xref: /freebsd/sys/kern/vfs_vnops.c (revision f55bd0e5798dd17251302f73904da76bdfd257b9)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/prng.h>
74 #include <sys/sx.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
78 #include <sys/conf.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
81 #include <sys/user.h>
82 
83 #include <security/audit/audit.h>
84 #include <security/mac/mac_framework.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_extern.h>
88 #include <vm/pmap.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pager.h>
93 
94 #ifdef HWPMC_HOOKS
95 #include <sys/pmckern.h>
96 #endif
97 
98 static fo_rdwr_t	vn_read;
99 static fo_rdwr_t	vn_write;
100 static fo_rdwr_t	vn_io_fault;
101 static fo_truncate_t	vn_truncate;
102 static fo_ioctl_t	vn_ioctl;
103 static fo_poll_t	vn_poll;
104 static fo_kqfilter_t	vn_kqfilter;
105 static fo_close_t	vn_closefile;
106 static fo_mmap_t	vn_mmap;
107 static fo_fallocate_t	vn_fallocate;
108 
109 struct 	fileops vnops = {
110 	.fo_read = vn_io_fault,
111 	.fo_write = vn_io_fault,
112 	.fo_truncate = vn_truncate,
113 	.fo_ioctl = vn_ioctl,
114 	.fo_poll = vn_poll,
115 	.fo_kqfilter = vn_kqfilter,
116 	.fo_stat = vn_statfile,
117 	.fo_close = vn_closefile,
118 	.fo_chmod = vn_chmod,
119 	.fo_chown = vn_chown,
120 	.fo_sendfile = vn_sendfile,
121 	.fo_seek = vn_seek,
122 	.fo_fill_kinfo = vn_fill_kinfo,
123 	.fo_mmap = vn_mmap,
124 	.fo_fallocate = vn_fallocate,
125 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
126 };
127 
128 const u_int io_hold_cnt = 16;
129 static int vn_io_fault_enable = 1;
130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
131     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
132 static int vn_io_fault_prefault = 0;
133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
134     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
135 static int vn_io_pgcache_read_enable = 1;
136 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
137     &vn_io_pgcache_read_enable, 0,
138     "Enable copying from page cache for reads, avoiding fs");
139 static u_long vn_io_faults_cnt;
140 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
141     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
142 
143 static int vfs_allow_read_dir = 0;
144 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
145     &vfs_allow_read_dir, 0,
146     "Enable read(2) of directory by root for filesystems that support it");
147 
148 /*
149  * Returns true if vn_io_fault mode of handling the i/o request should
150  * be used.
151  */
152 static bool
153 do_vn_io_fault(struct vnode *vp, struct uio *uio)
154 {
155 	struct mount *mp;
156 
157 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
158 	    (mp = vp->v_mount) != NULL &&
159 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
160 }
161 
162 /*
163  * Structure used to pass arguments to vn_io_fault1(), to do either
164  * file- or vnode-based I/O calls.
165  */
166 struct vn_io_fault_args {
167 	enum {
168 		VN_IO_FAULT_FOP,
169 		VN_IO_FAULT_VOP
170 	} kind;
171 	struct ucred *cred;
172 	int flags;
173 	union {
174 		struct fop_args_tag {
175 			struct file *fp;
176 			fo_rdwr_t *doio;
177 		} fop_args;
178 		struct vop_args_tag {
179 			struct vnode *vp;
180 		} vop_args;
181 	} args;
182 };
183 
184 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
185     struct vn_io_fault_args *args, struct thread *td);
186 
187 int
188 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
189 {
190 	struct thread *td = ndp->ni_cnd.cn_thread;
191 
192 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
193 }
194 
195 static uint64_t
196 open2nameif(int fmode, u_int vn_open_flags)
197 {
198 	uint64_t res;
199 
200 	res = ISOPEN | LOCKLEAF;
201 	if ((fmode & O_RESOLVE_BENEATH) != 0)
202 		res |= RBENEATH;
203 	if ((fmode & O_EMPTY_PATH) != 0)
204 		res |= EMPTYPATH;
205 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
206 		res |= AUDITVNODE1;
207 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
208 		res |= NOCAPCHECK;
209 	return (res);
210 }
211 
212 /*
213  * Common code for vnode open operations via a name lookup.
214  * Lookup the vnode and invoke VOP_CREATE if needed.
215  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
216  *
217  * Note that this does NOT free nameidata for the successful case,
218  * due to the NDINIT being done elsewhere.
219  */
220 int
221 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
222     struct ucred *cred, struct file *fp)
223 {
224 	struct vnode *vp;
225 	struct mount *mp;
226 	struct thread *td = ndp->ni_cnd.cn_thread;
227 	struct vattr vat;
228 	struct vattr *vap = &vat;
229 	int fmode, error;
230 	bool first_open;
231 
232 restart:
233 	first_open = false;
234 	fmode = *flagp;
235 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
236 	    O_EXCL | O_DIRECTORY) ||
237 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
238 		return (EINVAL);
239 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
240 		ndp->ni_cnd.cn_nameiop = CREATE;
241 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
242 		/*
243 		 * Set NOCACHE to avoid flushing the cache when
244 		 * rolling in many files at once.
245 		 *
246 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
247 		 * exist despite NOCACHE.
248 		 */
249 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
250 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
251 			ndp->ni_cnd.cn_flags |= FOLLOW;
252 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
253 			bwillwrite();
254 		if ((error = namei(ndp)) != 0)
255 			return (error);
256 		if (ndp->ni_vp == NULL) {
257 			VATTR_NULL(vap);
258 			vap->va_type = VREG;
259 			vap->va_mode = cmode;
260 			if (fmode & O_EXCL)
261 				vap->va_vaflags |= VA_EXCLUSIVE;
262 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
263 				NDFREE(ndp, NDF_ONLY_PNBUF);
264 				vput(ndp->ni_dvp);
265 				if ((error = vn_start_write(NULL, &mp,
266 				    V_XSLEEP | PCATCH)) != 0)
267 					return (error);
268 				NDREINIT(ndp);
269 				goto restart;
270 			}
271 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
272 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
273 #ifdef MAC
274 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
275 			    &ndp->ni_cnd, vap);
276 			if (error == 0)
277 #endif
278 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
279 				    &ndp->ni_cnd, vap);
280 			vp = ndp->ni_vp;
281 			if (error == 0 && (fmode & O_EXCL) != 0 &&
282 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
283 				VI_LOCK(vp);
284 				vp->v_iflag |= VI_FOPENING;
285 				VI_UNLOCK(vp);
286 				first_open = true;
287 			}
288 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
289 			    false);
290 			vn_finished_write(mp);
291 			if (error) {
292 				NDFREE(ndp, NDF_ONLY_PNBUF);
293 				if (error == ERELOOKUP) {
294 					NDREINIT(ndp);
295 					goto restart;
296 				}
297 				return (error);
298 			}
299 			fmode &= ~O_TRUNC;
300 		} else {
301 			if (ndp->ni_dvp == ndp->ni_vp)
302 				vrele(ndp->ni_dvp);
303 			else
304 				vput(ndp->ni_dvp);
305 			ndp->ni_dvp = NULL;
306 			vp = ndp->ni_vp;
307 			if (fmode & O_EXCL) {
308 				error = EEXIST;
309 				goto bad;
310 			}
311 			if (vp->v_type == VDIR) {
312 				error = EISDIR;
313 				goto bad;
314 			}
315 			fmode &= ~O_CREAT;
316 		}
317 	} else {
318 		ndp->ni_cnd.cn_nameiop = LOOKUP;
319 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
320 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
321 		    FOLLOW;
322 		if ((fmode & FWRITE) == 0)
323 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
324 		if ((error = namei(ndp)) != 0)
325 			return (error);
326 		vp = ndp->ni_vp;
327 	}
328 	error = vn_open_vnode(vp, fmode, cred, td, fp);
329 	if (first_open) {
330 		VI_LOCK(vp);
331 		vp->v_iflag &= ~VI_FOPENING;
332 		wakeup(vp);
333 		VI_UNLOCK(vp);
334 	}
335 	if (error)
336 		goto bad;
337 	*flagp = fmode;
338 	return (0);
339 bad:
340 	NDFREE(ndp, NDF_ONLY_PNBUF);
341 	vput(vp);
342 	*flagp = fmode;
343 	ndp->ni_vp = NULL;
344 	return (error);
345 }
346 
347 static int
348 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
349 {
350 	struct flock lf;
351 	int error, lock_flags, type;
352 
353 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
354 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
355 		return (0);
356 	KASSERT(fp != NULL, ("open with flock requires fp"));
357 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
358 		return (EOPNOTSUPP);
359 
360 	lock_flags = VOP_ISLOCKED(vp);
361 	VOP_UNLOCK(vp);
362 
363 	lf.l_whence = SEEK_SET;
364 	lf.l_start = 0;
365 	lf.l_len = 0;
366 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
367 	type = F_FLOCK;
368 	if ((fmode & FNONBLOCK) == 0)
369 		type |= F_WAIT;
370 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
371 		type |= F_FIRSTOPEN;
372 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
373 	if (error == 0)
374 		fp->f_flag |= FHASLOCK;
375 
376 	vn_lock(vp, lock_flags | LK_RETRY);
377 	return (error);
378 }
379 
380 /*
381  * Common code for vnode open operations once a vnode is located.
382  * Check permissions, and call the VOP_OPEN routine.
383  */
384 int
385 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
386     struct thread *td, struct file *fp)
387 {
388 	accmode_t accmode;
389 	int error;
390 
391 	if (vp->v_type == VLNK) {
392 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
393 			return (EMLINK);
394 	}
395 	if (vp->v_type == VSOCK)
396 		return (EOPNOTSUPP);
397 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
398 		return (ENOTDIR);
399 
400 	accmode = 0;
401 	if ((fmode & O_PATH) == 0) {
402 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
403 			if (vp->v_type == VDIR)
404 				return (EISDIR);
405 			accmode |= VWRITE;
406 		}
407 		if ((fmode & FREAD) != 0)
408 			accmode |= VREAD;
409 		if ((fmode & O_APPEND) && (fmode & FWRITE))
410 			accmode |= VAPPEND;
411 #ifdef MAC
412 		if ((fmode & O_CREAT) != 0)
413 			accmode |= VCREAT;
414 #endif
415 	}
416 	if ((fmode & FEXEC) != 0)
417 		accmode |= VEXEC;
418 #ifdef MAC
419 	if ((fmode & O_VERIFY) != 0)
420 		accmode |= VVERIFY;
421 	error = mac_vnode_check_open(cred, vp, accmode);
422 	if (error != 0)
423 		return (error);
424 
425 	accmode &= ~(VCREAT | VVERIFY);
426 #endif
427 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
428 		error = VOP_ACCESS(vp, accmode, cred, td);
429 		if (error != 0)
430 			return (error);
431 	}
432 	if ((fmode & O_PATH) != 0) {
433 		if (vp->v_type == VFIFO)
434 			error = EPIPE;
435 		else
436 			error = VOP_ACCESS(vp, VREAD, cred, td);
437 		if (error == 0)
438 			fp->f_flag |= FKQALLOWED;
439 		return (0);
440 	}
441 
442 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
443 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
444 	error = VOP_OPEN(vp, fmode, cred, td, fp);
445 	if (error != 0)
446 		return (error);
447 
448 	error = vn_open_vnode_advlock(vp, fmode, fp);
449 	if (error == 0 && (fmode & FWRITE) != 0) {
450 		error = VOP_ADD_WRITECOUNT(vp, 1);
451 		if (error == 0) {
452 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
453 			     __func__, vp, vp->v_writecount);
454 		}
455 	}
456 
457 	/*
458 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
459 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
460 	 */
461 	if (error != 0) {
462 		if (fp != NULL) {
463 			/*
464 			 * Arrange the call by having fdrop() to use
465 			 * vn_closefile().  This is to satisfy
466 			 * filesystems like devfs or tmpfs, which
467 			 * override fo_close().
468 			 */
469 			fp->f_flag |= FOPENFAILED;
470 			fp->f_vnode = vp;
471 			if (fp->f_ops == &badfileops) {
472 				fp->f_type = DTYPE_VNODE;
473 				fp->f_ops = &vnops;
474 			}
475 			vref(vp);
476 		} else {
477 			/*
478 			 * If there is no fp, due to kernel-mode open,
479 			 * we can call VOP_CLOSE() now.
