV_DONTNEED);
1193 unlock:
1194 	return (error);
1195 }
1196 
1197 /*
1198  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1199  * prevent the following deadlock:
1200  *
1201  * Assume that the thread A reads from the vnode vp1 into userspace
1202  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1203  * currently not resident, then system ends up with the call chain
1204  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1205  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1206  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1207  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1208  * backed by the pages of vnode vp1, and some page in buf2 is not
1209  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1210  *
1211  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1212  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1213  * Instead, it first tries to do the whole range i/o with pagefaults
1214  * disabled. If all pages in the i/o buffer are resident and mapped,
1215  * VOP will succeed (ignoring the genuine filesystem errors).
1216  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1217  * i/o in chunks, with all pages in the chunk prefaulted and held
1218  * using vm_fault_quick_hold_pages().
1219  *
1220  * Filesystems using this deadlock avoidance scheme should use the
1221  * array of the held pages from uio, saved in the curthread->td_ma,
1222  * instead of doing uiomove().  A helper function
1223  * vn_io_fault_uiomove() converts uiomove request into
1224  * uiomove_fromphys() over td_ma array.
1225  *
1226  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1227  * make the current i/o request atomic with respect to other i/os and
1228  * truncations.
1229  */
1230 
1231 /*
1232  * Decode vn_io_fault_args and perform the corresponding i/o.
1233  */
1234 static int
1235 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1236     struct thread *td)
1237 {
1238 	int error, save;
1239 
1240 	error = 0;
1241 	save = vm_fault_disable_pagefaults();
1242 	switch (args->kind) {
1243 	case VN_IO_FAULT_FOP:
1244 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1245 		    uio, args->cred, args->flags, td);
1246 		break;
1247 	case VN_IO_FAULT_VOP:
1248 		if (uio->uio_rw == UIO_READ) {
1249 			error = VOP_READ(args->args.vop_args.vp, uio,
1250 			    args->flags, args->cred);
1251 		} else if (uio->uio_rw == UIO_WRITE) {
1252 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1253 			    args->flags, args->cred);
1254 		}
1255 		break;
1256 	default:
1257 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1258 		    args->kind, uio->uio_rw);
1259 	}
1260 	vm_fault_enable_pagefaults(save);
1261 	return (error);
1262 }
1263 
1264 static int
1265 vn_io_fault_touch(char *base, const struct uio *uio)
1266 {
1267 	int r;
1268 
1269 	r = fubyte(base);
1270 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1271 		return (EFAULT);
1272 	return (0);
1273 }
1274 
1275 static int
1276 vn_io_fault_prefault_user(const struct uio *uio)
1277 {
1278 	char *base;
1279 	const struct iovec *iov;
1280 	size_t len;
1281 	ssize_t resid;
1282 	int error, i;
1283 
1284 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1285 	    ("vn_io_fault_prefault userspace"));
1286 
1287 	error = i = 0;
1288 	iov = uio->uio_iov;
1289 	resid = uio->uio_resid;
1290 	base = iov->iov_base;
1291 	len = iov->iov_len;
1292 	while (resid > 0) {
1293 		error = vn_io_fault_touch(base, uio);
1294 		if (error != 0)
1295 			break;
1296 		if (len < PAGE_SIZE) {
1297 			if (len != 0) {
1298 				error = vn_io_fault_touch(base + len - 1, uio);
1299 				if (error != 0)
1300 					break;
1301 				resid -= len;
1302 			}
1303 			if (++i >= uio->uio_iovcnt)
1304 				break;
1305 			iov = uio->uio_iov + i;
1306 			base = iov->iov_base;
1307 			len = iov->iov_len;
1308 		} else {
1309 			len -= PAGE_SIZE;
1310 			base += PAGE_SIZE;
1311 			resid -= PAGE_SIZE;
1312 		}
1313 	}
1314 	return (error);
1315 }
1316 
1317 /*
1318  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1319  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1320  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1321  * into args and call vn_io_fault1() to handle faults during the user
1322  * mode buffer accesses.
1323  */
1324 static int
1325 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1326     struct thread *td)
1327 {
1328 	vm_page_t ma[io_hold_cnt + 2];
1329 	struct uio *uio_clone, short_uio;
1330 	struct iovec short_iovec[1];
1331 	vm_page_t *prev_td_ma;
1332 	vm_prot_t prot;
1333 	vm_offset_t addr, end;
1334 	size_t len, resid;
1335 	ssize_t adv;
1336 	int error, cnt, saveheld, prev_td_ma_cnt;
1337 
1338 	if (vn_io_fault_prefault) {
1339 		error = vn_io_fault_prefault_user(uio);
1340 		if (error != 0)
1341 			return (error); /* Or ignore ? */
1342 	}
1343 
1344 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1345 
1346 	/*
1347 	 * The UFS follows IO_UNIT directive and replays back both
1348 	 * uio_offset and uio_resid if an error is encountered during the
1349 	 * operation.  But, since the iovec may be already advanced,
1350 	 * uio is still in an inconsistent state.
1351 	 *
1352 	 * Cache a copy of the original uio, which is advanced to the redo
1353 	 * point using UIO_NOCOPY below.
1354 	 */
1355 	uio_clone = cloneuio(uio);
1356 	resid = uio->uio_resid;
1357 
1358 	short_uio.uio_segflg = UIO_USERSPACE;
1359 	short_uio.uio_rw = uio->uio_rw;
1360 	short_uio.uio_td = uio->uio_td;
1361 
1362 	error = vn_io_fault_doio(args, uio, td);
1363 	if (error != EFAULT)
1364 		goto out;
1365 
1366 	atomic_add_long(&vn_io_faults_cnt, 1);
1367 	uio_clone->uio_segflg = UIO_NOCOPY;
1368 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1369 	uio_clone->uio_segflg = uio->uio_segflg;
1370 
1371 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1372 	prev_td_ma = td->td_ma;
1373 	prev_td_ma_cnt = td->td_ma_cnt;
1374 
1375 	while (uio_clone->uio_resid != 0) {
1376 		len = uio_clone->uio_iov->iov_len;
1377 		if (len == 0) {
1378 			KASSERT(uio_clone->uio_iovcnt >= 1,
1379 			    ("iovcnt underflow"));
1380 			uio_clone->uio_iov++;
1381 			uio_clone->uio_iovcnt--;
1382 			continue;
1383 		}
1384 		if (len > ptoa(io_hold_cnt))
1385 			len = ptoa(io_hold_cnt);
1386 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1387 		end = round_page(addr + len);
1388 		if (end < addr) {
1389 			error = EFAULT;
1390 			break;
1391 		}
1392 		/*
1393 		 * A perfectly misaligned address and length could cause
1394 		 * both the start and the end of the chunk to use partial
1395 		 * page.  +2 accounts for such a situation.
1396 		 */
1397 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1398 		    addr, len, prot, ma, io_hold_cnt + 2);
1399 		if (cnt == -1) {
1400 			error = EFAULT;
1401 			break;
1402 		}
1403 		short_uio.uio_iov = &short_iovec[0];
1404 		short_iovec[0].iov_base = (void *)addr;
1405 		short_uio.uio_iovcnt = 1;
1406 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1407 		short_uio.uio_offset = uio_clone->uio_offset;
1408 		td->td_ma = ma;
1409 		td->td_ma_cnt = cnt;
1410 
1411 		error = vn_io_fault_doio(args, &short_uio, td);
1412 		vm_page_unhold_pages(ma, cnt);
1413 		adv = len - short_uio.uio_resid;
1414 
1415 		uio_clone->uio_iov->iov_base =
1416 		    (char *)uio_clone->uio_iov->iov_base + adv;
1417 		uio_clone->uio_iov->iov_len -= adv;
1418 		uio_clone->uio_resid -= adv;
1419 		uio_clone->uio_offset += adv;
1420 
1421 		uio->uio_resid -= adv;
1422 		uio->uio_offset += adv;
1423 
1424 		if (error != 0 || adv == 0)
1425 			break;
1426 	}
1427 	td->td_ma = prev_td_ma;
1428 	td->td_ma_cnt = prev_td_ma_cnt;
1429 	curthread_pflags_restore(saveheld);
1430 out:
1431 	free(uio_clone, M_IOV);
1432 	return (error);
1433 }
1434 
1435 static int
1436 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1437     int flags, struct thread *td)
1438 {
1439 	fo_rdwr_t *doio;
1440 	struct vnode *vp;
1441 	void *rl_cookie;
1442 	struct vn_io_fault_args args;
1443 	int error;
1444 	bool do_io_fault, do_rangelock;
1445 
1446 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1447 	vp = fp->f_vnode;
1448 
1449 	/*
1450 	 * The ability to read(2) on a directory has historically been
1451 	 * allowed for all users, but this can and has been the source of
1452 	 * at least one security issue in the past.  As such, it is now hidden
1453 	 * away behind a sysctl for those that actually need it to use it, and
1454 	 * restricted to root when it's turned on to make it relatively safe to
1455 	 * leave on for longer sessions of need.
1456 	 */
1457 	if (vp->v_type == VDIR) {
1458 		KASSERT(uio->uio_rw == UIO_READ,
1459 		    ("illegal write attempted on a directory"));
1460 		if (!vfs_allow_read_dir)
1461 			return (EISDIR);
1462 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1463 			return (EISDIR);
1464 	}
1465 
1466 	do_io_fault = do_vn_io_fault(vp, uio);
1467 	do_rangelock = do_io_fault || (vn_irflag_read(vp) & VIRF_PGREAD) != 0;
1468 	foffset_lock_uio(fp, uio, flags);
1469 	if (do_rangelock) {
1470 		if (uio->uio_rw == UIO_READ) {
1471 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1472 			    uio->uio_offset + uio->uio_resid);
1473 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1474 		    (flags & FOF_OFFSET) == 0) {
1475 			/* For appenders, punt and lock the whole range. */
1476 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1477 		} else {
1478 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1479 			    uio->uio_offset + uio->uio_resid);
1480 		}
1481 	}
1482 	if (do_io_fault) {
1483 		args.kind = VN_IO_FAULT_FOP;
1484 		args.args.fop_args.fp = fp;
1485 		args.args.fop_args.doio = doio;
1486 		args.cred = active_cred;
1487 		args.flags = flags | FOF_OFFSET;
1488 		error = vn_io_fault1(vp, uio, &args, td);
1489 	} else {
1490 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1491 	}
1492 	if (do_rangelock)
1493 		vn_rangelock_unlock(vp, rl_cookie);
1494 	foffset_unlock_uio(fp, uio, flags);
1495 	return (error);
1496 }
1497 
1498 /*
1499  * Helper function to perform the requested uiomove operation using
1500  * the held pages for io->uio_iov[0].iov_base buffer instead of
1501  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1502  * instead of iov_base prevents page faults that could occur due to
1503  * pmap_collect() invalidating the mapping created by
1504  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1505  * object cleanup revoking the write access from page mappings.
1506  *
1507  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1508  * instead of plain uiomove().
1509  */
1510 int
1511 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1512 {
1513 	struct uio transp_uio;
1514 	struct iovec transp_iov[1];
1515 	struct thread *td;
1516 	size_t adv;
1517 	int error, pgadv;
1518 
1519 	td = curthread;
1520 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1521 	    uio->uio_segflg != UIO_USERSPACE)
1522 		return (uiomove(data, xfersize, uio));
1523 
1524 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1525 	transp_iov[0].iov_base = data;
1526 	transp_uio.uio_iov = &transp_iov[0];
1527 	transp_uio.uio_iovcnt = 1;
1528 	if (xfersize > uio->uio_resid)
1529 		xfersize = uio->uio_resid;
1530 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1531 	transp_uio.uio_offset = 0;
1532 	transp_uio.uio_segflg = UIO_SYSSPACE;
1533 	/*
1534 	 * Since transp_iov points to data, and td_ma page array
1535 	 * corresponds to original uio->uio_iov, we need to invert the
1536 	 * direction of the i/o operation as passed to
1537 	 * uiomove_fromphys().
1538 	 */
1539 	switch (uio->uio_rw) {
1540 	case UIO_WRITE:
1541 		transp_uio.uio_rw = UIO_READ;
1542 		break;
1543 	case UIO_READ:
1544 		transp_uio.uio_rw = UIO_WRITE;
1545 		break;
1546 	}
1547 	transp_uio.uio_td = uio->uio_td;
1548 	error = uiomove_fromphys(td->td_ma,
1549 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1550 	    xfersize, &transp_uio);
1551 	adv = xfersize - transp_uio.uio_resid;
1552 	pgadv =
1553 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1554 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1555 	td->td_ma += pgadv;
1556 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1557 	    pgadv));
1558 	td->td_ma_cnt -= pgadv;
1559 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1560 	uio->uio_iov->iov_len -= adv;
1561 	uio->uio_resid -= adv;
1562 	uio->uio_offset += adv;
1563 	return (error);
1564 }
1565 
1566 int
1567 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1568     struct uio *uio)
1569 {
1570 	struct thread *td;
1571 	vm_offset_t iov_base;
1572 	int cnt, pgadv;
1573 
1574 	td = curthread;
1575 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1576 	    uio->uio_segflg != UIO_USERSPACE)
1577 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1578 
1579 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1580 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1581 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1582 	switch (uio->uio_rw) {
1583 	case UIO_WRITE:
1584 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1585 		    offset, cnt);
1586 		break;
1587 	case UIO_READ:
1588 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1589 		    cnt);
1590 		break;
1591 	}
1592 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1593 	td->td_ma += pgadv;
1594 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1595 	    pgadv));
1596 	td->td_ma_cnt -= pgadv;
1597 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1598 	uio->uio_iov->iov_len -= cnt;
1599 	uio->uio_resid -= cnt;
1600 	uio->uio_offset += cnt;
1601 	return (0);
1602 }
1603 
1604 /*
1605  * File table truncate routine.
1606  */
1607 static int
1608 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1609     struct thread *td)
1610 {
1611 	struct mount *mp;
1612 	struct vnode *vp;
1613 	void *rl_cookie;
1614 	int error;
1615 
1616 	vp = fp->f_vnode;
1617 
1618 retry:
1619 	/*
1620 	 * Lock the whole range for truncation.  Otherwise split i/o
1621 	 * might happen partly before and partly after the truncation.
1622 	 */
1623 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1624 	error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
1625 	if (error)
1626 		goto out1;
1627 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1628 	AUDIT_ARG_VNODE1(vp);
1629 	if (vp->v_type == VDIR) {
1630 		error = EISDIR;
1631 		goto out;
1632 	}
1633 #ifdef MAC
1634 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1635 	if (error)
1636 		goto out;
1637 #endif
1638 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1639 	    fp->f_cred);
1640 out:
1641 	VOP_UNLOCK(vp);
1642 	vn_finished_write(mp);
1643 out1:
1644 	vn_rangelock_unlock(vp, rl_cookie);
1645 	if (error == ERELOOKUP)
1646 		goto retry;
1647 	return (error);
1648 }
1649 
1650 /*
1651  * Truncate a file that is already locked.
1652  */
1653 int
1654 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1655     struct ucred *cred)
1656 {
1657 	struct vattr vattr;
1658 	int error;
1659 
1660 	error = VOP_ADD_WRITECOUNT(vp, 1);
1661 	if (error == 0) {
1662 		VATTR_NULL(&vattr);
1663 		vattr.va_size = length;
1664 		if (sync)
1665 			vattr.va_vaflags |= VA_SYNC;
1666 		error = VOP_SETATTR(vp, &vattr, cred);
1667 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1668 	}
1669 	return (error);
1670 }
1671 
1672 /*
1673  * File table vnode stat routine.
1674  */
1675 int
1676 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred)
1677 {
1678 	struct vnode *vp = fp->f_vnode;
1679 	int error;
1680 
1681 	vn_lock(vp, LK_SHARED | LK_RETRY);
1682 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred);
1683 	VOP_UNLOCK(vp);
1684 
1685 	return (error);
1686 }
1687 
1688 /*
1689  * File table vnode ioctl routine.
1690  */
1691 static int
1692 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1693     struct thread *td)
1694 {
1695 	struct vnode *vp;
1696 	struct fiobmap2_arg *bmarg;
1697 	off_t size;
1698 	int error;
1699 
1700 	vp = fp->f_vnode;
1701 	switch (vp->v_type) {
1702 	case VDIR:
1703 	case VREG:
1704 		switch (com) {
1705 		case FIONREAD:
1706 			error = vn_getsize(vp, &size, active_cred);
1707 			if (error == 0)
1708 				*(int *)data = size - fp->f_offset;
1709 			return (error);
1710 		case FIOBMAP2:
1711 			bmarg = (struct fiobmap2_arg *)data;
1712 			vn_lock(vp, LK_SHARED | LK_RETRY);
1713 #ifdef MAC
1714 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1715 			    vp);
1716 			if (error == 0)
1717 #endif
1718 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1719 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1720 			VOP_UNLOCK(vp);
1721 			return (error);
1722 		case FIONBIO:
1723 		case FIOASYNC:
1724 			return (0);
1725 		default:
1726 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1727 			    active_cred, td));
1728 		}
1729 		break;
1730 	case VCHR:
1731 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1732 		    active_cred, td));
1733 	default:
1734 		return (ENOTTY);
1735 	}
1736 }
1737 
1738 /*
1739  * File table vnode poll routine.
1740  */
1741 static int
1742 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1743     struct thread *td)
1744 {
1745 	struct vnode *vp;
1746 	int error;
1747 
1748 	vp = fp->f_vnode;
1749 #if defined(MAC) || defined(AUDIT)
1750 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1751 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1752 		AUDIT_ARG_VNODE1(vp);
1753 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1754 		VOP_UNLOCK(vp);
1755 		if (error != 0)
1756 			return (error);
1757 	}
1758 #endif
1759 	error = VOP_POLL(vp, events, fp->f_cred, td);
1760 	return (error);
1761 }
1762 
1763 /*
1764  * Acquire the requested lock and then check for validity.  LK_RETRY
1765  * permits vn_lock to return doomed vnodes.
1766  */
1767 static int __noinline
1768 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1769     int error)
1770 {
1771 
1772 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1773 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1774 
1775 	if (error == 0)
1776 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1777 
1778 	if ((flags & LK_RETRY) == 0) {
1779 		if (error == 0) {
1780 			VOP_UNLOCK(vp);
1781 			error = ENOENT;
1782 		}
1783 		return (error);
1784 	}
1785 
1786 	/*
1787 	 * LK_RETRY case.
1788 	 *
1789 	 * Nothing to do if we got the lock.
1790 	 */
1791 	if (error == 0)
1792 		return (0);
1793 
1794 	/*
1795 	 * Interlock was dropped by the call in _vn_lock.
1796 	 */
1797 	flags &= ~LK_INTERLOCK;
1798 	do {
1799 		error = VOP_LOCK1(vp, flags, file, line);
1800 	} while (error != 0);
1801 	return (0);
1802 }
1803 
1804 int
1805 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1806 {
1807 	int error;
1808 
1809 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1810 	    ("vn_lock: no locktype (%d passed)", flags));
1811 	VNPASS(vp->v_holdcnt > 0, vp);
1812 	error = VOP_LOCK1(vp, flags, file, line);
1813 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1814 		return (_vn_lock_fallback(vp, flags, file, line, error));
1815 	return (0);
1816 }
1817 
1818 /*
1819  * File table vnode close routine.
1820  */
1821 static int
1822 vn_closefile(struct file *fp, struct thread *td)
1823 {
1824 	struct vnode *vp;
1825 	struct flock lf;
1826 	int error;
1827 	bool ref;
1828 
1829 	vp = fp->f_vnode;
1830 	fp->f_ops = &badfileops;
1831 	ref = (fp->f_flag & FHASLOCK) != 0;
1832 
1833 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1834 
1835 	if (__predict_false(ref)) {
1836 		lf.l_whence = SEEK_SET;
1837 		lf.l_start = 0;
1838 		lf.l_len = 0;
1839 		lf.l_type = F_UNLCK;
1840 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1841 		vrele(vp);
1842 	}
1843 	return (error);
1844 }
1845 
1846 /*
1847  * Preparing to start a filesystem write operation. If the operation is
1848  * permitted, then we bump the count of operations in progress and
1849  * proceed. If a suspend request is in progress, we wait until the
1850  * suspension is over, and then proceed.
1851  */
1852 static int
1853 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1854 {
1855 	struct mount_pcpu *mpcpu;
1856 	int error, mflags;
1857 
1858 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1859 	    vfs_op_thread_enter(mp, mpcpu)) {
1860 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1861 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1862 		vfs_op_thread_exit(mp, mpcpu);
1863 		return (0);
1864 	}
1865 
1866 	if (mplocked)
1867 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1868 	else
1869 		MNT_ILOCK(mp);
1870 
1871 	error = 0;
1872 
1873 	/*
1874 	 * Check on status of suspension.
1875 	 */
1876 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1877 	    mp->mnt_susp_owner != curthread) {
1878 		mflags = 0;
1879 		if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
1880 			if (flags & V_PCATCH)
1881 				mflags |= PCATCH;
1882 		}
1883 		mflags |= (PUSER - 1);
1884 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1885 			if ((flags & V_NOWAIT) != 0) {
1886 				error = EWOULDBLOCK;
1887 				goto unlock;
1888 			}
1889 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1890 			    "suspfs", 0);
1891 			if (error != 0)
1892 				goto unlock;
1893 		}
1894 	}
1895 	if ((flags & V_XSLEEP) != 0)
1896 		goto unlock;
1897 	mp->mnt_writeopcount++;
1898 unlock:
1899 	if (error != 0 || (flags & V_XSLEEP) != 0)
1900 		MNT_REL(mp);
1901 	MNT_IUNLOCK(mp);
1902 	return (error);
1903 }
1904 
1905 int
1906 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1907 {
1908 	struct mount *mp;
1909 	int error;
1910 
1911 	KASSERT((flags & ~V_VALID_FLAGS) == 0,
1912 	    ("%s: invalid flags passed %d\n", __func__, flags));
1913 
1914 	error = 0;
1915 	/*
1916 	 * If a vnode is provided, get and return the mount point that
1917 	 * to which it will write.
1918 	 */
1919 	if (vp != NULL) {
1920 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1921 			*mpp = NULL;
1922 			if (error != EOPNOTSUPP)
1923 				return (error);
1924 			return (0);
1925 		}
1926 	}
1927 	if ((mp = *mpp) == NULL)
1928 		return (0);
1929 
1930 	/*
1931 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1932 	 * a vfs_ref().
1933 	 * As long as a vnode is not provided we need to acquire a
1934 	 * refcount for the provided mountpoint too, in order to
1935 	 * emulate a vfs_ref().
1936 	 */
1937 	if (vp == NULL)
1938 		vfs_ref(mp);
1939 
1940 	error = vn_start_write_refed(mp, flags, false);
1941 	if (error != 0 && (flags & V_NOWAIT) == 0)
1942 		*mpp = NULL;
1943 	return (error);
1944 }
1945 
1946 /*
1947  * Secondary suspension. Used by operations such as vop_inactive
1948  * routines that are needed by the higher level functions. These
1949  * are allowed to proceed until all the higher level functions have
1950  * completed (indicated by mnt_writeopcount dropping to zero). At that
1951  * time, these operations are halted until the suspension is over.