480 			 */
481 			if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
482 			    !MNT_EXTENDED_SHARED(vp->v_mount) &&
483 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
484 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
485 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
486 			    cred, td);
487 		}
488 	}
489 
490 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
491 	return (error);
492 
493 }
494 
495 /*
496  * Check for write permissions on the specified vnode.
497  * Prototype text segments cannot be written.
498  * It is racy.
499  */
500 int
501 vn_writechk(struct vnode *vp)
502 {
503 
504 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
505 	/*
506 	 * If there's shared text associated with
507 	 * the vnode, try to free it up once.  If
508 	 * we fail, we can't allow writing.
509 	 */
510 	if (VOP_IS_TEXT(vp))
511 		return (ETXTBSY);
512 
513 	return (0);
514 }
515 
516 /*
517  * Vnode close call
518  */
519 static int
520 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
521     struct thread *td, bool keep_ref)
522 {
523 	struct mount *mp;
524 	int error, lock_flags;
525 
526 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
527 	    MNT_EXTENDED_SHARED(vp->v_mount))
528 		lock_flags = LK_SHARED;
529 	else
530 		lock_flags = LK_EXCLUSIVE;
531 
532 	vn_start_write(vp, &mp, V_WAIT);
533 	vn_lock(vp, lock_flags | LK_RETRY);
534 	AUDIT_ARG_VNODE1(vp);
535 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
536 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
537 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
538 		    __func__, vp, vp->v_writecount);
539 	}
540 	error = VOP_CLOSE(vp, flags, file_cred, td);
541 	if (keep_ref)
542 		VOP_UNLOCK(vp);
543 	else
544 		vput(vp);
545 	vn_finished_write(mp);
546 	return (error);
547 }
548 
549 int
550 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
551     struct thread *td)
552 {
553 
554 	return (vn_close1(vp, flags, file_cred, td, false));
555 }
556 
557 /*
558  * Heuristic to detect sequential operation.
559  */
560 static int
561 sequential_heuristic(struct uio *uio, struct file *fp)
562 {
563 	enum uio_rw rw;
564 
565 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
566 
567 	rw = uio->uio_rw;
568 	if (fp->f_flag & FRDAHEAD)
569 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
570 
571 	/*
572 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
573 	 * that the first I/O is normally considered to be slightly
574 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
575 	 * unless previous seeks have reduced f_seqcount to 0, in which
576 	 * case offset 0 is not special.
577 	 */
578 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
579 	    uio->uio_offset == fp->f_nextoff[rw]) {
580 		/*
581 		 * f_seqcount is in units of fixed-size blocks so that it
582 		 * depends mainly on the amount of sequential I/O and not
583 		 * much on the number of sequential I/O's.  The fixed size
584 		 * of 16384 is hard-coded here since it is (not quite) just
585 		 * a magic size that works well here.  This size is more
586 		 * closely related to the best I/O size for real disks than
587 		 * to any block size used by software.
588 		 */
589 		if (uio->uio_resid >= IO_SEQMAX * 16384)
590 			fp->f_seqcount[rw] = IO_SEQMAX;
591 		else {
592 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
593 			if (fp->f_seqcount[rw] > IO_SEQMAX)
594 				fp->f_seqcount[rw] = IO_SEQMAX;
595 		}
596 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
597 	}
598 
599 	/* Not sequential.  Quickly draw-down sequentiality. */
600 	if (fp->f_seqcount[rw] > 1)
601 		fp->f_seqcount[rw] = 1;
602 	else
603 		fp->f_seqcount[rw] = 0;
604 	return (0);
605 }
606 
607 /*
608  * Package up an I/O request on a vnode into a uio and do it.
609  */
610 int
611 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
612     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
613     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
614 {
615 	struct uio auio;
616 	struct iovec aiov;
617 	struct mount *mp;
618 	struct ucred *cred;
619 	void *rl_cookie;
620 	struct vn_io_fault_args args;
621 	int error, lock_flags;
622 
623 	if (offset < 0 && vp->v_type != VCHR)
624 		return (EINVAL);
625 	auio.uio_iov = &aiov;
626 	auio.uio_iovcnt = 1;
627 	aiov.iov_base = base;
628 	aiov.iov_len = len;
629 	auio.uio_resid = len;
630 	auio.uio_offset = offset;
631 	auio.uio_segflg = segflg;
632 	auio.uio_rw = rw;
633 	auio.uio_td = td;
634 	error = 0;
635 
636 	if ((ioflg & IO_NODELOCKED) == 0) {
637 		if ((ioflg & IO_RANGELOCKED) == 0) {
638 			if (rw == UIO_READ) {
639 				rl_cookie = vn_rangelock_rlock(vp, offset,
640 				    offset + len);
641 			} else if ((ioflg & IO_APPEND) != 0) {
642 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
643 			} else {
644 				rl_cookie = vn_rangelock_wlock(vp, offset,
645 				    offset + len);
646 			}
647 		} else
648 			rl_cookie = NULL;
649 		mp = NULL;
650 		if (rw == UIO_WRITE) {
651 			if (vp->v_type != VCHR &&
652 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
653 			    != 0)
654 				goto out;
655 			if (MNT_SHARED_WRITES(mp) ||
656 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
657 				lock_flags = LK_SHARED;
658 			else
659 				lock_flags = LK_EXCLUSIVE;
660 		} else
661 			lock_flags = LK_SHARED;
662 		vn_lock(vp, lock_flags | LK_RETRY);
663 	} else
664 		rl_cookie = NULL;
665 
666 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
667 #ifdef MAC
668 	if ((ioflg & IO_NOMACCHECK) == 0) {
669 		if (rw == UIO_READ)
670 			error = mac_vnode_check_read(active_cred, file_cred,
671 			    vp);
672 		else
673 			error = mac_vnode_check_write(active_cred, file_cred,
674 			    vp);
675 	}
676 #endif
677 	if (error == 0) {
678 		if (file_cred != NULL)
679 			cred = file_cred;
680 		else
681 			cred = active_cred;
682 		if (do_vn_io_fault(vp, &auio)) {
683 			args.kind = VN_IO_FAULT_VOP;
684 			args.cred = cred;
685 			args.flags = ioflg;
686 			args.args.vop_args.vp = vp;
687 			error = vn_io_fault1(vp, &auio, &args, td);
688 		} else if (rw == UIO_READ) {
689 			error = VOP_READ(vp, &auio, ioflg, cred);
690 		} else /* if (rw == UIO_WRITE) */ {
691 			error = VOP_WRITE(vp, &auio, ioflg, cred);
692 		}
693 	}
694 	if (aresid)
695 		*aresid = auio.uio_resid;
696 	else
697 		if (auio.uio_resid && error == 0)
698 			error = EIO;
699 	if ((ioflg & IO_NODELOCKED) == 0) {
700 		VOP_UNLOCK(vp);
701 		if (mp != NULL)
702 			vn_finished_write(mp);
703 	}
704  out:
705 	if (rl_cookie != NULL)
706 		vn_rangelock_unlock(vp, rl_cookie);
707 	return (error);
708 }
709 
710 /*
711  * Package up an I/O request on a vnode into a uio and do it.  The I/O
712  * request is split up into smaller chunks and we try to avoid saturating
713  * the buffer cache while potentially holding a vnode locked, so we
714  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
715  * to give other processes a chance to lock the vnode (either other processes
716  * core'ing the same binary, or unrelated processes scanning the directory).
717  */
718 int
719 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
720     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
721     struct ucred *file_cred, size_t *aresid, struct thread *td)
722 {
723 	int error = 0;
724 	ssize_t iaresid;
725 
726 	do {
727 		int chunk;
728 
729 		/*
730 		 * Force `offset' to a multiple of MAXBSIZE except possibly
731 		 * for the first chunk, so that filesystems only need to
732 		 * write full blocks except possibly for the first and last
733 		 * chunks.
734 		 */
735 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
736 
737 		if (chunk > len)
738 			chunk = len;
739 		if (rw != UIO_READ && vp->v_type == VREG)
740 			bwillwrite();
741 		iaresid = 0;
742 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
743 		    ioflg, active_cred, file_cred, &iaresid, td);
744 		len -= chunk;	/* aresid calc already includes length */
745 		if (error)
746 			break;
747 		offset += chunk;
748 		base = (char *)base + chunk;
749 		kern_yield(PRI_USER);
750 	} while (len);
751 	if (aresid)
752 		*aresid = len + iaresid;
753 	return (error);
754 }
755 
756 #if OFF_MAX <= LONG_MAX
757 off_t
758 foffset_lock(struct file *fp, int flags)
759 {
760 	volatile short *flagsp;
761 	off_t res;
762 	short state;
763 
764 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
765 
766 	if ((flags & FOF_NOLOCK) != 0)
767 		return (atomic_load_long(&fp->f_offset));
768 
769 	/*
770 	 * According to McKusick the vn lock was protecting f_offset here.
771 	 * It is now protected by the FOFFSET_LOCKED flag.
772 	 */
773 	flagsp = &fp->f_vnread_flags;
774 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
775 		return (atomic_load_long(&fp->f_offset));
776 
777 	sleepq_lock(&fp->f_vnread_flags);
778 	state = atomic_load_16(flagsp);
779 	for (;;) {
780 		if ((state & FOFFSET_LOCKED) == 0) {
781 			if (!atomic_fcmpset_acq_16(flagsp, &state,
782 			    FOFFSET_LOCKED))
783 				continue;
784 			break;
785 		}
786 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
787 			if (!atomic_fcmpset_acq_16(flagsp, &state,
788 			    state | FOFFSET_LOCK_WAITING))
789 				continue;
790 		}
791 		DROP_GIANT();
792 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
793 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
794 		PICKUP_GIANT();
795 		sleepq_lock(&fp->f_vnread_flags);
796 		state = atomic_load_16(flagsp);
797 	}
798 	res = atomic_load_long(&fp->f_offset);
799 	sleepq_release(&fp->f_vnread_flags);
800 	return (res);
801 }
802 
803 void
804 foffset_unlock(struct file *fp, off_t val, int flags)
805 {
806 	volatile short *flagsp;
807 	short state;
808 
809 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
810 
811 	if ((flags & FOF_NOUPDATE) == 0)
812 		atomic_store_long(&fp->f_offset, val);
813 	if ((flags & FOF_NEXTOFF_R) != 0)
814 		fp->f_nextoff[UIO_READ] = val;
815 	if ((flags & FOF_NEXTOFF_W) != 0)
816 		fp->f_nextoff[UIO_WRITE] = val;
817 
818 	if ((flags & FOF_NOLOCK) != 0)
819 		return;
820 
821 	flagsp = &fp->f_vnread_flags;
822 	state = atomic_load_16(flagsp);
823 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
824 	    atomic_cmpset_rel_16(flagsp, state, 0))
825 		return;
826 
827 	sleepq_lock(&fp->f_vnread_flags);
828 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
829 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
830 	fp->f_vnread_flags = 0;
831 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
832 	sleepq_release(&fp->f_vnread_flags);
833 }
834 #else
835 off_t
836 foffset_lock(struct file *fp, int flags)
837 {
838 	struct mtx *mtxp;
839 	off_t res;
840 
841 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
842 
843 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
844 	mtx_lock(mtxp);
845 	if ((flags & FOF_NOLOCK) == 0) {
846 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
847 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
848 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
849 			    "vofflock", 0);
850 		}
851 		fp->f_vnread_flags |= FOFFSET_LOCKED;
852 	}
853 	res = fp->f_offset;
854 	mtx_unlock(mtxp);
855 	return (res);
856 }
857 
858 void
859 foffset_unlock(struct file *fp, off_t val, int flags)
860 {
861 	struct mtx *mtxp;
862 
863 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
864 
865 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
866 	mtx_lock(mtxp);
867 	if ((flags & FOF_NOUPDATE) == 0)
868 		fp->f_offset = val;
869 	if ((flags & FOF_NEXTOFF_R) != 0)
870 		fp->f_nextoff[UIO_READ] = val;
871 	if ((flags & FOF_NEXTOFF_W) != 0)
872 		fp->f_nextoff[UIO_WRITE] = val;
873 	if ((flags & FOF_NOLOCK) == 0) {
874 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
875 		    ("Lost FOFFSET_LOCKED"));
876 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
877 			wakeup(&fp->f_vnread_flags);
878 		fp->f_vnread_flags = 0;
879 	}
880 	mtx_unlock(mtxp);
881 }
882 #endif
883 
884 void
885 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
886 {
887 
888 	if ((flags & FOF_OFFSET) == 0)
889 		uio->uio_offset = foffset_lock(fp, flags);
890 }
891 
892 void
893 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
894 {
895 
896 	if ((flags & FOF_OFFSET) == 0)
897 		foffset_unlock(fp, uio->uio_offset, flags);
898 }
899 
900 static int
901 get_advice(struct file *fp, struct uio *uio)
902 {
903 	struct mtx *mtxp;
904 	int ret;
905 
906 	ret = POSIX_FADV_NORMAL;
907 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
908 		return (ret);
909 
910 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
911 	mtx_lock(mtxp);
912 	if (fp->f_advice != NULL &&
913 	    uio->uio_offset >= fp->f_advice->fa_start &&
914 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
915 		ret = fp->f_advice->fa_advice;
916 	mtx_unlock(mtxp);
917 	return (ret);
918 }
919 
920 int
921 vn_read_from_obj(struct vnode *vp, struct uio *uio)
922 {
923 	vm_object_t obj;
924 	vm_page_t ma[io_hold_cnt + 2];
925 	off_t off, vsz;
926 	ssize_t resid;
927 	int error, i, j;
928 
929 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
930 	obj = atomic_load_ptr(&vp->v_object);
931 	if (obj == NULL)
932 		return (EJUSTRETURN);
933 
934 	/*
935 	 * Depends on type stability of vm_objects.