1952  */
1953 int
1954 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1955 {
1956 	struct mount *mp;
1957 	int error, mflags;
1958 
1959 	KASSERT((flags & (~V_VALID_FLAGS | V_XSLEEP)) == 0,
1960 	    ("%s: invalid flags passed %d\n", __func__, flags));
1961 
1962  retry:
1963 	if (vp != NULL) {
1964 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1965 			*mpp = NULL;
1966 			if (error != EOPNOTSUPP)
1967 				return (error);
1968 			return (0);
1969 		}
1970 	}
1971 	/*
1972 	 * If we are not suspended or have not yet reached suspended
1973 	 * mode, then let the operation proceed.
1974 	 */
1975 	if ((mp = *mpp) == NULL)
1976 		return (0);
1977 
1978 	/*
1979 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1980 	 * a vfs_ref().
1981 	 * As long as a vnode is not provided we need to acquire a
1982 	 * refcount for the provided mountpoint too, in order to
1983 	 * emulate a vfs_ref().
1984 	 */
1985 	MNT_ILOCK(mp);
1986 	if (vp == NULL)
1987 		MNT_REF(mp);
1988 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1989 		mp->mnt_secondary_writes++;
1990 		mp->mnt_secondary_accwrites++;
1991 		MNT_IUNLOCK(mp);
1992 		return (0);
1993 	}
1994 	if ((flags & V_NOWAIT) != 0) {
1995 		MNT_REL(mp);
1996 		MNT_IUNLOCK(mp);
1997 		*mpp = NULL;
1998 		return (EWOULDBLOCK);
1999 	}
2000 	/*
2001 	 * Wait for the suspension to finish.
2002 	 */
2003 	mflags = 0;
2004 	if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) {
2005 		if ((flags & V_PCATCH) != 0)
2006 			mflags |= PCATCH;
2007 	}
2008 	mflags |= (PUSER - 1) | PDROP;
2009 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0);
2010 	vfs_rel(mp);
2011 	if (error == 0)
2012 		goto retry;
2013 	*mpp = NULL;
2014 	return (error);
2015 }
2016 
2017 /*
2018  * Filesystem write operation has completed. If we are suspending and this
2019  * operation is the last one, notify the suspender that the suspension is
2020  * now in effect.
2021  */
2022 void
2023 vn_finished_write(struct mount *mp)
2024 {
2025 	struct mount_pcpu *mpcpu;
2026 	int c;
2027 
2028 	if (mp == NULL)
2029 		return;
2030 
2031 	if (vfs_op_thread_enter(mp, mpcpu)) {
2032 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2033 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2034 		vfs_op_thread_exit(mp, mpcpu);
2035 		return;
2036 	}
2037 
2038 	MNT_ILOCK(mp);
2039 	vfs_assert_mount_counters(mp);
2040 	MNT_REL(mp);
2041 	c = --mp->mnt_writeopcount;
2042 	if (mp->mnt_vfs_ops == 0) {
2043 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2044 		MNT_IUNLOCK(mp);
2045 		return;
2046 	}
2047 	if (c < 0)
2048 		vfs_dump_mount_counters(mp);
2049 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2050 		wakeup(&mp->mnt_writeopcount);
2051 	MNT_IUNLOCK(mp);
2052 }
2053 
2054 /*
2055  * Filesystem secondary write operation has completed. If we are
2056  * suspending and this operation is the last one, notify the suspender
2057  * that the suspension is now in effect.
2058  */
2059 void
2060 vn_finished_secondary_write(struct mount *mp)
2061 {
2062 	if (mp == NULL)
2063 		return;
2064 	MNT_ILOCK(mp);
2065 	MNT_REL(mp);
2066 	mp->mnt_secondary_writes--;
2067 	if (mp->mnt_secondary_writes < 0)
2068 		panic("vn_finished_secondary_write: neg cnt");
2069 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2070 	    mp->mnt_secondary_writes <= 0)
2071 		wakeup(&mp->mnt_secondary_writes);
2072 	MNT_IUNLOCK(mp);
2073 }
2074 
2075 /*
2076  * Request a filesystem to suspend write operations.
2077  */
2078 int
2079 vfs_write_suspend(struct mount *mp, int flags)
2080 {
2081 	int error;
2082 
2083 	vfs_op_enter(mp);
2084 
2085 	MNT_ILOCK(mp);
2086 	vfs_assert_mount_counters(mp);
2087 	if (mp->mnt_susp_owner == curthread) {
2088 		vfs_op_exit_locked(mp);
2089 		MNT_IUNLOCK(mp);
2090 		return (EALREADY);
2091 	}
2092 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2093 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2094 
2095 	/*
2096 	 * Unmount holds a write reference on the mount point.  If we
2097 	 * own busy reference and drain for writers, we deadlock with
2098 	 * the reference draining in the unmount path.  Callers of
2099 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2100 	 * vfs_busy() reference is owned and caller is not in the
2101 	 * unmount context.
2102 	 */
2103 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2104 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2105 		vfs_op_exit_locked(mp);
2106 		MNT_IUNLOCK(mp);
2107 		return (EBUSY);
2108 	}
2109 
2110 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2111 	mp->mnt_susp_owner = curthread;
2112 	if (mp->mnt_writeopcount > 0)
2113 		(void) msleep(&mp->mnt_writeopcount,
2114 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2115 	else
2116 		MNT_IUNLOCK(mp);
2117 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2118 		vfs_write_resume(mp, 0);
2119 		/* vfs_write_resume does vfs_op_exit() for us */
2120 	}
2121 	return (error);
2122 }
2123 
2124 /*
2125  * Request a filesystem to resume write operations.
2126  */
2127 void
2128 vfs_write_resume(struct mount *mp, int flags)
2129 {
2130 
2131 	MNT_ILOCK(mp);
2132 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2133 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2134 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2135 				       MNTK_SUSPENDED);
2136 		mp->mnt_susp_owner = NULL;
2137 		wakeup(&mp->mnt_writeopcount);
2138 		wakeup(&mp->mnt_flag);
2139 		curthread->td_pflags &= ~TDP_IGNSUSP;
2140 		if ((flags & VR_START_WRITE) != 0) {
2141 			MNT_REF(mp);
2142 			mp->mnt_writeopcount++;
2143 		}
2144 		MNT_IUNLOCK(mp);
2145 		if ((flags & VR_NO_SUSPCLR) == 0)
2146 			VFS_SUSP_CLEAN(mp);
2147 		vfs_op_exit(mp);
2148 	} else if ((flags & VR_START_WRITE) != 0) {
2149 		MNT_REF(mp);
2150 		vn_start_write_refed(mp, 0, true);
2151 	} else {
2152 		MNT_IUNLOCK(mp);
2153 	}
2154 }
2155 
2156 /*
2157  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2158  * methods.
2159  */
2160 int
2161 vfs_write_suspend_umnt(struct mount *mp)
2162 {
2163 	int error;
2164 
2165 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2166 	    ("vfs_write_suspend_umnt: recursed"));
2167 
2168 	/* dounmount() already called vn_start_write(). */
2169 	for (;;) {
2170 		vn_finished_write(mp);
2171 		error = vfs_write_suspend(mp, 0);
2172 		if (error != 0) {
2173 			vn_start_write(NULL, &mp, V_WAIT);
2174 			return (error);
2175 		}
2176 		MNT_ILOCK(mp);
2177 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2178 			break;
2179 		MNT_IUNLOCK(mp);
2180 		vn_start_write(NULL, &mp, V_WAIT);
2181 	}
2182 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2183 	wakeup(&mp->mnt_flag);
2184 	MNT_IUNLOCK(mp);
2185 	curthread->td_pflags |= TDP_IGNSUSP;
2186 	return (0);
2187 }
2188 
2189 /*
2190  * Implement kqueues for files by translating it to vnode operation.
2191  */
2192 static int
2193 vn_kqfilter(struct file *fp, struct knote *kn)
2194 {
2195 
2196 	return (VOP_KQFILTER(fp->f_vnode, kn));
2197 }
2198 
2199 int
2200 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2201 {
2202 	if ((fp->f_flag & FKQALLOWED) == 0)
2203 		return (EBADF);
2204 	return (vn_kqfilter(fp, kn));
2205 }
2206 
2207 /*
2208  * Simplified in-kernel wrapper calls for extended attribute access.
2209  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2210  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2211  */
2212 int
2213 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2214     const char *attrname, int *buflen, char *buf, struct thread *td)
2215 {
2216 	struct uio	auio;
2217 	struct iovec	iov;
2218 	int	error;
2219 
2220 	iov.iov_len = *buflen;
2221 	iov.iov_base = buf;
2222 
2223 	auio.uio_iov = &iov;
2224 	auio.uio_iovcnt = 1;
2225 	auio.uio_rw = UIO_READ;
2226 	auio.uio_segflg = UIO_SYSSPACE;
2227 	auio.uio_td = td;
2228 	auio.uio_offset = 0;
2229 	auio.uio_resid = *buflen;
2230 
2231 	if ((ioflg & IO_NODELOCKED) == 0)
2232 		vn_lock(vp, LK_SHARED | LK_RETRY);
2233 
2234 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2235 
2236 	/* authorize attribute retrieval as kernel */
2237 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2238 	    td);
2239 
2240 	if ((ioflg & IO_NODELOCKED) == 0)
2241 		VOP_UNLOCK(vp);
2242 
2243 	if (error == 0) {
2244 		*buflen = *buflen - auio.uio_resid;
2245 	}
2246 
2247 	return (error);
2248 }
2249 
2250 /*
2251  * XXX failure mode if partially written?
2252  */
2253 int
2254 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2255     const char *attrname, int buflen, char *buf, struct thread *td)
2256 {
2257 	struct uio	auio;
2258 	struct iovec	iov;
2259 	struct mount	*mp;
2260 	int	error;
2261 
2262 	iov.iov_len = buflen;
2263 	iov.iov_base = buf;
2264 
2265 	auio.uio_iov = &iov;
2266 	auio.uio_iovcnt = 1;
2267 	auio.uio_rw = UIO_WRITE;
2268 	auio.uio_segflg = UIO_SYSSPACE;
2269 	auio.uio_td = td;
2270 	auio.uio_offset = 0;
2271 	auio.uio_resid = buflen;
2272 
2273 	if ((ioflg & IO_NODELOCKED) == 0) {
2274 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2275 			return (error);
2276 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2277 	}
2278 
2279 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2280 
2281 	/* authorize attribute setting as kernel */
2282 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2283 
2284 	if ((ioflg & IO_NODELOCKED) == 0) {
2285 		vn_finished_write(mp);
2286 		VOP_UNLOCK(vp);
2287 	}
2288 
2289 	return (error);
2290 }
2291 
2292 int
2293 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2294     const char *attrname, struct thread *td)
2295 {
2296 	struct mount	*mp;
2297 	int	error;
2298 
2299 	if ((ioflg & IO_NODELOCKED) == 0) {
2300 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2301 			return (error);
2302 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2303 	}
2304 
2305 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2306 
2307 	/* authorize attribute removal as kernel */
2308 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2309 	if (error == EOPNOTSUPP)
2310 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2311 		    NULL, td);
2312 
2313 	if ((ioflg & IO_NODELOCKED) == 0) {
2314 		vn_finished_write(mp);
2315 		VOP_UNLOCK(vp);
2316 	}
2317 
2318 	return (error);
2319 }
2320 
2321 static int
2322 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2323     struct vnode **rvp)
2324 {
2325 
2326 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2327 }
2328 
2329 int
2330 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2331 {
2332 
2333 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2334 	    lkflags, rvp));
2335 }
2336 
2337 int
2338 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2339     int lkflags, struct vnode **rvp)
2340 {
2341 	struct mount *mp;
2342 	int ltype, error;
2343 
2344 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2345 	mp = vp->v_mount;
2346 	ltype = VOP_ISLOCKED(vp);
2347 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2348 	    ("vn_vget_ino: vp not locked"));
2349 	error = vfs_busy(mp, MBF_NOWAIT);
2350 	if (error != 0) {
2351 		vfs_ref(mp);
2352 		VOP_UNLOCK(vp);
2353 		error = vfs_busy(mp, 0);
2354 		vn_lock(vp, ltype | LK_RETRY);
2355 		vfs_rel(mp);
2356 		if (error != 0)
2357 			return (ENOENT);
2358 		if (VN_IS_DOOMED(vp)) {
2359 			vfs_unbusy(mp);
2360 			return (ENOENT);
2361 		}
2362 	}
2363 	VOP_UNLOCK(vp);
2364 	error = alloc(mp, alloc_arg, lkflags, rvp);
2365 	vfs_unbusy(mp);
2366 	if (error != 0 || *rvp != vp)
2367 		vn_lock(vp, ltype | LK_RETRY);
2368 	if (VN_IS_DOOMED(vp)) {
2369 		if (error == 0) {
2370 			if (*rvp == vp)
2371 				vunref(vp);
2372 			else
2373 				vput(*rvp);
2374 		}
2375 		error = ENOENT;
2376 	}
2377 	return (error);
2378 }
2379 
2380 static void
2381 vn_send_sigxfsz(struct proc *p)
2382 {
2383 	PROC_LOCK(p);
2384 	kern_psignal(p, SIGXFSZ);
2385 	PROC_UNLOCK(p);
2386 }
2387 
2388 int
2389 vn_rlimit_trunc(u_quad_t size, struct thread *td)
2390 {
2391 	if (size <= lim_cur(td, RLIMIT_FSIZE))
2392 		return (0);
2393 	vn_send_sigxfsz(td->td_proc);
2394 	return (EFBIG);
2395 }
2396 
2397 static int
2398 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2399     bool adj, struct thread *td)
2400 {
2401 	off_t lim;
2402 	bool ktr_write;
2403 
2404 	if (vp->v_type != VREG)
2405 		return (0);
2406 
2407 	/*
2408 	 * Handle file system maximum file size.