936 	 */
937 	vm_object_pip_add(obj, 1);
938 	if ((obj->flags & OBJ_DEAD) != 0) {
939 		/*
940 		 * Note that object might be already reused from the
941 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
942 		 * we recheck for DOOMED vnode state after all pages
943 		 * are busied, and retract then.
944 		 *
945 		 * But we check for OBJ_DEAD to ensure that we do not
946 		 * busy pages while vm_object_terminate_pages()
947 		 * processes the queue.
948 		 */
949 		error = EJUSTRETURN;
950 		goto out_pip;
951 	}
952 
953 	resid = uio->uio_resid;
954 	off = uio->uio_offset;
955 	for (i = 0; resid > 0; i++) {
956 		MPASS(i < io_hold_cnt + 2);
957 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
958 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
959 		    VM_ALLOC_NOWAIT);
960 		if (ma[i] == NULL)
961 			break;
962 
963 		/*
964 		 * Skip invalid pages.  Valid mask can be partial only
965 		 * at EOF, and we clip later.
966 		 */
967 		if (vm_page_none_valid(ma[i])) {
968 			vm_page_sunbusy(ma[i]);
969 			break;
970 		}
971 
972 		resid -= PAGE_SIZE;
973 		off += PAGE_SIZE;
974 	}
975 	if (i == 0) {
976 		error = EJUSTRETURN;
977 		goto out_pip;
978 	}
979 
980 	/*
981 	 * Check VIRF_DOOMED after we busied our pages.  Since
982 	 * vgonel() terminates the vnode' vm_object, it cannot
983 	 * process past pages busied by us.
984 	 */
985 	if (VN_IS_DOOMED(vp)) {
986 		error = EJUSTRETURN;
987 		goto out;
988 	}
989 
990 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
991 	if (resid > uio->uio_resid)
992 		resid = uio->uio_resid;
993 
994 	/*
995 	 * Unlocked read of vnp_size is safe because truncation cannot
996 	 * pass busied page.  But we load vnp_size into a local
997 	 * variable so that possible concurrent extension does not
998 	 * break calculation.
999 	 */
1000 #if defined(__powerpc__) && !defined(__powerpc64__)
1001 	vsz = obj->un_pager.vnp.vnp_size;
1002 #else
1003 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1004 #endif
1005 	if (uio->uio_offset >= vsz) {
1006 		error = EJUSTRETURN;
1007 		goto out;
1008 	}
1009 	if (uio->uio_offset + resid > vsz)
1010 		resid = vsz - uio->uio_offset;
1011 
1012 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1013 
1014 out:
1015 	for (j = 0; j < i; j++) {
1016 		if (error == 0)
1017 			vm_page_reference(ma[j]);
1018 		vm_page_sunbusy(ma[j]);
1019 	}
1020 out_pip:
1021 	vm_object_pip_wakeup(obj);
1022 	if (error != 0)
1023 		return (error);
1024 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1025 }
1026 
1027 /*
1028  * File table vnode read routine.
1029  */
1030 static int
1031 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1032     struct thread *td)
1033 {
1034 	struct vnode *vp;
1035 	off_t orig_offset;
1036 	int error, ioflag;
1037 	int advice;
1038 
1039 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1040 	    uio->uio_td, td));
1041 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1042 	vp = fp->f_vnode;
1043 	ioflag = 0;
1044 	if (fp->f_flag & FNONBLOCK)
1045 		ioflag |= IO_NDELAY;
1046 	if (fp->f_flag & O_DIRECT)
1047 		ioflag |= IO_DIRECT;
1048 
1049 	/*
1050 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1051 	 * allows us to avoid unneeded work outright.
1052 	 */
1053 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1054 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1055 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1056 		if (error == 0) {
1057 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1058 			return (0);
1059 		}
1060 		if (error != EJUSTRETURN)
1061 			return (error);
1062 	}
1063 
1064 	advice = get_advice(fp, uio);
1065 	vn_lock(vp, LK_SHARED | LK_RETRY);
1066 
1067 	switch (advice) {
1068 	case POSIX_FADV_NORMAL:
1069 	case POSIX_FADV_SEQUENTIAL:
1070 	case POSIX_FADV_NOREUSE:
1071 		ioflag |= sequential_heuristic(uio, fp);
1072 		break;
1073 	case POSIX_FADV_RANDOM:
1074 		/* Disable read-ahead for random I/O. */
1075 		break;
1076 	}
1077 	orig_offset = uio->uio_offset;
1078 
1079 #ifdef MAC
1080 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1081 	if (error == 0)
1082 #endif
1083 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1084 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1085 	VOP_UNLOCK(vp);
1086 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1087 	    orig_offset != uio->uio_offset)
1088 		/*
1089 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1090 		 * for the backing file after a POSIX_FADV_NOREUSE
1091 		 * read(2).
1092 		 */
1093 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1094 		    POSIX_FADV_DONTNEED);
1095 	return (error);
1096 }
1097 
1098 /*
1099  * File table vnode write routine.
1100  */
1101 static int
1102 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1103     struct thread *td)
1104 {
1105 	struct vnode *vp;
1106 	struct mount *mp;
1107 	off_t orig_offset;
1108 	int error, ioflag, lock_flags;
1109 	int advice;
1110 
1111 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1112 	    uio->uio_td, td));
1113 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1114 	vp = fp->f_vnode;
1115 	if (vp->v_type == VREG)
1116 		bwillwrite();
1117 	ioflag = IO_UNIT;
1118 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1119 		ioflag |= IO_APPEND;
1120 	if (fp->f_flag & FNONBLOCK)
1121 		ioflag |= IO_NDELAY;
1122 	if (fp->f_flag & O_DIRECT)
1123 		ioflag |= IO_DIRECT;
1124 	if ((fp->f_flag & O_FSYNC) ||
1125 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1126 		ioflag |= IO_SYNC;
1127 	/*
1128 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1129 	 * implementations that don't understand IO_DATASYNC fall back to full
1130 	 * O_SYNC behavior.
1131 	 */
1132 	if (fp->f_flag & O_DSYNC)
1133 		ioflag |= IO_SYNC | IO_DATASYNC;
1134 	mp = NULL;
1135 	if (vp->v_type != VCHR &&
1136 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1137 		goto unlock;
1138 
1139 	advice = get_advice(fp, uio);
1140 
1141 	if (MNT_SHARED_WRITES(mp) ||
1142 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1143 		lock_flags = LK_SHARED;
1144 	} else {
1145 		lock_flags = LK_EXCLUSIVE;
1146 	}
1147 
1148 	vn_lock(vp, lock_flags | LK_RETRY);
1149 	switch (advice) {
1150 	case POSIX_FADV_NORMAL:
1151 	case POSIX_FADV_SEQUENTIAL:
1152 	case POSIX_FADV_NOREUSE:
1153 		ioflag |= sequential_heuristic(uio, fp);
1154 		break;
1155 	case POSIX_FADV_RANDOM:
1156 		/* XXX: Is this correct? */
1157 		break;
1158 	}
1159 	orig_offset = uio->uio_offset;
1160 
1161 #ifdef MAC
1162 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1163 	if (error == 0)
1164 #endif
1165 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1166 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1167 	VOP_UNLOCK(vp);
1168 	if (vp->v_type != VCHR)
1169 		vn_finished_write(mp);
1170 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1171 	    orig_offset != uio->uio_offset)
1172 		/*
1173 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1174 		 * for the backing file after a POSIX_FADV_NOREUSE
1175 		 * write(2).
1176 		 */
1177 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1178 		    POSIX_FADV_DONTNEED);
1179 unlock:
1180 	return (error);
1181 }
1182 
1183 /*
1184  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1185  * prevent the following deadlock:
1186  *
1187  * Assume that the thread A reads from the vnode vp1 into userspace
1188  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1189  * currently not resident, then system ends up with the call chain
1190  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1191  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1192  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1193  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1194  * backed by the pages of vnode vp1, and some page in buf2 is not
1195  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1196  *
1197  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1198  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1199  * Instead, it first tries to do the whole range i/o with pagefaults
1200  * disabled. If all pages in the i/o buffer are resident and mapped,
1201  * VOP will succeed (ignoring the genuine filesystem errors).
1202  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1203  * i/o in chunks, with all pages in the chunk prefaulted and held
1204  * using vm_fault_quick_hold_pages().
1205  *
1206  * Filesystems using this deadlock avoidance scheme should use the
1207  * array of the held pages from uio, saved in the curthread->td_ma,
1208  * instead of doing uiomove().  A helper function
1209  * vn_io_fault_uiomove() converts uiomove request into
1210  * uiomove_fromphys() over td_ma array.
1211  *
1212  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1213  * make the current i/o request atomic with respect to other i/os and
1214  * truncations.
1215  */
1216 
1217 /*
1218  * Decode vn_io_fault_args and perform the corresponding i/o.
1219  */
1220 static int
1221 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1222     struct thread *td)
1223 {
1224 	int error, save;
1225 
1226 	error = 0;
1227 	save = vm_fault_disable_pagefaults();
1228 	switch (args->kind) {
1229 	case VN_IO_FAULT_FOP:
1230 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1231 		    uio, args->cred, args->flags, td);
1232 		break;
1233 	case VN_IO_FAULT_VOP:
1234 		if (uio->uio_rw == UIO_READ) {
1235 			error = VOP_READ(args->args.vop_args.vp, uio,
1236 			    args->flags, args->cred);
1237 		} else if (uio->uio_rw == UIO_WRITE) {
1238 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1239 			    args->flags, args->cred);
1240 		}
1241 		break;
1242 	default:
1243 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1244 		    args->kind, uio->uio_rw);
1245 	}
1246 	vm_fault_enable_pagefaults(save);
1247 	return (error);
1248 }
1249 
1250 static int
1251 vn_io_fault_touch(char *base, const struct uio *uio)
1252 {
1253 	int r;
1254 
1255 	r = fubyte(base);
1256 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1257 		return (EFAULT);
1258 	return (0);
1259 }
1260 
1261 static int
1262 vn_io_fault_prefault_user(const struct uio *uio)
1263 {
1264 	char *base;
1265 	const struct iovec *iov;
1266 	size_t len;
1267 	ssize_t resid;
1268 	int error, i;
1269 
1270 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1271 	    ("vn_io_fault_prefault userspace"));
1272 
1273 	error = i = 0;
1274 	iov = uio->uio_iov;
1275 	resid = uio->uio_resid;
1276 	base = iov->iov_base;
1277 	len = iov->iov_len;
1278 	while (resid > 0) {
1279 		error = vn_io_fault_touch(base, uio);
1280 		if (error != 0)
1281 			break;
1282 		if (len < PAGE_SIZE) {
1283 			if (len != 0) {
1284 				error = vn_io_fault_touch(base + len - 1, uio);
1285 				if (error != 0)
1286 					break;
1287 				resid -= len;
1288 			}
1289 			if (++i >= uio->uio_iovcnt)
1290 				break;
1291 			iov = uio->uio_iov + i;
1292 			base = iov->iov_base;
1293 			len = iov->iov_len;
1294 		} else {
1295 			len -= PAGE_SIZE;
1296 			base += PAGE_SIZE;
1297 			resid -= PAGE_SIZE;
1298 		}
1299 	}
1300 	return (error);
1301 }
1302 
1303 /*
1304  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1305  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1306  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1307  * into args and call vn_io_fault1() to handle faults during the user
1308  * mode buffer accesses.