2409 	 */
2410 	if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) {
2411 		if (!adj || uio->uio_offset >= maxfsz)
2412 			return (EFBIG);
2413 		uio->uio_resid = maxfsz - uio->uio_offset;
2414 	}
2415 
2416 	/*
2417 	 * This is kernel write (e.g. vnode_pager) or accounting
2418 	 * write, ignore limit.
2419 	 */
2420 	if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0)
2421 		return (0);
2422 
2423 	/*
2424 	 * Calculate file size limit.
2425 	 */
2426 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2427 	lim = __predict_false(ktr_write) ? td->td_ktr_io_lim :
2428 	    lim_cur(td, RLIMIT_FSIZE);
2429 
2430 	/*
2431 	 * Is the limit reached?
2432 	 */
2433 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2434 		return (0);
2435 
2436 	/*
2437 	 * Prepared filesystems can handle writes truncated to the
2438 	 * file size limit.
2439 	 */
2440 	if (adj && (uoff_t)uio->uio_offset < lim) {
2441 		uio->uio_resid = lim - (uoff_t)uio->uio_offset;
2442 		return (0);
2443 	}
2444 
2445 	if (!ktr_write || ktr_filesize_limit_signal)
2446 		vn_send_sigxfsz(td->td_proc);
2447 	return (EFBIG);
2448 }
2449 
2450 /*
2451  * Helper for VOP_WRITE() implementations, the common code to
2452  * handle maximum supported file size on the filesystem, and
2453  * RLIMIT_FSIZE, except for special writes from accounting subsystem
2454  * and ktrace.
2455  *
2456  * For maximum file size (maxfsz argument):
2457  * - return EFBIG if uio_offset is beyond it
2458  * - otherwise, clamp uio_resid if write would extend file beyond maxfsz.
2459  *
2460  * For RLIMIT_FSIZE:
2461  * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit
2462  * - otherwise, clamp uio_resid if write would extend file beyond limit.
2463  *
2464  * If clamping occured, the adjustment for uio_resid is stored in
2465  * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return
2466  * from the VOP.
2467  */
2468 int
2469 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz,
2470     ssize_t *resid_adj, struct thread *td)
2471 {
2472 	ssize_t resid_orig;
2473 	int error;
2474 	bool adj;
2475 
2476 	resid_orig = uio->uio_resid;
2477 	adj = resid_adj != NULL;
2478 	error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td);
2479 	if (adj)
2480 		*resid_adj = resid_orig - uio->uio_resid;
2481 	return (error);
2482 }
2483 
2484 void
2485 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj)
2486 {
2487 	uio->uio_resid += resid_adj;
2488 }
2489 
2490 int
2491 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2492     struct thread *td)
2493 {
2494 	return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL,
2495 	    td));
2496 }
2497 
2498 int
2499 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2500     struct thread *td)
2501 {
2502 	struct vnode *vp;
2503 
2504 	vp = fp->f_vnode;
2505 #ifdef AUDIT
2506 	vn_lock(vp, LK_SHARED | LK_RETRY);
2507 	AUDIT_ARG_VNODE1(vp);
2508 	VOP_UNLOCK(vp);
2509 #endif
2510 	return (setfmode(td, active_cred, vp, mode));
2511 }
2512 
2513 int
2514 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2515     struct thread *td)
2516 {
2517 	struct vnode *vp;
2518 
2519 	vp = fp->f_vnode;
2520 #ifdef AUDIT
2521 	vn_lock(vp, LK_SHARED | LK_RETRY);
2522 	AUDIT_ARG_VNODE1(vp);
2523 	VOP_UNLOCK(vp);
2524 #endif
2525 	return (setfown(td, active_cred, vp, uid, gid));
2526 }
2527 
2528 /*
2529  * Remove pages in the range ["start", "end") from the vnode's VM object.  If
2530  * "end" is 0, then the range extends to the end of the object.
2531  */
2532 void
2533 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2534 {
2535 	vm_object_t object;
2536 
2537 	if ((object = vp->v_object) == NULL)
2538 		return;
2539 	VM_OBJECT_WLOCK(object);
2540 	vm_object_page_remove(object, start, end, 0);
2541 	VM_OBJECT_WUNLOCK(object);
2542 }
2543 
2544 /*
2545  * Like vn_pages_remove(), but skips invalid pages, which by definition are not
2546  * mapped into any process' address space.  Filesystems may use this in
2547  * preference to vn_pages_remove() to avoid blocking on pages busied in
2548  * preparation for a VOP_GETPAGES.
2549  */
2550 void
2551 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2552 {
2553 	vm_object_t object;
2554 
2555 	if ((object = vp->v_object) == NULL)
2556 		return;
2557 	VM_OBJECT_WLOCK(object);
2558 	vm_object_page_remove(object, start, end, OBJPR_VALIDONLY);
2559 	VM_OBJECT_WUNLOCK(object);
2560 }
2561 
2562 int
2563 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off,
2564     struct ucred *cred)
2565 {
2566 	vm_object_t obj;
2567 	off_t size;
2568 	daddr_t bn, bnp;
2569 	uint64_t bsize;
2570 	off_t noff;
2571 	int error;
2572 
2573 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2574 	    ("%s: Wrong command %lu", __func__, cmd));
2575 	ASSERT_VOP_ELOCKED(vp, "vn_bmap_seekhole_locked");
2576 
2577 	if (vp->v_type != VREG) {
2578 		error = ENOTTY;
2579 		goto out;
2580 	}
2581 	error = vn_getsize_locked(vp, &size, cred);
2582 	if (error != 0)
2583 		goto out;
2584 	noff = *off;
2585 	if (noff < 0 || noff >= size) {
2586 		error = ENXIO;
2587 		goto out;
2588 	}
2589 
2590 	/* See the comment in ufs_bmap_seekdata(). */
2591 	obj = vp->v_object;
2592 	if (obj != NULL) {
2593 		VM_OBJECT_WLOCK(obj);
2594 		vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
2595 		VM_OBJECT_WUNLOCK(obj);
2596 	}
2597 
2598 	bsize = vp->v_mount->mnt_stat.f_iosize;
2599 	for (bn = noff / bsize; noff < size; bn++, noff += bsize -
2600 	    noff % bsize) {
2601 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2602 		if (error == EOPNOTSUPP) {
2603 			error = ENOTTY;
2604 			goto out;
2605 		}
2606 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2607 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2608 			noff = bn * bsize;
2609 			if (noff < *off)
2610 				noff = *off;
2611 			goto out;
2612 		}
2613 	}
2614 	if (noff > size)
2615 		noff = size;
2616 	/* noff == size. There is an implicit hole at the end of file. */
2617 	if (cmd == FIOSEEKDATA)
2618 		error = ENXIO;
2619 out:
2620 	if (error == 0)
2621 		*off = noff;
2622 	return (error);
2623 }
2624 
2625 int
2626 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2627 {
2628 	int error;
2629 
2630 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2631 	    ("%s: Wrong command %lu", __func__, cmd));
2632 
2633 	if (vn_lock(vp, LK_EXCLUSIVE) != 0)
2634 		return (EBADF);
2635 	error = vn_bmap_seekhole_locked(vp, cmd, off, cred);
2636 	VOP_UNLOCK(vp);
2637 	return (error);
2638 }
2639 
2640 int
2641 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2642 {
2643 	struct ucred *cred;
2644 	struct vnode *vp;
2645 	off_t foffset, fsize, size;
2646 	int error, noneg;
2647 
2648 	cred = td->td_ucred;
2649 	vp = fp->f_vnode;
2650 	foffset = foffset_lock(fp, 0);
2651 	noneg = (vp->v_type != VCHR);
2652 	error = 0;
2653 	switch (whence) {
2654 	case L_INCR:
2655 		if (noneg &&
2656 		    (foffset < 0 ||
2657 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2658 			error = EOVERFLOW;
2659 			break;
2660 		}
2661 		offset += foffset;
2662 		break;
2663 	case L_XTND:
2664 		error = vn_getsize(vp, &fsize, cred);
2665 		if (error != 0)
2666 			break;
2667 
2668 		/*
2669 		 * If the file references a disk device, then fetch
2670 		 * the media size and use that to determine the ending
2671 		 * offset.
2672 		 */
2673 		if (fsize == 0 && vp->v_type == VCHR &&
2674 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2675 			fsize = size;
2676 		if (noneg && offset > 0 && fsize > OFF_MAX - offset) {
2677 			error = EOVERFLOW;
2678 			break;
2679 		}
2680 		offset += fsize;
2681 		break;
2682 	case L_SET:
2683 		break;
2684 	case SEEK_DATA:
2685 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2686 		if (error == ENOTTY)
2687 			error = EINVAL;
2688 		break;
2689 	case SEEK_HOLE:
2690 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2691 		if (error == ENOTTY)
2692 			error = EINVAL;
2693 		break;
2694 	default:
2695 		error = EINVAL;
2696 	}
2697 	if (error == 0 && noneg && offset < 0)
2698 		error = EINVAL;
2699 	if (error != 0)
2700 		goto drop;
2701 	VFS_KNOTE_UNLOCKED(vp, 0);
2702 	td->td_uretoff.tdu_off = offset;
2703 drop:
2704 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2705 	return (error);
2706 }
2707 
2708 int
2709 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2710     struct thread *td)
2711 {
2712 	int error;
2713 
2714 	/*
2715 	 * Grant permission if the caller is the owner of the file, or
2716 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2717 	 * on the file.  If the time pointer is null, then write
2718 	 * permission on the file is also sufficient.
2719 	 *
2720 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2721 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2722 	 * will be allowed to set the times [..] to the current
2723 	 * server time.
2724 	 */
2725 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2726 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2727 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2728 	return (error);
2729 }
2730 
2731 int
2732 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2733 {
2734 	struct vnode *vp;
2735 	int error;
2736 
2737 	if (fp->f_type == DTYPE_FIFO)
2738 		kif->kf_type = KF_TYPE_FIFO;
2739 	else
2740 		kif->kf_type = KF_TYPE_VNODE;
2741 	vp = fp->f_vnode;
2742 	vref(vp);
2743 	FILEDESC_SUNLOCK(fdp);
2744 	error = vn_fill_kinfo_vnode(vp, kif);
2745 	vrele(vp);
2746 	FILEDESC_SLOCK(fdp);
2747 	return (error);
2748 }
2749 
2750 static inline void
2751 vn_fill_junk(struct kinfo_file *kif)
2752 {
2753 	size_t len, olen;
2754 
2755 	/*
2756 	 * Simulate vn_fullpath returning changing values for a given
2757 	 * vp during e.g. coredump.
2758 	 */
2759 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2760 	olen = strlen(kif->kf_path);
2761 	if (len < olen)
2762 		strcpy(&kif->kf_path[len - 1], "$");
2763 	else
2764 		for (; olen < len; olen++)
2765 			strcpy(&kif->kf_path[olen], "A");
2766 }
2767 
2768 int
2769 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2770 {
2771 	struct vattr va;
2772 	char *fullpath, *freepath;
2773 	int error;
2774 
2775 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2776 	freepath = NULL;
2777 	fullpath = "-";
2778 	error = vn_fullpath(vp, &fullpath, &freepath);
2779 	if (error == 0) {
2780 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2781 	}
2782 	if (freepath != NULL)
2783 		free(freepath, M_TEMP);
2784 
2785 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2786 		vn_fill_junk(kif);
2787 	);
2788 
2789 	/*
2790 	 * Retrieve vnode attributes.
2791 	 */
2792 	va.va_fsid = VNOVAL;
2793 	va.va_rdev = NODEV;
2794 	vn_lock(vp, LK_SHARED | LK_RETRY);
2795 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2796 	VOP_UNLOCK(vp);
2797 	if (error != 0)
2798 		return (error);
2799 	if (va.va_fsid != VNOVAL)
2800 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2801 	else
2802 		kif->kf_un.kf_file.kf_file_fsid =
2803 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2804 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2805 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2806 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2807 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2808 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2809 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2810 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2811 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2812 	kif->kf_un.kf_file.kf_file_nlink = va.va_nlink;
2813 	return (0);
2814 }
2815 
2816 int
2817 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2818     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2819     struct thread *td)
2820 {
2821 #ifdef HWPMC_HOOKS
2822 	struct pmckern_map_in pkm;
2823 #endif
2824 	struct mount *mp;
2825 	struct vnode *vp;
2826 	vm_object_t object;
2827 	vm_prot_t maxprot;
2828 	boolean_t writecounted;
2829 	int error;
2830 
2831 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2832     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2833 	/*
2834 	 * POSIX shared-memory objects are defined to have
2835 	 * kernel persistence, and are not defined to support
2836 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2837 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2838 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2839 	 * flag to request this behavior.
2840 	 */
2841 	if ((fp->f_flag & FPOSIXSHM) != 0)
2842 		flags |= MAP_NOSYNC;
2843 #endif
2844 	vp = fp->f_vnode;
2845 
2846 	/*
2847 	 * Ensure that file and memory protections are
2848 	 * compatible.  Note that we only worry about
2849 	 * writability if mapping is shared; in this case,
2850 	 * current and max prot are dictated by the open file.
2851 	 * XXX use the vnode instead?  Problem is: what
2852 	 * credentials do we use for determination? What if
2853 	 * proc does a setuid?
2854 	 */
2855 	mp = vp->v_mount;
2856 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2857 		maxprot = VM_PROT_NONE;
2858 		if ((prot & VM_PROT_EXECUTE) != 0)
2859 			return (EACCES);
2860 	} else
2861 		maxprot = VM_PROT_EXECUTE;
2862 	if ((fp->f_flag & FREAD) != 0)
2863 		maxprot |= VM_PROT_READ;
2864 	else if ((prot & VM_PROT_READ) != 0)
2865 		return (EACCES);
2866 
2867 	/*
2868 	 * If we are sharing potential changes via MAP_SHARED and we
2869 	 * are trying to get write permission although we opened it
2870 	 * without asking for it, bail out.
2871 	 */
2872 	if ((flags & MAP_SHARED) != 0) {
2873 		if ((fp->f_flag & FWRITE) != 0)
2874 			maxprot |= VM_PROT_WRITE;
2875 		else if ((prot & VM_PROT_WRITE) != 0)
2876 			return (EACCES);
2877 	} else {
2878 		maxprot |= VM_PROT_WRITE;
2879 		cap_maxprot |= VM_PROT_WRITE;
2880 	}
2881 	maxprot &= cap_maxprot;
2882 
2883 	/*
2884 	 * For regular files and shared memory, POSIX requires that
2885 	 * the value of foff be a legitimate offset within the data
2886 	 * object.  In particular, negative offsets are invalid.
2887 	 * Blocking negative offsets and overflows here avoids
2888 	 * possible wraparound or user-level access into reserved
2889 	 * ranges of the data object later.  In contrast, POSIX does
2890 	 * not dictate how offsets are used by device drivers, so in
2891 	 * the case of a device mapping a negative offset is passed
2892 	 * on.
2893 	 */
2894 	if (
2895 #ifdef _LP64
2896 	    size > OFF_MAX ||
2897 #endif
2898 	    foff > OFF_MAX - size)
2899 		return (EINVAL);
2900 
2901 	writecounted = FALSE;
2902 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2903 	    &foff, &object, &writecounted);
2904 	if (error != 0)
2905 		return (error);
2906 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2907 	    foff, writecounted, td);
2908 	if (error != 0) {
2909 		/*
2910 		 * If this mapping was accounted for in the vnode's
2911 		 * writecount, then undo that now.
2912 		 */
2913 		if (writecounted)
2914 			vm_pager_release_writecount(object, 0, size);
2915 		vm_object_deallocate(object);
2916 	}
2917 #ifdef HWPMC_HOOKS
2918 	/* Inform hwpmc(4) if an executable is being mapped. */
2919 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2920 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2921 			pkm.pm_file = vp;
2922 			pkm.pm_address = (uintptr_t) *addr;
2923 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2924 		}
2925 	}
2926 #endif
2927 	return (error);
2928 }
2929 
2930 void
2931 vn_fsid(struct vnode *vp, struct vattr *va)
2932 {
2933 	fsid_t *f;
2934 
2935 	f = &vp->v_mount->mnt_stat.f_fsid;
2936 	va->va_fsid = (uint32_t)f->val[1];
2937 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2938 	va->va_fsid += (uint32_t)f->val[0];
2939 }
2940 
2941 int
2942 vn_fsync_buf(struct vnode *vp, int waitfor)
2943 {
2944 	struct buf *bp, *nbp;
2945 	struct bufobj *bo;
2946 	struct mount *mp;
2947 	int error, maxretry;
2948 
2949 	error = 0;
2950 	maxretry = 10000;     /* large, arbitrarily chosen */
2951 	mp = NULL;
2952 	if (vp->v_type == VCHR) {
2953 		VI_LOCK(vp);
2954 		mp = vp->v_rdev->si_mountpt;
2955 		VI_UNLOCK(vp);
2956 	}
2957 	bo = &vp->v_bufobj;
2958 	BO_LOCK(bo);
2959 loop1:
2960 	/*
2961 	 * MARK/SCAN initialization to avoid infinite loops.
2962 	 */
2963         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2964 		bp->b_vflags &= ~BV_SCANNED;
2965 		bp->b_error = 0;
2966 	}
2967 
2968 	/*
2969 	 * Flush all dirty buffers associated with a vnode.
2970 	 */
2971 loop2:
2972 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2973 		if ((bp->b_vflags & BV_SCANNED) != 0)
2974 			continue;
2975 		bp->b_vflags |= BV_SCANNED;
2976 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2977 			if (waitfor != MNT_WAIT)
2978 				continue;
2979 			if (BUF_LOCK(bp,
2980 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2981 			    BO_LOCKPTR(bo)) != 0) {
2982 				BO_LOCK(bo);
2983 				goto loop1;
2984 			}
2985 			BO_LOCK(bo);
2986 		}
2987 		BO_UNLOCK(bo);
2988 		KASSERT(bp->b_bufobj == bo,
2989 		    ("bp %p wrong b_bufobj %p should be %p",
2990 		    bp, bp->b_bufobj, bo));
2991 		if ((bp->b_flags & B_DELWRI) == 0)
2992 			panic("fsync: not dirty");
2993 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2994 			vfs_bio_awrite(bp);
2995 		} else {
2996 			bremfree(bp);
2997 			bawrite(bp);
2998 		}
2999 		if (maxretry < 1000)
3000 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
3001 		BO_LOCK(bo);
3002 		goto loop2;
3003 	}
3004 
3005 	/*
3006 	 * If synchronous the caller expects us to completely resolve all
3007 	 * dirty buffers in the system.  Wait for in-progress I/O to
3008 	 * complete (which could include background bitmap writes), then
3009 	 * retry if dirty blocks still exist.
3010 	 */
3011 	if (waitfor == MNT_WAIT) {
3012 		bufobj_wwait(bo, 0, 0);
3013 		if (bo->bo_dirty.bv_cnt > 0) {
3014 			/*
3015 			 * If we are unable to write any of these buffers
3016 			 * then we fail now rather than trying endlessly
3017 			 * to write them out.
3018 			 */
3019 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
3020 				if ((error = bp->b_error) != 0)
3021 					break;
3022 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
3023 			    (error == 0 && --maxretry >= 0))
3024 				goto loop1;
3025 			if (error == 0)
3026 				error = EAGAIN;
3027 		}
3028 	}
3029 	BO_UNLOCK(bo);
3030 	if (error != 0)
3031 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
3032 
3033 	return (error);
3034 }
3035 
3036 /*
3037  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
3038  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
3039  * to do the actual copy.
3040  * vn_generic_copy_file_range() is factored out, so it can be called
3041  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
3042  * different file systems.
3043  */
3044 int
3045 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
3046     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
3047     struct ucred *outcred, struct thread *fsize_td)
3048 {
3049 	int error;
3050 	size_t len;
3051 	uint64_t uval;
3052 
3053 	len = *lenp;
3054 	*lenp = 0;		/* For error returns. */
3055 	error = 0;
3056 
3057 	/* Do some sanity checks on the arguments. */
3058 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
3059 		error = EISDIR;
3060 	else if (*inoffp < 0 || *outoffp < 0 ||
3061 	    invp->v_type != VREG || outvp->v_type != VREG)
3062 		error = EINVAL;
3063 	if (error != 0)
3064 		goto out;
3065 
3066 	/* Ensure offset + len does not wrap around. */
3067 	uval = *inoffp;
3068 	uval += len;
3069 	if (uval > INT64_MAX)
3070 		len = INT64_MAX - *inoffp;
3071 	uval = *outoffp;
3072 	uval += len;
3073 	if (uval > INT64_MAX)
3074 		len = INT64_MAX - *outoffp;
3075 	if (len == 0)
3076 		goto out;
3077 
3078 	/*
3079 	 * If the two vnodes are for the same file system type, call
3080 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
3081 	 * which can handle copies across multiple file system types.
3082 	 */
3083 	*lenp = len;
3084 	if (invp->v_mount == outvp->v_mount ||
3085 	    strcmp(invp->v_mount->mnt_vfc->vfc_name,
3086 	    outvp->v_mount->mnt_vfc->vfc_name) == 0)
3087 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
3088 		    lenp, flags, incred, outcred, fsize_td);
3089 	else
3090 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
3091 		    outoffp, lenp, flags, incred, outcred, fsize_td);
3092 out:
3093 	return (error);
3094 }
3095 
3096 /*
3097  * Test len bytes of data starting at dat for all bytes == 0.