1309  */
1310 static int
1311 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1312     struct thread *td)
1313 {
1314 	vm_page_t ma[io_hold_cnt + 2];
1315 	struct uio *uio_clone, short_uio;
1316 	struct iovec short_iovec[1];
1317 	vm_page_t *prev_td_ma;
1318 	vm_prot_t prot;
1319 	vm_offset_t addr, end;
1320 	size_t len, resid;
1321 	ssize_t adv;
1322 	int error, cnt, saveheld, prev_td_ma_cnt;
1323 
1324 	if (vn_io_fault_prefault) {
1325 		error = vn_io_fault_prefault_user(uio);
1326 		if (error != 0)
1327 			return (error); /* Or ignore ? */
1328 	}
1329 
1330 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1331 
1332 	/*
1333 	 * The UFS follows IO_UNIT directive and replays back both
1334 	 * uio_offset and uio_resid if an error is encountered during the
1335 	 * operation.  But, since the iovec may be already advanced,
1336 	 * uio is still in an inconsistent state.
1337 	 *
1338 	 * Cache a copy of the original uio, which is advanced to the redo
1339 	 * point using UIO_NOCOPY below.
1340 	 */
1341 	uio_clone = cloneuio(uio);
1342 	resid = uio->uio_resid;
1343 
1344 	short_uio.uio_segflg = UIO_USERSPACE;
1345 	short_uio.uio_rw = uio->uio_rw;
1346 	short_uio.uio_td = uio->uio_td;
1347 
1348 	error = vn_io_fault_doio(args, uio, td);
1349 	if (error != EFAULT)
1350 		goto out;
1351 
1352 	atomic_add_long(&vn_io_faults_cnt, 1);
1353 	uio_clone->uio_segflg = UIO_NOCOPY;
1354 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1355 	uio_clone->uio_segflg = uio->uio_segflg;
1356 
1357 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1358 	prev_td_ma = td->td_ma;
1359 	prev_td_ma_cnt = td->td_ma_cnt;
1360 
1361 	while (uio_clone->uio_resid != 0) {
1362 		len = uio_clone->uio_iov->iov_len;
1363 		if (len == 0) {
1364 			KASSERT(uio_clone->uio_iovcnt >= 1,
1365 			    ("iovcnt underflow"));
1366 			uio_clone->uio_iov++;
1367 			uio_clone->uio_iovcnt--;
1368 			continue;
1369 		}
1370 		if (len > ptoa(io_hold_cnt))
1371 			len = ptoa(io_hold_cnt);
1372 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1373 		end = round_page(addr + len);
1374 		if (end < addr) {
1375 			error = EFAULT;
1376 			break;
1377 		}
1378 		cnt = atop(end - trunc_page(addr));
1379 		/*
1380 		 * A perfectly misaligned address and length could cause
1381 		 * both the start and the end of the chunk to use partial
1382 		 * page.  +2 accounts for such a situation.
1383 		 */
1384 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1385 		    addr, len, prot, ma, io_hold_cnt + 2);
1386 		if (cnt == -1) {
1387 			error = EFAULT;
1388 			break;
1389 		}
1390 		short_uio.uio_iov = &short_iovec[0];
1391 		short_iovec[0].iov_base = (void *)addr;
1392 		short_uio.uio_iovcnt = 1;
1393 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1394 		short_uio.uio_offset = uio_clone->uio_offset;
1395 		td->td_ma = ma;
1396 		td->td_ma_cnt = cnt;
1397 
1398 		error = vn_io_fault_doio(args, &short_uio, td);
1399 		vm_page_unhold_pages(ma, cnt);
1400 		adv = len - short_uio.uio_resid;
1401 
1402 		uio_clone->uio_iov->iov_base =
1403 		    (char *)uio_clone->uio_iov->iov_base + adv;
1404 		uio_clone->uio_iov->iov_len -= adv;
1405 		uio_clone->uio_resid -= adv;
1406 		uio_clone->uio_offset += adv;
1407 
1408 		uio->uio_resid -= adv;
1409 		uio->uio_offset += adv;
1410 
1411 		if (error != 0 || adv == 0)
1412 			break;
1413 	}
1414 	td->td_ma = prev_td_ma;
1415 	td->td_ma_cnt = prev_td_ma_cnt;
1416 	curthread_pflags_restore(saveheld);
1417 out:
1418 	free(uio_clone, M_IOV);
1419 	return (error);
1420 }
1421 
1422 static int
1423 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1424     int flags, struct thread *td)
1425 {
1426 	fo_rdwr_t *doio;
1427 	struct vnode *vp;
1428 	void *rl_cookie;
1429 	struct vn_io_fault_args args;
1430 	int error;
1431 
1432 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1433 	vp = fp->f_vnode;
1434 
1435 	/*
1436 	 * The ability to read(2) on a directory has historically been
1437 	 * allowed for all users, but this can and has been the source of
1438 	 * at least one security issue in the past.  As such, it is now hidden
1439 	 * away behind a sysctl for those that actually need it to use it, and
1440 	 * restricted to root when it's turned on to make it relatively safe to
1441 	 * leave on for longer sessions of need.
1442 	 */
1443 	if (vp->v_type == VDIR) {
1444 		KASSERT(uio->uio_rw == UIO_READ,
1445 		    ("illegal write attempted on a directory"));
1446 		if (!vfs_allow_read_dir)
1447 			return (EISDIR);
1448 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1449 			return (EISDIR);
1450 	}
1451 
1452 	foffset_lock_uio(fp, uio, flags);
1453 	if (do_vn_io_fault(vp, uio)) {
1454 		args.kind = VN_IO_FAULT_FOP;
1455 		args.args.fop_args.fp = fp;
1456 		args.args.fop_args.doio = doio;
1457 		args.cred = active_cred;
1458 		args.flags = flags | FOF_OFFSET;
1459 		if (uio->uio_rw == UIO_READ) {
1460 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1461 			    uio->uio_offset + uio->uio_resid);
1462 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1463 		    (flags & FOF_OFFSET) == 0) {
1464 			/* For appenders, punt and lock the whole range. */
1465 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1466 		} else {
1467 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1468 			    uio->uio_offset + uio->uio_resid);
1469 		}
1470 		error = vn_io_fault1(vp, uio, &args, td);
1471 		vn_rangelock_unlock(vp, rl_cookie);
1472 	} else {
1473 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1474 	}
1475 	foffset_unlock_uio(fp, uio, flags);
1476 	return (error);
1477 }
1478 
1479 /*
1480  * Helper function to perform the requested uiomove operation using
1481  * the held pages for io->uio_iov[0].iov_base buffer instead of
1482  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1483  * instead of iov_base prevents page faults that could occur due to
1484  * pmap_collect() invalidating the mapping created by
1485  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1486  * object cleanup revoking the write access from page mappings.
1487  *
1488  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1489  * instead of plain uiomove().
1490  */
1491 int
1492 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1493 {
1494 	struct uio transp_uio;
1495 	struct iovec transp_iov[1];
1496 	struct thread *td;
1497 	size_t adv;
1498 	int error, pgadv;
1499 
1500 	td = curthread;
1501 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1502 	    uio->uio_segflg != UIO_USERSPACE)
1503 		return (uiomove(data, xfersize, uio));
1504 
1505 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1506 	transp_iov[0].iov_base = data;
1507 	transp_uio.uio_iov = &transp_iov[0];
1508 	transp_uio.uio_iovcnt = 1;
1509 	if (xfersize > uio->uio_resid)
1510 		xfersize = uio->uio_resid;
1511 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1512 	transp_uio.uio_offset = 0;
1513 	transp_uio.uio_segflg = UIO_SYSSPACE;
1514 	/*
1515 	 * Since transp_iov points to data, and td_ma page array
1516 	 * corresponds to original uio->uio_iov, we need to invert the
1517 	 * direction of the i/o operation as passed to
1518 	 * uiomove_fromphys().
1519 	 */
1520 	switch (uio->uio_rw) {
1521 	case UIO_WRITE:
1522 		transp_uio.uio_rw = UIO_READ;
1523 		break;
1524 	case UIO_READ:
1525 		transp_uio.uio_rw = UIO_WRITE;
1526 		break;
1527 	}
1528 	transp_uio.uio_td = uio->uio_td;
1529 	error = uiomove_fromphys(td->td_ma,
1530 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1531 	    xfersize, &transp_uio);
1532 	adv = xfersize - transp_uio.uio_resid;
1533 	pgadv =
1534 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1535 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1536 	td->td_ma += pgadv;
1537 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1538 	    pgadv));
1539 	td->td_ma_cnt -= pgadv;
1540 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1541 	uio->uio_iov->iov_len -= adv;
1542 	uio->uio_resid -= adv;
1543 	uio->uio_offset += adv;
1544 	return (error);
1545 }
1546 
1547 int
1548 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1549     struct uio *uio)
1550 {
1551 	struct thread *td;
1552 	vm_offset_t iov_base;
1553 	int cnt, pgadv;
1554 
1555 	td = curthread;
1556 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1557 	    uio->uio_segflg != UIO_USERSPACE)
1558 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1559 
1560 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1561 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1562 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1563 	switch (uio->uio_rw) {
1564 	case UIO_WRITE:
1565 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1566 		    offset, cnt);
1567 		break;
1568 	case UIO_READ:
1569 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1570 		    cnt);
1571 		break;
1572 	}
1573 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1574 	td->td_ma += pgadv;
1575 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1576 	    pgadv));
1577 	td->td_ma_cnt -= pgadv;
1578 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1579 	uio->uio_iov->iov_len -= cnt;
1580 	uio->uio_resid -= cnt;
1581 	uio->uio_offset += cnt;
1582 	return (0);
1583 }
1584 
1585 /*
1586  * File table truncate routine.
1587  */
1588 static int
1589 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1590     struct thread *td)
1591 {
1592 	struct mount *mp;
1593 	struct vnode *vp;
1594 	void *rl_cookie;
1595 	int error;
1596 
1597 	vp = fp->f_vnode;
1598 
1599 retry:
1600 	/*
1601 	 * Lock the whole range for truncation.  Otherwise split i/o
1602 	 * might happen partly before and partly after the truncation.
1603 	 */
1604 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1605 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1606 	if (error)
1607 		goto out1;
1608 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1609 	AUDIT_ARG_VNODE1(vp);
1610 	if (vp->v_type == VDIR) {
1611 		error = EISDIR;
1612 		goto out;
1613 	}
1614 #ifdef MAC
1615 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1616 	if (error)
1617 		goto out;
1618 #endif
1619 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1620 	    fp->f_cred);
1621 out:
1622 	VOP_UNLOCK(vp);
1623 	vn_finished_write(mp);
1624 out1:
1625 	vn_rangelock_unlock(vp, rl_cookie);
1626 	if (error == ERELOOKUP)
1627 		goto retry;
1628 	return (error);
1629 }
1630 
1631 /*
1632  * Truncate a file that is already locked.
1633  */
1634 int
1635 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1636     struct ucred *cred)
1637 {
1638 	struct vattr vattr;
1639 	int error;
1640 
1641 	error = VOP_ADD_WRITECOUNT(vp, 1);
1642 	if (error == 0) {
1643 		VATTR_NULL(&vattr);
1644 		vattr.va_size = length;
1645 		if (sync)
1646 			vattr.va_vaflags |= VA_SYNC;
1647 		error = VOP_SETATTR(vp, &vattr, cred);
1648 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1649 	}
1650 	return (error);
1651 }
1652 
1653 /*
1654  * File table vnode stat routine.
1655  */
1656 int
1657 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1658     struct thread *td)
1659 {
1660 	struct vnode *vp = fp->f_vnode;
1661 	int error;
1662 
1663 	vn_lock(vp, LK_SHARED | LK_RETRY);
1664 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1665 	VOP_UNLOCK(vp);
1666 
1667 	return (error);
1668 }
1669 
1670 /*
1671  * File table vnode ioctl routine.
1672  */
1673 static int
1674 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1675     struct thread *td)
1676 {
1677 	struct vattr vattr;
1678 	struct vnode *vp;
1679 	struct fiobmap2_arg *bmarg;
1680 	int error;
1681 
1682 	vp = fp->f_vnode;
1683 	switch (vp->v_type) {
1684 	case VDIR:
1685 	case VREG:
1686 		switch (com) {
1687 		case FIONREAD:
1688 			vn_lock(vp, LK_SHARED | LK_RETRY);
1689 			error = VOP_GETATTR(vp, &vattr, active_cred);
1690 			VOP_UNLOCK(vp);
1691 			if (error == 0)
1692 				*(int *)data = vattr.va_size - fp->f_offset;
1693 			return (error);
1694 		case FIOBMAP2:
1695 			bmarg = (struct fiobmap2_arg *)data;
1696 			vn_lock(vp, LK_SHARED | LK_RETRY);
1697 #ifdef MAC
1698 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1699 			    vp);
1700 			if (error == 0)
1701 #endif
1702 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1703 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1704 			VOP_UNLOCK(vp);
1705 			return (error);
1706 		case FIONBIO:
1707 		case FIOASYNC:
1708 			return (0);
1709 		default:
1710 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1711 			    active_cred, td));
1712 		}
1713 		break;
1714 	case VCHR:
1715 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1716 		    active_cred, td));
1717 	default:
1718 		return (ENOTTY);
1719 	}
1720 }
1721 
1722 /*
1723  * File table vnode poll routine.
1724  */
1725 static int
1726 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1727     struct thread *td)
1728 {
1729 	struct vnode *vp;
1730 	int error;
1731 
1732 	vp = fp->f_vnode;
1733 #if defined(MAC) || defined(AUDIT)
1734 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1735 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1736 		AUDIT_ARG_VNODE1(vp);
1737 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1738 		VOP_UNLOCK(vp);
1739 		if (error != 0)
1740 			return (error);
1741 	}
1742 #endif
1743 	error = VOP_POLL(vp, events, fp->f_cred, td);
1744 	return (error);
1745 }
1746 
1747 /*
1748  * Acquire the requested lock and then check for validity.  LK_RETRY
1749  * permits vn_lock to return doomed vnodes.
1750  */
1751 static int __noinline
1752 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1753     int error)
1754 {
1755 
1756 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1757 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1758 
1759 	if (error == 0)
1760 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1761 
1762 	if ((flags & LK_RETRY) == 0) {
1763 		if (error == 0) {
1764 			VOP_UNLOCK(vp);
1765 			error = ENOENT;
1766 		}
1767 		return (error);
1768 	}
1769 
1770 	/*
1771 	 * LK_RETRY case.
1772 	 *
1773 	 * Nothing to do if we got the lock.
1774 	 */
1775 	if (error == 0)
1776 		return (0);
1777 
1778 	/*
1779 	 * Interlock was dropped by the call in _vn_lock.
1780 	 */
1781 	flags &= ~LK_INTERLOCK;
1782 	do {
1783 		error = VOP_LOCK1(vp, flags, file, line);
1784 	} while (error != 0);
1785 	return (0);
1786 }
1787 
1788 int
1789 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1790 {
1791 	int error;
1792 
1793 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1794 	    ("vn_lock: no locktype (%d passed)", flags));
1795 	VNPASS(vp->v_holdcnt > 0, vp);
1796 	error = VOP_LOCK1(vp, flags, file, line);
1797 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1798 		return (_vn_lock_fallback(vp, flags, file, line, error));
1799 	return (0);
1800 }
1801 
1802 /*
1803  * File table vnode close routine.
1804  */
1805 static int
1806 vn_closefile(struct file *fp, struct thread *td)
1807 {
1808 	struct vnode *vp;
1809 	struct flock lf;
1810 	int error;
1811 	bool ref;
1812 
1813 	vp = fp->f_vnode;
1814 	fp->f_ops = &badfileops;
1815 	ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1816 
1817 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1818 
1819 	if (__predict_false(ref)) {
1820 		lf.l_whence = SEEK_SET;
1821 		lf.l_start = 0;
1822 		lf.l_len = 0;
1823 		lf.l_type = F_UNLCK;
1824 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1825 		vrele(vp);
1826 	}
1827 	return (error);
1828 }
1829 
1830 /*
1831  * Preparing to start a filesystem write operation. If the operation is
1832  * permitted, then we bump the count of operations in progress and
1833  * proceed. If a suspend request is in progress, we wait until the
1834  * suspension is over, and then proceed.
1835  */
1836 static int
1837 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1838 {
1839 	struct mount_pcpu *mpcpu;
1840 	int error, mflags;
1841 
1842 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1843 	    vfs_op_thread_enter(mp, mpcpu)) {
1844 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1845 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1846 		vfs_op_thread_exit(mp, mpcpu);
1847 		return (0);
1848 	}
1849 
1850 	if (mplocked)
1851 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1852 	else
1853 		MNT_ILOCK(mp);
1854 
1855 	error = 0;
1856 
1857 	/*
1858 	 * Check on status of suspension.
1859 	 */
1860 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1861 	    mp->mnt_susp_owner != curthread) {
1862 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1863 		    (flags & PCATCH) : 0) | (PUSER - 1);
1864 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1865 			if (flags & V_NOWAIT) {
1866 				error = EWOULDBLOCK;
1867 				goto unlock;
1868 			}
1869 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1870 			    "suspfs", 0);
1871 			if (error)
1872 				goto unlock;
1873 		}
1874 	}
1875 	if (flags & V_XSLEEP)
1876 		goto unlock;
1877 	mp->mnt_writeopcount++;
1878 unlock:
1879 	if (error != 0 || (flags & V_XSLEEP) != 0)
1880 		MNT_REL(mp);
1881 	MNT_IUNLOCK(mp);
1882 	return (error);
1883 }
1884 
1885 int
1886 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1887 {
1888 	struct mount *mp;
1889 	int error;
1890 
1891 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1892 	    ("V_MNTREF requires mp"));
1893 
1894 	error = 0;
1895 	/*
1896 	 * If a vnode is provided, get and return the mount point that
1897 	 * to which it will write.
1898 	 */
1899 	if (vp != NULL) {
1900 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1901 			*mpp = NULL;
1902 			if (error != EOPNOTSUPP)
1903 				return (error);
1904 			return (0);
1905 		}
1906 	}
1907 	if ((mp = *mpp) == NULL)
1908 		return (0);
1909 
1910 	/*
1911 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1912 	 * a vfs_ref().
1913 	 * As long as a vnode is not provided we need to acquire a
1914 	 * refcount for the provided mountpoint too, in order to
1915 	 * emulate a vfs_ref().
1916 	 */
1917 	if (vp == NULL && (flags & V_MNTREF) == 0)
1918 		vfs_ref(mp);
1919 
1920 	return (vn_start_write_refed(mp, flags, false));
1921 }
1922 
1923 /*
1924  * Secondary suspension. Used by operations such as vop_inactive
1925  * routines that are needed by the higher level functions. These
1926  * are allowed to proceed until all the higher level functions have
1927  * completed (indicated by mnt_writeopcount dropping to zero). At that
1928  * time, these operations are halted until the suspension is over.
1929  */
1930 int
1931 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1932 {
1933 	struct mount *mp;
1934 	int error;
1935 
1936 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1937 	    ("V_MNTREF requires mp"));
1938 
1939  retry:
1940 	if (vp != NULL) {
1941 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1942 			*mpp = NULL;
1943 			if (error != EOPNOTSUPP)
1944 				return (error);
1945 			return (0);
1946 		}
1947 	}
1948 	/*
1949 	 * If we are not suspended or have not yet reached suspended
1950 	 * mode, then let the operation proceed.
1951 	 */
1952 	if ((mp = *mpp) == NULL)
1953 		return (0);
1954 
1955 	/*
1956 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1957 	 * a vfs_ref().
1958 	 * As long as a vnode is not provided we need to acquire a
1959 	 * refcount for the provided mountpoint too, in order to
1960 	 * emulate a vfs_ref().
1961 	 */
1962 	MNT_ILOCK(mp);
1963 	if (vp == NULL && (flags & V_MNTREF) == 0)
1964 		MNT_REF(mp);
1965 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1966 		mp->mnt_secondary_writes++;
1967 		mp->mnt_secondary_accwrites++;
1968 		MNT_IUNLOCK(mp);
1969 		return (0);
1970 	}
1971 	if (flags & V_NOWAIT) {
1972 		MNT_REL(mp);
1973 		MNT_IUNLOCK(mp);
1974 		return (EWOULDBLOCK);
1975 	}
1976 	/*
1977 	 * Wait for the suspension to finish.
1978 	 */
1979 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1980 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1981 	    "suspfs", 0);
1982 	vfs_rel(mp);
1983 	if (error == 0)
1984 		goto retry;
1985 	return (error);
1986 }
1987 
1988 /*
1989  * Filesystem write operation has completed. If we are suspending and this
1990  * operation is the last one, notify the suspender that the suspension is
1991  * now in effect.
1992  */
1993 void
1994 vn_finished_write(struct mount *mp)
1995 {
1996 	struct mount_pcpu *mpcpu;
1997 	int c;
1998 
1999 	if (mp == NULL)
2000 		return;
2001 
2002 	if (vfs_op_thread_enter(mp, mpcpu)) {
2003 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2004 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2005 		vfs_op_thread_exit(mp, mpcpu);
2006 		return;
2007 	}
2008 
2009 	MNT_ILOCK(mp);
2010 	vfs_assert_mount_counters(mp);
2011 	MNT_REL(mp);
2012 	c = --mp->mnt_writeopcount;
2013 	if (mp->mnt_vfs_ops == 0) {
2014 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2015 		MNT_IUNLOCK(mp);
2016 		return;
2017 	}
2018 	if (c < 0)
2019 		vfs_dump_mount_counters(mp);
2020 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2021 		wakeup(&mp->mnt_writeopcount);
2022 	MNT_IUNLOCK(mp);
2023 }
2024 
2025 /*
2026  * Filesystem secondary write operation has completed. If we are
2027  * suspending and this operation is the last one, notify the suspender
2028  * that the suspension is now in effect.
2029  */
2030 void
2031 vn_finished_secondary_write(struct mount *mp)
2032 {
2033 	if (mp == NULL)
2034 		return;
2035 	MNT_ILOCK(mp);
2036 	MNT_REL(mp);
2037 	mp->mnt_secondary_writes--;
2038 	if (mp->mnt_secondary_writes < 0)
2039 		panic("vn_finished_secondary_write: neg cnt");
2040 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2041 	    mp->mnt_secondary_writes <= 0)
2042 		wakeup(&mp->mnt_secondary_writes);
2043 	MNT_IUNLOCK(mp);
2044 }
2045 
2046 /*
2047  * Request a filesystem to suspend write operations.
2048  */
2049 int
2050 vfs_write_suspend(struct mount *mp, int flags)
2051 {
2052 	int error;
2053 
2054 	vfs_op_enter(mp);
2055 
2056 	MNT_ILOCK(mp);
2057 	vfs_assert_mount_counters(mp);
2058 	if (mp->mnt_susp_owner == curthread) {
2059 		vfs_op_exit_locked(mp);
2060 		MNT_IUNLOCK(mp);
2061 		return (EALREADY);
2062 	}
2063 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2064 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2065 
2066 	/*
2067 	 * Unmount holds a write reference on the mount point.  If we
2068 	 * own busy reference and drain for writers, we deadlock with
2069 	 * the reference draining in the unmount path.  Callers of
2070 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2071 	 * vfs_busy() reference is owned and caller is not in the
2072 	 * unmount context.