3098  * Return true if all bytes are zero, false otherwise.
3099  * Expects dat to be well aligned.
3100  */
3101 static bool
3102 mem_iszero(void *dat, int len)
3103 {
3104 	int i;
3105 	const u_int *p;
3106 	const char *cp;
3107 
3108 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
3109 		if (len >= sizeof(*p)) {
3110 			if (*p != 0)
3111 				return (false);
3112 		} else {
3113 			cp = (const char *)p;
3114 			for (i = 0; i < len; i++, cp++)
3115 				if (*cp != '\0')
3116 					return (false);
3117 		}
3118 	}
3119 	return (true);
3120 }
3121 
3122 /*
3123  * Look for a hole in the output file and, if found, adjust *outoffp
3124  * and *xferp to skip past the hole.
3125  * *xferp is the entire hole length to be written and xfer2 is how many bytes
3126  * to be written as 0's upon return.
3127  */
3128 static off_t
3129 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
3130     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
3131 {
3132 	int error;
3133 	off_t delta;
3134 
3135 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
3136 		*dataoffp = *outoffp;
3137 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
3138 		    curthread);
3139 		if (error == 0) {
3140 			*holeoffp = *dataoffp;
3141 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
3142 			    curthread);
3143 		}
3144 		if (error != 0 || *holeoffp == *dataoffp) {
3145 			/*
3146 			 * Since outvp is unlocked, it may be possible for
3147 			 * another thread to do a truncate(), lseek(), write()
3148 			 * creating a hole at startoff between the above
3149 			 * VOP_IOCTL() calls, if the other thread does not do
3150 			 * rangelocking.
3151 			 * If that happens, *holeoffp == *dataoffp and finding
3152 			 * the hole has failed, so disable vn_skip_hole().
3153 			 */
3154 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3155 			return (xfer2);
3156 		}
3157 		KASSERT(*dataoffp >= *outoffp,
3158 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3159 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3160 		KASSERT(*holeoffp > *dataoffp,
3161 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3162 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3163 	}
3164 
3165 	/*
3166 	 * If there is a hole before the data starts, advance *outoffp and
3167 	 * *xferp past the hole.
3168 	 */
3169 	if (*dataoffp > *outoffp) {
3170 		delta = *dataoffp - *outoffp;
3171 		if (delta >= *xferp) {
3172 			/* Entire *xferp is a hole. */
3173 			*outoffp += *xferp;
3174 			*xferp = 0;
3175 			return (0);
3176 		}
3177 		*xferp -= delta;
3178 		*outoffp += delta;
3179 		xfer2 = MIN(xfer2, *xferp);
3180 	}
3181 
3182 	/*
3183 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3184 	 * that the write ends at the start of the hole.
3185 	 * *holeoffp should always be greater than *outoffp, but for the
3186 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3187 	 * value.
3188 	 */
3189 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3190 		xfer2 = *holeoffp - *outoffp;
3191 	return (xfer2);
3192 }
3193 
3194 /*
3195  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3196  * dat is a maximum of blksize in length and can be written repeatedly in
3197  * the chunk.
3198  * If growfile == true, just grow the file via vn_truncate_locked() instead
3199  * of doing actual writes.
3200  * If checkhole == true, a hole is being punched, so skip over any hole
3201  * already in the output file.
3202  */
3203 static int
3204 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3205     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3206 {
3207 	struct mount *mp;
3208 	off_t dataoff, holeoff, xfer2;
3209 	int error;
3210 
3211 	/*
3212 	 * Loop around doing writes of blksize until write has been completed.
3213 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3214 	 * done for each iteration, since the xfer argument can be very
3215 	 * large if there is a large hole to punch in the output file.
3216 	 */
3217 	error = 0;
3218 	holeoff = 0;
3219 	do {
3220 		xfer2 = MIN(xfer, blksize);
3221 		if (checkhole) {
3222 			/*
3223 			 * Punching a hole.  Skip writing if there is
3224 			 * already a hole in the output file.
3225 			 */
3226 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3227 			    &dataoff, &holeoff, cred);
3228 			if (xfer == 0)
3229 				break;
3230 			if (holeoff < 0)
3231 				checkhole = false;
3232 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3233 			    (intmax_t)xfer2));
3234 		}
3235 		bwillwrite();
3236 		mp = NULL;
3237 		error = vn_start_write(outvp, &mp, V_WAIT);
3238 		if (error != 0)
3239 			break;
3240 		if (growfile) {
3241 			error = vn_lock(outvp, LK_EXCLUSIVE);
3242 			if (error == 0) {
3243 				error = vn_truncate_locked(outvp, outoff + xfer,
3244 				    false, cred);
3245 				VOP_UNLOCK(outvp);
3246 			}
3247 		} else {
3248 			error = vn_lock(outvp, vn_lktype_write(mp, outvp));
3249 			if (error == 0) {
3250 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3251 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3252 				    curthread->td_ucred, cred, NULL, curthread);
3253 				outoff += xfer2;
3254 				xfer -= xfer2;
3255 				VOP_UNLOCK(outvp);
3256 			}
3257 		}
3258 		if (mp != NULL)
3259 			vn_finished_write(mp);
3260 	} while (!growfile && xfer > 0 && error == 0);
3261 	return (error);
3262 }
3263 
3264 /*
3265  * Copy a byte range of one file to another.  This function can handle the
3266  * case where invp and outvp are on different file systems.
3267  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3268  * is no better file system specific way to do it.
3269  */
3270 int
3271 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3272     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3273     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3274 {
3275 	struct mount *mp;
3276 	off_t startoff, endoff, xfer, xfer2;
3277 	u_long blksize;
3278 	int error, interrupted;
3279 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3280 	ssize_t aresid, r = 0;
3281 	size_t copylen, len, savlen;
3282 	off_t insize, outsize;
3283 	char *dat;
3284 	long holein, holeout;
3285 	struct timespec curts, endts;
3286 
3287 	holein = holeout = 0;
3288 	savlen = len = *lenp;
3289 	error = 0;
3290 	interrupted = 0;
3291 	dat = NULL;
3292 
3293 	error = vn_lock(invp, LK_SHARED);
3294 	if (error != 0)
3295 		goto out;
3296 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3297 		holein = 0;
3298 	if (holein > 0)
3299 		error = vn_getsize_locked(invp, &insize, incred);
3300 	VOP_UNLOCK(invp);
3301 	if (error != 0)
3302 		goto out;
3303 
3304 	mp = NULL;
3305 	error = vn_start_write(outvp, &mp, V_WAIT);
3306 	if (error == 0)
3307 		error = vn_lock(outvp, LK_EXCLUSIVE);
3308 	if (error == 0) {
3309 		/*
3310 		 * If fsize_td != NULL, do a vn_rlimit_fsizex() call,
3311 		 * now that outvp is locked.
3312 		 */
3313 		if (fsize_td != NULL) {
3314 			struct uio io;
3315 
3316 			io.uio_offset = *outoffp;
3317 			io.uio_resid = len;
3318 			error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td);
3319 			len = savlen = io.uio_resid;
3320 			/*
3321 			 * No need to call vn_rlimit_fsizex_res before return,
3322 			 * since the uio is local.
3323 			 */
3324 		}
3325 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3326 			holeout = 0;
3327 		/*
3328 		 * Holes that are past EOF do not need to be written as a block
3329 		 * of zero bytes.  So, truncate the output file as far as
3330 		 * possible and then use size to decide if writing 0
3331 		 * bytes is necessary in the loop below.
3332 		 */
3333 		if (error == 0)
3334 			error = vn_getsize_locked(outvp, &outsize, outcred);
3335 		if (error == 0 && outsize > *outoffp && outsize <= *outoffp + len) {
3336 #ifdef MAC
3337 			error = mac_vnode_check_write(curthread->td_ucred,
3338 			    outcred, outvp);
3339 			if (error == 0)
3340 #endif
3341 				error = vn_truncate_locked(outvp, *outoffp,
3342 				    false, outcred);
3343 			if (error == 0)
3344 				outsize = *outoffp;
3345 		}
3346 		VOP_UNLOCK(outvp);
3347 	}
3348 	if (mp != NULL)
3349 		vn_finished_write(mp);
3350 	if (error != 0)
3351 		goto out;
3352 
3353 	if (holein == 0 && holeout > 0) {
3354 		/*
3355 		 * For this special case, the input data will be scanned
3356 		 * for blocks of all 0 bytes.  For these blocks, the
3357 		 * write can be skipped for the output file to create
3358 		 * an unallocated region.
3359 		 * Therefore, use the appropriate size for the output file.
3360 		 */
3361 		blksize = holeout;
3362 		if (blksize <= 512) {
3363 			/*
3364 			 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE
3365 			 * of 512, although it actually only creates
3366 			 * unallocated regions for blocks >= f_iosize.
3367 			 */
3368 			blksize = outvp->v_mount->mnt_stat.f_iosize;
3369 		}
3370 	} else {
3371 		/*
3372 		 * Use the larger of the two f_iosize values.  If they are
3373 		 * not the same size, one will normally be an exact multiple of
3374 		 * the other, since they are both likely to be a power of 2.
3375 		 */
3376 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3377 		    outvp->v_mount->mnt_stat.f_iosize);
3378 	}
3379 
3380 	/* Clip to sane limits. */
3381 	if (blksize < 4096)
3382 		blksize = 4096;
3383 	else if (blksize > maxphys)
3384 		blksize = maxphys;
3385 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3386 
3387 	/*
3388 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3389 	 * to find holes.  Otherwise, just scan the read block for all 0s
3390 	 * in the inner loop where the data copying is done.
3391 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3392 	 * support holes on the server, but do not support FIOSEEKHOLE.
3393 	 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate
3394 	 * that this function should return after 1second with a partial
3395 	 * completion.
3396 	 */
3397 	if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) {
3398 		getnanouptime(&endts);
3399 		endts.tv_sec++;
3400 	} else
3401 		timespecclear(&endts);
3402 	holetoeof = eof = false;
3403 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3404 		endoff = 0;			/* To shut up compilers. */
3405 		cantseek = true;
3406 		startoff = *inoffp;
3407 		copylen = len;
3408 
3409 		/*
3410 		 * Find the next data area.  If there is just a hole to EOF,
3411 		 * FIOSEEKDATA should fail with ENXIO.
3412 		 * (I do not know if any file system will report a hole to
3413 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3414 		 *  will fail for those file systems.)
3415 		 *
3416 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3417 		 * the code just falls through to the inner copy loop.
3418 		 */
3419 		error = EINVAL;
3420 		if (holein > 0) {
3421 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3422 			    incred, curthread);
3423 			if (error == ENXIO) {
3424 				startoff = endoff = insize;
3425 				eof = holetoeof = true;
3426 				error = 0;
3427 			}
3428 		}
3429 		if (error == 0 && !holetoeof) {
3430 			endoff = startoff;
3431 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3432 			    incred, curthread);
3433 			/*
3434 			 * Since invp is unlocked, it may be possible for
3435 			 * another thread to do a truncate(), lseek(), write()
3436 			 * creating a hole at startoff between the above
3437 			 * VOP_IOCTL() calls, if the other thread does not do
3438 			 * rangelocking.
3439 			 * If that happens, startoff == endoff and finding
3440 			 * the hole has failed, so set an error.