2073 	 */
2074 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2075 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2076 		vfs_op_exit_locked(mp);
2077 		MNT_IUNLOCK(mp);
2078 		return (EBUSY);
2079 	}
2080 
2081 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2082 	mp->mnt_susp_owner = curthread;
2083 	if (mp->mnt_writeopcount > 0)
2084 		(void) msleep(&mp->mnt_writeopcount,
2085 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2086 	else
2087 		MNT_IUNLOCK(mp);
2088 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2089 		vfs_write_resume(mp, 0);
2090 		/* vfs_write_resume does vfs_op_exit() for us */
2091 	}
2092 	return (error);
2093 }
2094 
2095 /*
2096  * Request a filesystem to resume write operations.
2097  */
2098 void
2099 vfs_write_resume(struct mount *mp, int flags)
2100 {
2101 
2102 	MNT_ILOCK(mp);
2103 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2104 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2105 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2106 				       MNTK_SUSPENDED);
2107 		mp->mnt_susp_owner = NULL;
2108 		wakeup(&mp->mnt_writeopcount);
2109 		wakeup(&mp->mnt_flag);
2110 		curthread->td_pflags &= ~TDP_IGNSUSP;
2111 		if ((flags & VR_START_WRITE) != 0) {
2112 			MNT_REF(mp);
2113 			mp->mnt_writeopcount++;
2114 		}
2115 		MNT_IUNLOCK(mp);
2116 		if ((flags & VR_NO_SUSPCLR) == 0)
2117 			VFS_SUSP_CLEAN(mp);
2118 		vfs_op_exit(mp);
2119 	} else if ((flags & VR_START_WRITE) != 0) {
2120 		MNT_REF(mp);
2121 		vn_start_write_refed(mp, 0, true);
2122 	} else {
2123 		MNT_IUNLOCK(mp);
2124 	}
2125 }
2126 
2127 /*
2128  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2129  * methods.
2130  */
2131 int
2132 vfs_write_suspend_umnt(struct mount *mp)
2133 {
2134 	int error;
2135 
2136 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2137 	    ("vfs_write_suspend_umnt: recursed"));
2138 
2139 	/* dounmount() already called vn_start_write(). */
2140 	for (;;) {
2141 		vn_finished_write(mp);
2142 		error = vfs_write_suspend(mp, 0);
2143 		if (error != 0) {
2144 			vn_start_write(NULL, &mp, V_WAIT);
2145 			return (error);
2146 		}
2147 		MNT_ILOCK(mp);
2148 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2149 			break;
2150 		MNT_IUNLOCK(mp);
2151 		vn_start_write(NULL, &mp, V_WAIT);
2152 	}
2153 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2154 	wakeup(&mp->mnt_flag);
2155 	MNT_IUNLOCK(mp);
2156 	curthread->td_pflags |= TDP_IGNSUSP;
2157 	return (0);
2158 }
2159 
2160 /*
2161  * Implement kqueues for files by translating it to vnode operation.
2162  */
2163 static int
2164 vn_kqfilter(struct file *fp, struct knote *kn)
2165 {
2166 
2167 	return (VOP_KQFILTER(fp->f_vnode, kn));
2168 }
2169 
2170 int
2171 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2172 {
2173 	if ((fp->f_flag & FKQALLOWED) == 0)
2174 		return (EBADF);
2175 	return (vn_kqfilter(fp, kn));
2176 }
2177 
2178 /*
2179  * Simplified in-kernel wrapper calls for extended attribute access.
2180  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2181  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2182  */
2183 int
2184 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2185     const char *attrname, int *buflen, char *buf, struct thread *td)
2186 {
2187 	struct uio	auio;
2188 	struct iovec	iov;
2189 	int	error;
2190 
2191 	iov.iov_len = *buflen;
2192 	iov.iov_base = buf;
2193 
2194 	auio.uio_iov = &iov;
2195 	auio.uio_iovcnt = 1;
2196 	auio.uio_rw = UIO_READ;
2197 	auio.uio_segflg = UIO_SYSSPACE;
2198 	auio.uio_td = td;
2199 	auio.uio_offset = 0;
2200 	auio.uio_resid = *buflen;
2201 
2202 	if ((ioflg & IO_NODELOCKED) == 0)
2203 		vn_lock(vp, LK_SHARED | LK_RETRY);
2204 
2205 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2206 
2207 	/* authorize attribute retrieval as kernel */
2208 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2209 	    td);
2210 
2211 	if ((ioflg & IO_NODELOCKED) == 0)
2212 		VOP_UNLOCK(vp);
2213 
2214 	if (error == 0) {
2215 		*buflen = *buflen - auio.uio_resid;
2216 	}
2217 
2218 	return (error);
2219 }
2220 
2221 /*
2222  * XXX failure mode if partially written?
2223  */
2224 int
2225 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2226     const char *attrname, int buflen, char *buf, struct thread *td)
2227 {
2228 	struct uio	auio;
2229 	struct iovec	iov;
2230 	struct mount	*mp;
2231 	int	error;
2232 
2233 	iov.iov_len = buflen;
2234 	iov.iov_base = buf;
2235 
2236 	auio.uio_iov = &iov;
2237 	auio.uio_iovcnt = 1;
2238 	auio.uio_rw = UIO_WRITE;
2239 	auio.uio_segflg = UIO_SYSSPACE;
2240 	auio.uio_td = td;
2241 	auio.uio_offset = 0;
2242 	auio.uio_resid = buflen;
2243 
2244 	if ((ioflg & IO_NODELOCKED) == 0) {
2245 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2246 			return (error);
2247 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2248 	}
2249 
2250 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2251 
2252 	/* authorize attribute setting as kernel */
2253 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2254 
2255 	if ((ioflg & IO_NODELOCKED) == 0) {
2256 		vn_finished_write(mp);
2257 		VOP_UNLOCK(vp);
2258 	}
2259 
2260 	return (error);
2261 }
2262 
2263 int
2264 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2265     const char *attrname, struct thread *td)
2266 {
2267 	struct mount	*mp;
2268 	int	error;
2269 
2270 	if ((ioflg & IO_NODELOCKED) == 0) {
2271 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2272 			return (error);
2273 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2274 	}
2275 
2276 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2277 
2278 	/* authorize attribute removal as kernel */
2279 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2280 	if (error == EOPNOTSUPP)
2281 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2282 		    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 static int
2293 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2294     struct vnode **rvp)
2295 {
2296 
2297 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2298 }
2299 
2300 int
2301 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2302 {
2303 
2304 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2305 	    lkflags, rvp));
2306 }
2307 
2308 int
2309 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2310     int lkflags, struct vnode **rvp)
2311 {
2312 	struct mount *mp;
2313 	int ltype, error;
2314 
2315 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2316 	mp = vp->v_mount;
2317 	ltype = VOP_ISLOCKED(vp);
2318 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2319 	    ("vn_vget_ino: vp not locked"));
2320 	error = vfs_busy(mp, MBF_NOWAIT);
2321 	if (error != 0) {
2322 		vfs_ref(mp);
2323 		VOP_UNLOCK(vp);
2324 		error = vfs_busy(mp, 0);
2325 		vn_lock(vp, ltype | LK_RETRY);
2326 		vfs_rel(mp);
2327 		if (error != 0)
2328 			return (ENOENT);
2329 		if (VN_IS_DOOMED(vp)) {
2330 			vfs_unbusy(mp);
2331 			return (ENOENT);
2332 		}
2333 	}
2334 	VOP_UNLOCK(vp);
2335 	error = alloc(mp, alloc_arg, lkflags, rvp);
2336 	vfs_unbusy(mp);
2337 	if (error != 0 || *rvp != vp)
2338 		vn_lock(vp, ltype | LK_RETRY);
2339 	if (VN_IS_DOOMED(vp)) {
2340 		if (error == 0) {
2341 			if (*rvp == vp)
2342 				vunref(vp);
2343 			else
2344 				vput(*rvp);
2345 		}
2346 		error = ENOENT;
2347 	}
2348 	return (error);
2349 }
2350 
2351 int
2352 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2353     struct thread *td)
2354 {
2355 
2356 	if (vp->v_type != VREG || td == NULL)
2357 		return (0);
2358 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2359 	    lim_cur(td, RLIMIT_FSIZE)) {
2360 		PROC_LOCK(td->td_proc);
2361 		kern_psignal(td->td_proc, SIGXFSZ);
2362 		PROC_UNLOCK(td->td_proc);
2363 		return (EFBIG);
2364 	}
2365 	return (0);
2366 }
2367 
2368 int
2369 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2370     struct thread *td)
2371 {
2372 	struct vnode *vp;
2373 
2374 	vp = fp->f_vnode;
2375 #ifdef AUDIT
2376 	vn_lock(vp, LK_SHARED | LK_RETRY);
2377 	AUDIT_ARG_VNODE1(vp);
2378 	VOP_UNLOCK(vp);
2379 #endif
2380 	return (setfmode(td, active_cred, vp, mode));
2381 }
2382 
2383 int
2384 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2385     struct thread *td)
2386 {
2387 	struct vnode *vp;
2388 
2389 	vp = fp->f_vnode;
2390 #ifdef AUDIT
2391 	vn_lock(vp, LK_SHARED | LK_RETRY);
2392 	AUDIT_ARG_VNODE1(vp);
2393 	VOP_UNLOCK(vp);
2394 #endif
2395 	return (setfown(td, active_cred, vp, uid, gid));
2396 }
2397 
2398 void
2399 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2400 {
2401 	vm_object_t object;
2402 
2403 	if ((object = vp->v_object) == NULL)
2404 		return;
2405 	VM_OBJECT_WLOCK(object);
2406 	vm_object_page_remove(object, start, end, 0);
2407 	VM_OBJECT_WUNLOCK(object);
2408 }
2409 
2410 int
2411 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2412 {
2413 	struct vattr va;
2414 	daddr_t bn, bnp;
2415 	uint64_t bsize;
2416 	off_t noff;
2417 	int error;
2418 
2419 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2420 	    ("Wrong command %lu", cmd));
2421 
2422 	if (vn_lock(vp, LK_SHARED) != 0)
2423 		return (EBADF);
2424 	if (vp->v_type != VREG) {
2425 		error = ENOTTY;
2426 		goto unlock;
2427 	}
2428 	error = VOP_GETATTR(vp, &va, cred);
2429 	if (error != 0)
2430 		goto unlock;
2431 	noff = *off;
2432 	if (noff >= va.va_size) {
2433 		error = ENXIO;
2434 		goto unlock;
2435 	}
2436 	bsize = vp->v_mount->mnt_stat.f_iosize;
2437 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2438 	    noff % bsize) {
2439 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2440 		if (error == EOPNOTSUPP) {
2441 			error = ENOTTY;
2442 			goto unlock;
2443 		}
2444 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2445 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2446 			noff = bn * bsize;
2447 			if (noff < *off)
2448 				noff = *off;
2449 			goto unlock;
2450 		}
2451 	}
2452 	if (noff > va.va_size)
2453 		noff = va.va_size;
2454 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2455 	if (cmd == FIOSEEKDATA)
2456 		error = ENXIO;
2457 unlock:
2458 	VOP_UNLOCK(vp);
2459 	if (error == 0)
2460 		*off = noff;
2461 	return (error);
2462 }
2463 
2464 int
2465 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2466 {
2467 	struct ucred *cred;
2468 	struct vnode *vp;
2469 	struct vattr vattr;
2470 	off_t foffset, size;
2471 	int error, noneg;
2472 
2473 	cred = td->td_ucred;
2474 	vp = fp->f_vnode;
2475 	foffset = foffset_lock(fp, 0);
2476 	noneg = (vp->v_type != VCHR);
2477 	error = 0;
2478 	switch (whence) {
2479 	case L_INCR:
2480 		if (noneg &&
2481 		    (foffset < 0 ||
2482 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2483 			error = EOVERFLOW;
2484 			break;
2485 		}
2486 		offset += foffset;
2487 		break;
2488 	case L_XTND:
2489 		vn_lock(vp, LK_SHARED | LK_RETRY);
2490 		error = VOP_GETATTR(vp, &vattr, cred);
2491 		VOP_UNLOCK(vp);
2492 		if (error)
2493 			break;
2494 
2495 		/*
2496 		 * If the file references a disk device, then fetch
2497 		 * the media size and use that to determine the ending
2498 		 * offset.
2499 		 */
2500 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2501 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2502 			vattr.va_size = size;
2503 		if (noneg &&
2504 		    (vattr.va_size > OFF_MAX ||
2505 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2506 			error = EOVERFLOW;
2507 			break;
2508 		}
2509 		offset += vattr.va_size;
2510 		break;
2511 	case L_SET:
2512 		break;
2513 	case SEEK_DATA:
2514 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2515 		if (error == ENOTTY)
2516 			error = EINVAL;
2517 		break;
2518 	case SEEK_HOLE:
2519 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2520 		if (error == ENOTTY)
2521 			error = EINVAL;
2522 		break;
2523 	default:
2524 		error = EINVAL;
2525 	}
2526 	if (error == 0 && noneg && offset < 0)
2527 		error = EINVAL;
2528 	if (error != 0)
2529 		goto drop;
2530 	VFS_KNOTE_UNLOCKED(vp, 0);
2531 	td->td_uretoff.tdu_off = offset;
2532 drop:
2533 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2534 	return (error);
2535 }
2536 
2537 int
2538 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2539     struct thread *td)
2540 {
2541 	int error;
2542 
2543 	/*
2544 	 * Grant permission if the caller is the owner of the file, or
2545 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2546 	 * on the file.  If the time pointer is null, then write
2547 	 * permission on the file is also sufficient.
2548 	 *
2549 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2550 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2551 	 * will be allowed to set the times [..] to the current
2552 	 * server time.
2553 	 */
2554 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2555 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2556 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2557 	return (error);
2558 }
2559 
2560 int
2561 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2562 {
2563 	struct vnode *vp;
2564 	int error;
2565 
2566 	if (fp->f_type == DTYPE_FIFO)
2567 		kif->kf_type = KF_TYPE_FIFO;
2568 	else
2569 		kif->kf_type = KF_TYPE_VNODE;
2570 	vp = fp->f_vnode;
2571 	vref(vp);
2572 	FILEDESC_SUNLOCK(fdp);
2573 	error = vn_fill_kinfo_vnode(vp, kif);
2574 	vrele(vp);
2575 	FILEDESC_SLOCK(fdp);
2576 	return (error);
2577 }
2578 
2579 static inline void
2580 vn_fill_junk(struct kinfo_file *kif)
2581 {
2582 	size_t len, olen;
2583 
2584 	/*
2585 	 * Simulate vn_fullpath returning changing values for a given
2586 	 * vp during e.g. coredump.
2587 	 */
2588 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2589 	olen = strlen(kif->kf_path);
2590 	if (len < olen)
2591 		strcpy(&kif->kf_path[len - 1], "$");
2592 	else
2593 		for (; olen < len; olen++)
2594 			strcpy(&kif->kf_path[olen], "A");
2595 }
2596 
2597 int
2598 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2599 {
2600 	struct vattr va;
2601 	char *fullpath, *freepath;
2602 	int error;
2603 
2604 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2605 	freepath = NULL;
2606 	fullpath = "-";
2607 	error = vn_fullpath(vp, &fullpath, &freepath);
2608 	if (error == 0) {
2609 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2610 	}
2611 	if (freepath != NULL)
2612 		free(freepath, M_TEMP);
2613 
2614 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2615 		vn_fill_junk(kif);
2616 	);
2617 
2618 	/*
2619 	 * Retrieve vnode attributes.
2620 	 */
2621 	va.va_fsid = VNOVAL;
2622 	va.va_rdev = NODEV;
2623 	vn_lock(vp, LK_SHARED | LK_RETRY);
2624 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2625 	VOP_UNLOCK(vp);
2626 	if (error != 0)
2627 		return (error);
2628 	if (va.va_fsid != VNOVAL)
2629 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2630 	else
2631 		kif->kf_un.kf_file.kf_file_fsid =
2632 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2633 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2634 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2635 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2636 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2637 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2638 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2639 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2640 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2641 	return (0);
2642 }
2643 
2644 int
2645 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2646     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2647     struct thread *td)
2648 {
2649 #ifdef HWPMC_HOOKS
2650 	struct pmckern_map_in pkm;
2651 #endif
2652 	struct mount *mp;
2653 	struct vnode *vp;
2654 	vm_object_t object;
2655 	vm_prot_t maxprot;
2656 	boolean_t writecounted;
2657 	int error;
2658 
2659 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2660     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2661 	/*
2662 	 * POSIX shared-memory objects are defined to have
2663 	 * kernel persistence, and are not defined to support
2664 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2665 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2666 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2667 	 * flag to request this behavior.
2668 	 */
2669 	if ((fp->f_flag & FPOSIXSHM) != 0)
2670 		flags |= MAP_NOSYNC;
2671 #endif
2672 	vp = fp->f_vnode;
2673 
2674 	/*
2675 	 * Ensure that file and memory protections are
2676 	 * compatible.  Note that we only worry about
2677 	 * writability if mapping is shared; in this case,
2678 	 * current and max prot are dictated by the open file.
2679 	 * XXX use the vnode instead?  Problem is: what
2680 	 * credentials do we use for determination? What if
2681 	 * proc does a setuid?
2682 	 */
2683 	mp = vp->v_mount;
2684 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2685 		maxprot = VM_PROT_NONE;
2686 		if ((prot & VM_PROT_EXECUTE) != 0)
2687 			return (EACCES);
2688 	} else
2689 		maxprot = VM_PROT_EXECUTE;
2690 	if ((fp->f_flag & FREAD) != 0)
2691 		maxprot |= VM_PROT_READ;
2692 	else if ((prot & VM_PROT_READ) != 0)
2693 		return (EACCES);
2694 
2695 	/*
2696 	 * If we are sharing potential changes via MAP_SHARED and we
2697 	 * are trying to get write permission although we opened it
2698 	 * without asking for it, bail out.
2699 	 */
2700 	if ((flags & MAP_SHARED) != 0) {
2701 		if ((fp->f_flag & FWRITE) != 0)
2702 			maxprot |= VM_PROT_WRITE;
2703 		else if ((prot & VM_PROT_WRITE) != 0)
2704 			return (EACCES);
2705 	} else {
2706 		maxprot |= VM_PROT_WRITE;
2707 		cap_maxprot |= VM_PROT_WRITE;
2708 	}
2709 	maxprot &= cap_maxprot;
2710 
2711 	/*
2712 	 * For regular files and shared memory, POSIX requires that
2713 	 * the value of foff be a legitimate offset within the data
2714 	 * object.  In particular, negative offsets are invalid.
2715 	 * Blocking negative offsets and overflows here avoids
2716 	 * possible wraparound or user-level access into reserved
2717 	 * ranges of the data object later.  In contrast, POSIX does
2718 	 * not dictate how offsets are used by device drivers, so in
2719 	 * the case of a device mapping a negative offset is passed
2720 	 * on.
2721 	 */
2722 	if (
2723 #ifdef _LP64
2724 	    size > OFF_MAX ||
2725 #endif
2726 	    foff > OFF_MAX - size)
2727 		return (EINVAL);
2728 
2729 	writecounted = FALSE;
2730 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2731 	    &foff, &object, &writecounted);
2732 	if (error != 0)
2733 		return (error);
2734 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2735 	    foff, writecounted, td);
2736 	if (error != 0) {
2737 		/*
2738 		 * If this mapping was accounted for in the vnode's
2739 		 * writecount, then undo that now.
2740 		 */
2741 		if (writecounted)
2742 			vm_pager_release_writecount(object, 0, size);
2743 		vm_object_deallocate(object);
2744 	}
2745 #ifdef HWPMC_HOOKS
2746 	/* Inform hwpmc(4) if an executable is being mapped. */
2747 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2748 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2749 			pkm.pm_file = vp;
2750 			pkm.pm_address = (uintptr_t) *addr;
2751 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2752 		}
2753 	}
2754 #endif
2755 	return (error);
2756 }
2757 
2758 void
2759 vn_fsid(struct vnode *vp, struct vattr *va)
2760 {
2761 	fsid_t *f;
2762 
2763 	f = &vp->v_mount->mnt_stat.f_fsid;
2764 	va->va_fsid = (uint32_t)f->val[1];
2765 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2766 	va->va_fsid += (uint32_t)f->val[0];
2767 }
2768 
2769 int
2770 vn_fsync_buf(struct vnode *vp, int waitfor)
2771 {
2772 	struct buf *bp, *nbp;
2773 	struct bufobj *bo;
2774 	struct mount *mp;
2775 	int error, maxretry;
2776 
2777 	error = 0;
2778 	maxretry = 10000;     /* large, arbitrarily chosen */
2779 	mp = NULL;
2780 	if (vp->v_type == VCHR) {
2781 		VI_LOCK(vp);
2782 		mp = vp->v_rdev->si_mountpt;
2783 		VI_UNLOCK(vp);
2784 	}
2785 	bo = &vp->v_bufobj;
2786 	BO_LOCK(bo);
2787 loop1:
2788 	/*
2789 	 * MARK/SCAN initialization to avoid infinite loops.
2790 	 */
2791         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2792 		bp->b_vflags &= ~BV_SCANNED;
2793 		bp->b_error = 0;
2794 	}
2795 
2796 	/*
2797 	 * Flush all dirty buffers associated with a vnode.
2798 	 */
2799 loop2:
2800 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2801 		if ((bp->b_vflags & BV_SCANNED) != 0)
2802 			continue;
2803 		bp->b_vflags |= BV_SCANNED;
2804 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2805 			if (waitfor != MNT_WAIT)
2806 				continue;
2807 			if (BUF_LOCK(bp,
2808 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2809 			    BO_LOCKPTR(bo)) != 0) {
2810 				BO_LOCK(bo);
2811 				goto loop1;
2812 			}
2813 			BO_LOCK(bo);
2814 		}
2815 		BO_UNLOCK(bo);
2816 		KASSERT(bp->b_bufobj == bo,
2817 		    ("bp %p wrong b_bufobj %p should be %p",
2818 		    bp, bp->b_bufobj, bo));
2819 		if ((bp->b_flags & B_DELWRI) == 0)
2820 			panic("fsync: not dirty");
2821 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2822 			vfs_bio_awrite(bp);
2823 		} else {
2824 			bremfree(bp);
2825 			bawrite(bp);
2826 		}
2827 		if (maxretry < 1000)
2828 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2829 		BO_LOCK(bo);
2830 		goto loop2;
2831 	}
2832 
2833 	/*
2834 	 * If synchronous the caller expects us to completely resolve all
2835 	 * dirty buffers in the system.  Wait for in-progress I/O to
2836 	 * complete (which could include background bitmap writes), then
2837 	 * retry if dirty blocks still exist.
2838 	 */
2839 	if (waitfor == MNT_WAIT) {
2840 		bufobj_wwait(bo, 0, 0);
2841 		if (bo->bo_dirty.bv_cnt > 0) {
2842 			/*
2843 			 * If we are unable to write any of these buffers
2844 			 * then we fail now rather than trying endlessly
2845 			 * to write them out.
2846 			 */
2847 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2848 				if ((error = bp->b_error) != 0)
2849 					break;
2850 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2851 			    (error == 0 && --maxretry >= 0))
2852 				goto loop1;
2853 			if (error == 0)
2854 				error = EAGAIN;
2855 		}
2856 	}
2857 	BO_UNLOCK(bo);
2858 	if (error != 0)
2859 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2860 
2861 	return (error);
2862 }
2863 
2864 /*
2865  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2866  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2867  * to do the actual copy.
2868  * vn_generic_copy_file_range() is factored out, so it can be called
2869  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2870  * different file systems.
2871  */
2872 int
2873 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2874     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2875     struct ucred *outcred, struct thread *fsize_td)
2876 {
2877 	int error;
2878 	size_t len;
2879 	uint64_t uval;
2880 
2881 	len = *lenp;
2882 	*lenp = 0;		/* For error returns. */
2883 	error = 0;
2884 
2885 	/* Do some sanity checks on the arguments. */
2886 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2887 		error = EISDIR;
2888 	else if (*inoffp < 0 || *outoffp < 0 ||
2889 	    invp->v_type != VREG || outvp->v_type != VREG)
2890 		error = EINVAL;
2891 	if (error != 0)
2892 		goto out;
2893 
2894 	/* Ensure offset + len does not wrap around. */
2895 	uval = *inoffp;
2896 	uval += len;
2897 	if (uval > INT64_MAX)
2898 		len = INT64_MAX - *inoffp;
2899 	uval = *outoffp;
2900 	uval += len;
2901 	if (uval > INT64_MAX)
2902 		len = INT64_MAX - *outoffp;
2903 	if (len == 0)
2904 		goto out;
2905 
2906 	/*
2907 	 * If the two vnode are for the same file system, call
2908 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2909 	 * which can handle copies across multiple file systems.