3441 			 */
3442 			if (error == 0 && startoff == endoff)
3443 				error = EINVAL; /* Any error. Reset to 0. */
3444 		}
3445 		if (error == 0) {
3446 			if (startoff > *inoffp) {
3447 				/* Found hole before data block. */
3448 				xfer = MIN(startoff - *inoffp, len);
3449 				if (*outoffp < outsize) {
3450 					/* Must write 0s to punch hole. */
3451 					xfer2 = MIN(outsize - *outoffp,
3452 					    xfer);
3453 					memset(dat, 0, MIN(xfer2, blksize));
3454 					error = vn_write_outvp(outvp, dat,
3455 					    *outoffp, xfer2, blksize, false,
3456 					    holeout > 0, outcred);
3457 				}
3458 
3459 				if (error == 0 && *outoffp + xfer >
3460 				    outsize && (xfer == len || holetoeof)) {
3461 					/* Grow output file (hole at end). */
3462 					error = vn_write_outvp(outvp, dat,
3463 					    *outoffp, xfer, blksize, true,
3464 					    false, outcred);
3465 				}
3466 				if (error == 0) {
3467 					*inoffp += xfer;
3468 					*outoffp += xfer;
3469 					len -= xfer;
3470 					if (len < savlen) {
3471 						interrupted = sig_intr();
3472 						if (timespecisset(&endts) &&
3473 						    interrupted == 0) {
3474 							getnanouptime(&curts);
3475 							if (timespeccmp(&curts,
3476 							    &endts, >=))
3477 								interrupted =
3478 								    EINTR;
3479 						}
3480 					}
3481 				}
3482 			}
3483 			copylen = MIN(len, endoff - startoff);
3484 			cantseek = false;
3485 		} else {
3486 			cantseek = true;
3487 			startoff = *inoffp;
3488 			copylen = len;
3489 			error = 0;
3490 		}
3491 
3492 		xfer = blksize;
3493 		if (cantseek) {
3494 			/*
3495 			 * Set first xfer to end at a block boundary, so that
3496 			 * holes are more likely detected in the loop below via
3497 			 * the for all bytes 0 method.
3498 			 */
3499 			xfer -= (*inoffp % blksize);
3500 		}
3501 		/* Loop copying the data block. */
3502 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3503 			if (copylen < xfer)
3504 				xfer = copylen;
3505 			error = vn_lock(invp, LK_SHARED);
3506 			if (error != 0)
3507 				goto out;
3508 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3509 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3510 			    curthread->td_ucred, incred, &aresid,
3511 			    curthread);
3512 			VOP_UNLOCK(invp);
3513 			lastblock = false;
3514 			if (error == 0 && aresid > 0) {
3515 				/* Stop the copy at EOF on the input file. */
3516 				xfer -= aresid;
3517 				eof = true;
3518 				lastblock = true;
3519 			}
3520 			if (error == 0) {
3521 				/*
3522 				 * Skip the write for holes past the initial EOF
3523 				 * of the output file, unless this is the last
3524 				 * write of the output file at EOF.
3525 				 */
3526 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3527 				    false;
3528 				if (xfer == len)
3529 					lastblock = true;
3530 				if (!cantseek || *outoffp < outsize ||
3531 				    lastblock || !readzeros)
3532 					error = vn_write_outvp(outvp, dat,
3533 					    *outoffp, xfer, blksize,
3534 					    readzeros && lastblock &&
3535 					    *outoffp >= outsize, false,
3536 					    outcred);
3537 				if (error == 0) {
3538 					*inoffp += xfer;
3539 					startoff += xfer;
3540 					*outoffp += xfer;
3541 					copylen -= xfer;
3542 					len -= xfer;
3543 					if (len < savlen) {
3544 						interrupted = sig_intr();
3545 						if (timespecisset(&endts) &&
3546 						    interrupted == 0) {
3547 							getnanouptime(&curts);
3548 							if (timespeccmp(&curts,
3549 							    &endts, >=))
3550 								interrupted =
3551 								    EINTR;
3552 						}
3553 					}
3554 				}
3555 			}
3556 			xfer = blksize;
3557 		}
3558 	}
3559 out:
3560 	*lenp = savlen - len;
3561 	free(dat, M_TEMP);
3562 	return (error);
3563 }
3564 
3565 static int
3566 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3567 {
3568 	struct mount *mp;
3569 	struct vnode *vp;
3570 	off_t olen, ooffset;
3571 	int error;
3572 #ifdef AUDIT
3573 	int audited_vnode1 = 0;
3574 #endif
3575 
3576 	vp = fp->f_vnode;
3577 	if (vp->v_type != VREG)
3578 		return (ENODEV);
3579 
3580 	/* Allocating blocks may take a long time, so iterate. */
3581 	for (;;) {
3582 		olen = len;
3583 		ooffset = offset;
3584 
3585 		bwillwrite();
3586 		mp = NULL;
3587 		error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH);
3588 		if (error != 0)
3589 			break;
3590 		error = vn_lock(vp, LK_EXCLUSIVE);
3591 		if (error != 0) {
3592 			vn_finished_write(mp);
3593 			break;
3594 		}
3595 #ifdef AUDIT
3596 		if (!audited_vnode1) {
3597 			AUDIT_ARG_VNODE1(vp);
3598 			audited_vnode1 = 1;
3599 		}
3600 #endif
3601 #ifdef MAC
3602 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3603 		if (error == 0)
3604 #endif
3605 			error = VOP_ALLOCATE(vp, &offset, &len, 0,
3606 			    td->td_ucred);
3607 		VOP_UNLOCK(vp);
3608 		vn_finished_write(mp);
3609 
3610 		if (olen + ooffset != offset + len) {
3611 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3612 			    ooffset, olen, offset, len);
3613 		}
3614 		if (error != 0 || len == 0)
3615 			break;
3616 		KASSERT(olen > len, ("Iteration did not make progress?"));
3617 		maybe_yield();
3618 	}
3619 
3620 	return (error);
3621 }
3622 
3623 static int
3624 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags,
3625     int ioflag, struct ucred *cred, struct ucred *active_cred,
3626     struct ucred *file_cred)
3627 {
3628 	struct mount *mp;
3629 	void *rl_cookie;
3630 	off_t off, len;
3631 	int error;
3632 #ifdef AUDIT
3633 	bool audited_vnode1 = false;
3634 #endif
3635 
3636 	rl_cookie = NULL;
3637 	error = 0;
3638 	mp = NULL;
3639 	off = *offset;
3640 	len = *length;
3641 
3642 	if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0)
3643 		rl_cookie = vn_rangelock_wlock(vp, off, off + len);
3644 	while (len > 0 && error == 0) {
3645 		/*
3646 		 * Try to deallocate the longest range in one pass.
3647 		 * In case a pass takes too long to be executed, it returns
3648 		 * partial result. The residue will be proceeded in the next
3649 		 * pass.
3650 		 */
3651 
3652 		if ((ioflag & IO_NODELOCKED) == 0) {
3653 			bwillwrite();
3654 			if ((error = vn_start_write(vp, &mp,
3655 			    V_WAIT | V_PCATCH)) != 0)
3656 				goto out;
3657 			vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY);
3658 		}
3659 #ifdef AUDIT
3660 		if (!audited_vnode1) {
3661 			AUDIT_ARG_VNODE1(vp);
3662 			audited_vnode1 = true;
3663 		}
3664 #endif
3665 
3666 #ifdef MAC
3667 		if ((ioflag & IO_NOMACCHECK) == 0)
3668 			error = mac_vnode_check_write(active_cred, file_cred,
3669 			    vp);
3670 #endif
3671 		if (error == 0)
3672 			error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag,
3673 			    cred);
3674 
3675 		if ((ioflag & IO_NODELOCKED) == 0) {
3676 			VOP_UNLOCK(vp);
3677 			if (mp != NULL) {
3678 				vn_finished_write(mp);
3679 				mp = NULL;
3680 			}
3681 		}
3682 		if (error == 0 && len != 0)
3683 			maybe_yield();
3684 	}
3685 out:
3686 	if (rl_cookie != NULL)
3687 		vn_rangelock_unlock(vp, rl_cookie);
3688 	*offset = off;
3689 	*length = len;
3690 	return (error);
3691 }
3692 
3693 /*
3694  * This function is supposed to be used in the situations where the deallocation
3695  * is not triggered by a user request.
3696  */
3697 int
3698 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags,
3699     int ioflag, struct ucred *active_cred, struct ucred *file_cred)
3700 {
3701 	struct ucred *cred;
3702 
3703 	if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset ||
3704 	    flags != 0)
3705 		return (EINVAL);
3706 	if (vp->v_type != VREG)
3707 		return (ENODEV);
3708 
3709 	cred = file_cred != NOCRED ? file_cred : active_cred;
3710 	return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred,
3711 	    active_cred, file_cred));
3712 }
3713 
3714 static int
3715 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
3716     struct ucred *active_cred, struct thread *td)
3717 {
3718 	int error;
3719 	struct vnode *vp;
3720 	int ioflag;
3721 
3722 	KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd"));
3723 	KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
3724 	    ("vn_fspacectl: non-zero flags"));
3725 	KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
3726 	    ("vn_fspacectl: offset/length overflow or underflow"));
3727 	vp = fp->f_vnode;
3728 
3729 	if (vp->v_type != VREG)
3730 		return (ENODEV);
3731 
3732 	ioflag = get_write_ioflag(fp);
3733 
3734 	switch (cmd) {
3735 	case SPACECTL_DEALLOC:
3736 		error = vn_deallocate_impl(vp, offset, length, flags, ioflag,
3737 		    active_cred, active_cred, fp->f_cred);
3738 		break;
3739 	default:
3740 		panic("vn_fspacectl: unknown cmd %d", cmd);
3741 	}
3742 
3743 	return (error);
3744 }
3745 
3746 /*
3747  * Keep this assert as long as sizeof(struct dirent) is used as the maximum
3748  * entry size.
3749  */
3750 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent),
3751     "'struct dirent' size must be a multiple of its alignment "
3752     "(see _GENERIC_DIRLEN())");
3753 
3754 /*
3755  * Returns successive directory entries through some caller's provided buffer.
3756  *
3757  * This function automatically refills the provided buffer with calls to
3758  * VOP_READDIR() (after MAC permission checks).
3759  *
3760  * 'td' is used for credentials and passed to uiomove().  'dirbuf' is the
3761  * caller's buffer to fill and 'dirbuflen' its allocated size.  'dirbuf' must
3762  * be properly aligned to access 'struct dirent' structures and 'dirbuflen'
3763  * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning
3764  * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always
3765  * be returned if this requirement is not verified).  '*dpp' points to the
3766  * current directory entry in the buffer and '*len' contains the remaining
3767  * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry).
3768  *
3769  * At first call (or when restarting the read), '*len' must have been set to 0,
3770  * '*off' to 0 (or any valid start offset) and '*eofflag' to 0.  There are no
3771  * more entries as soon as '*len' is 0 after a call that returned 0.  Calling
3772  * again this function after such a condition is considered an error and EINVAL
3773  * will be returned.  Other possible error codes are those of VOP_READDIR(),
3774  * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL
3775  * (bad call).  All errors are unrecoverable, i.e., the state ('*len', '*off'
3776  * and '*eofflag') must be re-initialized before a subsequent call.  On error
3777  * or at end of directory, '*dpp' is reset to NULL.
3778  *
3779  * '*len', '*off' and '*eofflag' are internal state the caller should not
3780  * tamper with except as explained above.  '*off' is the next directory offset
3781  * to read from to refill the buffer.  '*eofflag' is set to 0 or 1 by the last
3782  * internal call to VOP_READDIR() that returned without error, indicating
3783  * whether it reached the end of the directory, and to 2 by this function after
3784  * all entries have been read.
3785  */
3786 int
3787 vn_dir_next_dirent(struct vnode *vp, struct thread *td,
3788     char *dirbuf, size_t dirbuflen,
3789     struct dirent **dpp, size_t *len, off_t *off, int *eofflag)
3790 {
3791 	struct dirent *dp = NULL;
3792 	int reclen;
3793 	int error;
3794 	struct uio uio;
3795 	struct iovec iov;
3796 
3797 	ASSERT_VOP_LOCKED(vp, "vnode not locked");
3798 	VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory"));
3799 	MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen,
3800 	    "Address space overflow");
3801 
3802 	if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) {
3803 		/* Don't take any chances in this case */
3804 		error = EINVAL;
3805 		goto out;
3806 	}
3807 
3808 	if (*len != 0) {
3809 		dp = *dpp;
3810 
3811 		/*
3812 		 * The caller continued to call us after an error (we set dp to
3813 		 * NULL in a previous iteration).  Bail out right now.
3814 		 */
3815 		if (__predict_false(dp == NULL))
3816 			return (EINVAL);
3817 
3818 		MPASS(*len <= dirbuflen);
3819 		MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp &&
3820 		    (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen,
3821 		    "Filled range not inside buffer");
3822 
3823 		reclen = dp->d_reclen;
3824 		if (reclen >= *len) {
3825 			/* End of buffer reached */
3826 			*len = 0;
3827 		} else {
3828 			dp = (struct dirent *)((char *)dp + reclen);
3829 			*len -= reclen;
3830 		}
3831 	}
3832 
3833 	if (*len == 0) {
3834 		dp = NULL;
3835 
3836 		/* Have to refill. */
3837 		switch (*eofflag) {
3838 		case 0:
3839 			break;
3840 
3841 		case 1:
3842 			/* Nothing more to read. */
3843 			*eofflag = 2; /* Remember the caller reached EOF. */
3844 			goto success;
3845 
3846 		default:
3847 			/* The caller didn't test for EOF. */
3848 			error = EINVAL;
3849 			goto out;
3850 		}
3851 
3852 		iov.iov_base = dirbuf;
3853 		iov.iov_len = dirbuflen;
3854 
3855 		uio.uio_iov = &iov;
3856 		uio.uio_iovcnt = 1;
3857 		uio.uio_offset = *off;
3858 		uio.uio_resid = dirbuflen;
3859 		uio.uio_segflg = UIO_SYSSPACE;
3860 		uio.uio_rw = UIO_READ;
3861 		uio.uio_td = td;
3862 
3863 #ifdef MAC
3864 		error = mac_vnode_check_readdir(td->td_ucred, vp);
3865 		if (error == 0)
3866 #endif
3867 			error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag,
3868 			    NULL, NULL);
3869 		if (error != 0)
3870 			goto out;
3871 
3872 		*len = dirbuflen - uio.uio_resid;
3873 		*off = uio.uio_offset;
3874 
3875 		if (*len == 0) {
3876 			/* Sanity check on INVARIANTS. */
3877 			MPASS(*eofflag != 0);
3878 			*eofflag = 1;
3879 			goto success;
3880 		}
3881 
3882 		/*
3883 		 * Normalize the flag returned by VOP_READDIR(), since we use 2
3884 		 * as a sentinel value.
3885 		 */
3886 		if (*eofflag != 0)
3887 			*eofflag = 1;
3888 
3889 		dp = (struct dirent *)dirbuf;
3890 	}
3891 
3892 	if (__predict_false(*len < GENERIC_MINDIRSIZ ||
3893 	    dp->d_reclen < GENERIC_MINDIRSIZ)) {
3894 		error = EINTEGRITY;
3895 		dp = NULL;
3896 		goto out;
3897 	}
3898 
3899 success:
3900 	error = 0;
3901 out:
3902 	*dpp = dp;
3903 	return (error);
3904 }
3905 
3906 /*
3907  * Checks whether a directory is empty or not.
3908  *
3909  * If the directory is empty, returns 0, and if it is not, ENOTEMPTY.  Other
3910  * values are genuine errors preventing the check.
3911  */
3912 int
3913 vn_dir_check_empty(struct vnode *vp)
3914 {
3915 	struct thread *const td = curthread;
3916 	char *dirbuf;
3917 	size_t dirbuflen, len;
3918 	off_t off;
3919 	int eofflag, error;
3920 	struct dirent *dp;
3921 	struct vattr va;
3922 
3923 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
3924 	VNPASS(vp->v_type == VDIR, vp);
3925 
3926 	error = VOP_GETATTR(vp, &va, td->td_ucred);
3927 	if (error != 0)
3928 		return (error);
3929 
3930 	dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ);
3931 	if (dirbuflen < va.va_blocksize)
3932 		dirbuflen = va.va_blocksize;
3933 	dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK);
3934 
3935 	len = 0;
3936 	off = 0;
3937 	eofflag = 0;
3938 
3939 	for (;;) {
3940 		error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen,
3941 		    &dp, &len, &off, &eofflag);
3942 		if (error != 0)
3943 			goto end;
3944 
3945 		if (len == 0) {
3946 			/* EOF */
3947 			error = 0;
3948 			goto end;
3949 		}
3950 
3951 		/*
3952 		 * Skip whiteouts.  Unionfs operates on filesystems only and
3953 		 * not on hierarchies, so these whiteouts would be shadowed on
3954 		 * the system hierarchy but not for a union using the
3955 		 * filesystem of their directories as the upper layer.
3956 		 * Additionally, unionfs currently transparently exposes
3957 		 * union-specific metadata of its upper layer, meaning that
3958 		 * whiteouts can be seen through the union view in empty
3959 		 * directories.  Taking into account these whiteouts would then
3960 		 * prevent mounting another filesystem on such effectively
3961 		 * empty directories.
3962 		 */
3963 		if (dp->d_type == DT_WHT)
3964 			continue;
3965 
3966 		/*
3967 		 * Any file in the directory which is not '.' or '..' indicates
3968 		 * the directory is not empty.
3969 		 */
3970 		switch (dp->d_namlen) {
3971 		case 2:
3972 			if (dp->d_name[1] != '.') {
3973 				/* Can't be '..' (nor '.') */
3974 				error = ENOTEMPTY;
3975 				goto end;
3976 			}
3977 			/* FALLTHROUGH */
3978 		case 1:
3979 			if (dp->d_name[0] != '.') {
3980 				/* Can't be '..' nor '.' */
3981 				error = ENOTEMPTY;
3982 				goto end;
3983 			}
3984 			break;
3985 
3986 		default:
3987 			error = ENOTEMPTY;
3988 			goto end;
3989 		}
3990 	}
3991 
3992 end:
3993 	free(dirbuf, M_TEMP);
3994 	return (error);
3995 }
3996 
3997 
3998 static u_long vn_lock_pair_pause_cnt;
3999 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
4000     &vn_lock_pair_pause_cnt, 0,
4001     "Count of vn_lock_pair deadlocks");
4002 
4003 u_int vn_lock_pair_pause_max;
4004 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
4005     &vn_lock_pair_pause_max, 0,
4006     "Max ticks for vn_lock_pair deadlock avoidance sleep");
4007 
4008 static void
4009 vn_lock_pair_pause(const char *wmesg)
4010 {
4011 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
4012 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
4013 }
4014 
4015 /*
4016  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
4017  * vp1_locked indicates whether vp1 is locked; if not, vp1 must be
4018  * unlocked.  Same for vp2 and vp2_locked.  One of the vnodes can be
4019  * NULL.
4020  *
4021  * The function returns with both vnodes exclusively or shared locked,
4022  * according to corresponding lkflags, and guarantees that it does not
4023  * create lock order reversal with other threads during its execution.
4024  * Both vnodes could be unlocked temporary (and reclaimed).
4025  *
4026  * If requesting shared locking, locked vnode lock must not be recursed.
4027  *
4028  * Only one of LK_SHARED and LK_EXCLUSIVE must be specified.
4029  * LK_NODDLKTREAT can be optionally passed.
4030  */
4031 void
4032 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1,
4033     struct vnode *vp2, bool vp2_locked, int lkflags2)
4034 {
4035 	int error;
4036 
4037 	MPASS(((lkflags1 & LK_SHARED) != 0) ^ ((lkflags1 & LK_EXCLUSIVE) != 0));
4038 	MPASS((lkflags1 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
4039 	MPASS(((lkflags2 & LK_SHARED) != 0) ^ ((lkflags2 & LK_EXCLUSIVE) != 0));
4040 	MPASS((lkflags2 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0);
4041 
4042 	if (vp1 == NULL && vp2 == NULL)
4043 		return;
4044 
4045 	if (vp1 != NULL) {
4046 		if ((lkflags1 & LK_SHARED) != 0 &&
4047 		    (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
4048 			lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE;
4049 		if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) {
4050 			ASSERT_VOP_LOCKED(vp1, "vp1");
4051 			if ((lkflags1 & LK_EXCLUSIVE) != 0) {
4052 				VOP_UNLOCK(vp1);
4053 				ASSERT_VOP_UNLOCKED(vp1,
4054 				    "vp1 shared recursed");
4055 				vp1_locked = false;
4056 			}
4057 		} else if (!vp1_locked)
4058 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
4059 	} else {
4060 		vp1_locked = true;
4061 	}
4062 
4063 	if (vp2 != NULL) {
4064 		if ((lkflags2 & LK_SHARED) != 0 &&
4065 		    (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0)
4066 			lkflags2 = (lkflags2 & ~LK_SHARED) | LK_EXCLUSIVE;
4067 		if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) {
4068 			ASSERT_VOP_LOCKED(vp2, "vp2");
4069 			if ((lkflags2 & LK_EXCLUSIVE) != 0) {
4070 				VOP_UNLOCK(vp2);
4071 				ASSERT_VOP_UNLOCKED(vp2,
4072 				    "vp2 shared recursed");
4073 				vp2_locked = false;
4074 			}
4075 		} else if (!vp2_locked)
4076 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
4077 	} else {
4078 		vp2_locked = true;
4079 	}
4080 
4081 	if (!vp1_locked && !vp2_locked) {
4082 		vn_lock(vp1, lkflags1 | LK_RETRY);
4083 		vp1_locked = true;
4084 	}
4085 
4086 	while (!vp1_locked || !vp2_locked) {
4087 		if (vp1_locked && vp2 != NULL) {
4088 			if (vp1 != NULL) {
4089 				error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT,
4090 				    __FILE__, __LINE__);
4091 				if (error == 0)
4092 					break;
4093 				VOP_UNLOCK(vp1);
4094 				vp1_locked = false;
4095 				vn_lock_pair_pause("vlp1");
4096 			}
4097 			vn_lock(vp2, lkflags2 | LK_RETRY);
4098 			vp2_locked = true;
4099 		}
4100 		if (vp2_locked && vp1 != NULL) {
4101 			if (vp2 != NULL) {
4102 				error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT,
4103 				    __FILE__, __LINE__);
4104 				if (error == 0)
4105 					break;
4106 				VOP_UNLOCK(vp2);
4107 				vp2_locked = false;
4108 				vn_lock_pair_pause("vlp2");
4109 			}
4110 			vn_lock(vp1, lkflags1 | LK_RETRY);
4111 			vp1_locked = true;
4112 		}
4113 	}
4114 	if (vp1 != NULL) {
4115 		if (lkflags1 == LK_EXCLUSIVE)
4116 			ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
4117 		else
4118 			ASSERT_VOP_LOCKED(vp1, "vp1 ret");
4119 	}
4120 	if (vp2 != NULL) {
4121 		if (lkflags2 == LK_EXCLUSIVE)
4122 			ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
4123 		else
4124 			ASSERT_VOP_LOCKED(vp2, "vp2 ret");
4125 	}
4126 }
4127 
4128 int
4129 vn_lktype_write(struct mount *mp, struct vnode *vp)
4130 {
4131 	if (MNT_SHARED_WRITES(mp) ||
4132 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount)))
4133 		return (LK_SHARED);
4134 	return (LK_EXCLUSIVE);
4135 }
4136