2910 	 */
2911 	*lenp = len;
2912 	if (invp->v_mount == outvp->v_mount)
2913 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2914 		    lenp, flags, incred, outcred, fsize_td);
2915 	else
2916 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2917 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2918 out:
2919 	return (error);
2920 }
2921 
2922 /*
2923  * Test len bytes of data starting at dat for all bytes == 0.
2924  * Return true if all bytes are zero, false otherwise.
2925  * Expects dat to be well aligned.
2926  */
2927 static bool
2928 mem_iszero(void *dat, int len)
2929 {
2930 	int i;
2931 	const u_int *p;
2932 	const char *cp;
2933 
2934 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2935 		if (len >= sizeof(*p)) {
2936 			if (*p != 0)
2937 				return (false);
2938 		} else {
2939 			cp = (const char *)p;
2940 			for (i = 0; i < len; i++, cp++)
2941 				if (*cp != '\0')
2942 					return (false);
2943 		}
2944 	}
2945 	return (true);
2946 }
2947 
2948 /*
2949  * Look for a hole in the output file and, if found, adjust *outoffp
2950  * and *xferp to skip past the hole.
2951  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2952  * to be written as 0's upon return.
2953  */
2954 static off_t
2955 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2956     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2957 {
2958 	int error;
2959 	off_t delta;
2960 
2961 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2962 		*dataoffp = *outoffp;
2963 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2964 		    curthread);
2965 		if (error == 0) {
2966 			*holeoffp = *dataoffp;
2967 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2968 			    curthread);
2969 		}
2970 		if (error != 0 || *holeoffp == *dataoffp) {
2971 			/*
2972 			 * Since outvp is unlocked, it may be possible for
2973 			 * another thread to do a truncate(), lseek(), write()
2974 			 * creating a hole at startoff between the above
2975 			 * VOP_IOCTL() calls, if the other thread does not do
2976 			 * rangelocking.
2977 			 * If that happens, *holeoffp == *dataoffp and finding
2978 			 * the hole has failed, so disable vn_skip_hole().
2979 			 */
2980 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2981 			return (xfer2);
2982 		}
2983 		KASSERT(*dataoffp >= *outoffp,
2984 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2985 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2986 		KASSERT(*holeoffp > *dataoffp,
2987 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2988 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2989 	}
2990 
2991 	/*
2992 	 * If there is a hole before the data starts, advance *outoffp and
2993 	 * *xferp past the hole.
2994 	 */
2995 	if (*dataoffp > *outoffp) {
2996 		delta = *dataoffp - *outoffp;
2997 		if (delta >= *xferp) {
2998 			/* Entire *xferp is a hole. */
2999 			*outoffp += *xferp;
3000 			*xferp = 0;
3001 			return (0);
3002 		}
3003 		*xferp -= delta;
3004 		*outoffp += delta;
3005 		xfer2 = MIN(xfer2, *xferp);
3006 	}
3007 
3008 	/*
3009 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3010 	 * that the write ends at the start of the hole.
3011 	 * *holeoffp should always be greater than *outoffp, but for the
3012 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3013 	 * value.
3014 	 */
3015 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3016 		xfer2 = *holeoffp - *outoffp;
3017 	return (xfer2);
3018 }
3019 
3020 /*
3021  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3022  * dat is a maximum of blksize in length and can be written repeatedly in
3023  * the chunk.
3024  * If growfile == true, just grow the file via vn_truncate_locked() instead
3025  * of doing actual writes.
3026  * If checkhole == true, a hole is being punched, so skip over any hole
3027  * already in the output file.
3028  */
3029 static int
3030 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3031     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3032 {
3033 	struct mount *mp;
3034 	off_t dataoff, holeoff, xfer2;
3035 	int error, lckf;
3036 
3037 	/*
3038 	 * Loop around doing writes of blksize until write has been completed.
3039 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3040 	 * done for each iteration, since the xfer argument can be very
3041 	 * large if there is a large hole to punch in the output file.
3042 	 */
3043 	error = 0;
3044 	holeoff = 0;
3045 	do {
3046 		xfer2 = MIN(xfer, blksize);
3047 		if (checkhole) {
3048 			/*
3049 			 * Punching a hole.  Skip writing if there is
3050 			 * already a hole in the output file.
3051 			 */
3052 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3053 			    &dataoff, &holeoff, cred);
3054 			if (xfer == 0)
3055 				break;
3056 			if (holeoff < 0)
3057 				checkhole = false;
3058 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3059 			    (intmax_t)xfer2));
3060 		}
3061 		bwillwrite();
3062 		mp = NULL;
3063 		error = vn_start_write(outvp, &mp, V_WAIT);
3064 		if (error != 0)
3065 			break;
3066 		if (growfile) {
3067 			error = vn_lock(outvp, LK_EXCLUSIVE);
3068 			if (error == 0) {
3069 				error = vn_truncate_locked(outvp, outoff + xfer,
3070 				    false, cred);
3071 				VOP_UNLOCK(outvp);
3072 			}
3073 		} else {
3074 			if (MNT_SHARED_WRITES(mp))
3075 				lckf = LK_SHARED;
3076 			else
3077 				lckf = LK_EXCLUSIVE;
3078 			error = vn_lock(outvp, lckf);
3079 			if (error == 0) {
3080 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3081 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3082 				    curthread->td_ucred, cred, NULL, curthread);
3083 				outoff += xfer2;
3084 				xfer -= xfer2;
3085 				VOP_UNLOCK(outvp);
3086 			}
3087 		}
3088 		if (mp != NULL)
3089 			vn_finished_write(mp);
3090 	} while (!growfile && xfer > 0 && error == 0);
3091 	return (error);
3092 }
3093 
3094 /*
3095  * Copy a byte range of one file to another.  This function can handle the
3096  * case where invp and outvp are on different file systems.
3097  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3098  * is no better file system specific way to do it.
3099  */
3100 int
3101 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3102     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3103     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3104 {
3105 	struct vattr va, inva;
3106 	struct mount *mp;
3107 	struct uio io;
3108 	off_t startoff, endoff, xfer, xfer2;
3109 	u_long blksize;
3110 	int error, interrupted;
3111 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3112 	ssize_t aresid;
3113 	size_t copylen, len, rem, savlen;
3114 	char *dat;
3115 	long holein, holeout;
3116 
3117 	holein = holeout = 0;
3118 	savlen = len = *lenp;
3119 	error = 0;
3120 	interrupted = 0;
3121 	dat = NULL;
3122 
3123 	error = vn_lock(invp, LK_SHARED);
3124 	if (error != 0)
3125 		goto out;
3126 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3127 		holein = 0;
3128 	if (holein > 0)
3129 		error = VOP_GETATTR(invp, &inva, incred);
3130 	VOP_UNLOCK(invp);
3131 	if (error != 0)
3132 		goto out;
3133 
3134 	mp = NULL;
3135 	error = vn_start_write(outvp, &mp, V_WAIT);
3136 	if (error == 0)
3137 		error = vn_lock(outvp, LK_EXCLUSIVE);
3138 	if (error == 0) {
3139 		/*
3140 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3141 		 * now that outvp is locked.
3142 		 */
3143 		if (fsize_td != NULL) {
3144 			io.uio_offset = *outoffp;
3145 			io.uio_resid = len;
3146 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3147 			if (error != 0)
3148 				error = EFBIG;
3149 		}
3150 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3151 			holeout = 0;
3152 		/*
3153 		 * Holes that are past EOF do not need to be written as a block
3154 		 * of zero bytes.  So, truncate the output file as far as
3155 		 * possible and then use va.va_size to decide if writing 0
3156 		 * bytes is necessary in the loop below.
3157 		 */
3158 		if (error == 0)
3159 			error = VOP_GETATTR(outvp, &va, outcred);
3160 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3161 		    *outoffp + len) {
3162 #ifdef MAC
3163 			error = mac_vnode_check_write(curthread->td_ucred,
3164 			    outcred, outvp);
3165 			if (error == 0)
3166 #endif
3167 				error = vn_truncate_locked(outvp, *outoffp,
3168 				    false, outcred);
3169 			if (error == 0)
3170 				va.va_size = *outoffp;
3171 		}
3172 		VOP_UNLOCK(outvp);
3173 	}
3174 	if (mp != NULL)
3175 		vn_finished_write(mp);
3176 	if (error != 0)
3177 		goto out;
3178 
3179 	/*
3180 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3181 	 * If hole sizes aren't available, set the blksize to the larger
3182 	 * f_iosize of invp and outvp.
3183 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3184 	 * This value is clipped at 4Kbytes and 1Mbyte.
3185 	 */
3186 	blksize = MAX(holein, holeout);
3187 
3188 	/* Clip len to end at an exact multiple of hole size. */
3189 	if (blksize > 1) {
3190 		rem = *inoffp % blksize;
3191 		if (rem > 0)
3192 			rem = blksize - rem;
3193 		if (len > rem && len - rem > blksize)
3194 			len = savlen = rounddown(len - rem, blksize) + rem;
3195 	}
3196 
3197 	if (blksize <= 1)
3198 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3199 		    outvp->v_mount->mnt_stat.f_iosize);
3200 	if (blksize < 4096)
3201 		blksize = 4096;
3202 	else if (blksize > 1024 * 1024)
3203 		blksize = 1024 * 1024;
3204 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3205 
3206 	/*
3207 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3208 	 * to find holes.  Otherwise, just scan the read block for all 0s
3209 	 * in the inner loop where the data copying is done.
3210 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3211 	 * support holes on the server, but do not support FIOSEEKHOLE.
3212 	 */
3213 	holetoeof = eof = false;
3214 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3215 		endoff = 0;			/* To shut up compilers. */
3216 		cantseek = true;
3217 		startoff = *inoffp;
3218 		copylen = len;
3219 
3220 		/*
3221 		 * Find the next data area.  If there is just a hole to EOF,
3222 		 * FIOSEEKDATA should fail with ENXIO.
3223 		 * (I do not know if any file system will report a hole to
3224 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3225 		 *  will fail for those file systems.)
3226 		 *
3227 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3228 		 * the code just falls through to the inner copy loop.
3229 		 */
3230 		error = EINVAL;
3231 		if (holein > 0) {
3232 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3233 			    incred, curthread);
3234 			if (error == ENXIO) {
3235 				startoff = endoff = inva.va_size;
3236 				eof = holetoeof = true;
3237 				error = 0;
3238 			}
3239 		}
3240 		if (error == 0 && !holetoeof) {
3241 			endoff = startoff;
3242 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3243 			    incred, curthread);
3244 			/*
3245 			 * Since invp is unlocked, it may be possible for
3246 			 * another thread to do a truncate(), lseek(), write()
3247 			 * creating a hole at startoff between the above
3248 			 * VOP_IOCTL() calls, if the other thread does not do
3249 			 * rangelocking.
3250 			 * If that happens, startoff == endoff and finding
3251 			 * the hole has failed, so set an error.
3252 			 */
3253 			if (error == 0 && startoff == endoff)
3254 				error = EINVAL; /* Any error. Reset to 0. */
3255 		}
3256 		if (error == 0) {
3257 			if (startoff > *inoffp) {
3258 				/* Found hole before data block. */
3259 				xfer = MIN(startoff - *inoffp, len);
3260 				if (*outoffp < va.va_size) {
3261 					/* Must write 0s to punch hole. */
3262 					xfer2 = MIN(va.va_size - *outoffp,
3263 					    xfer);
3264 					memset(dat, 0, MIN(xfer2, blksize));
3265 					error = vn_write_outvp(outvp, dat,
3266 					    *outoffp, xfer2, blksize, false,
3267 					    holeout > 0, outcred);
3268 				}
3269 
3270 				if (error == 0 && *outoffp + xfer >
3271 				    va.va_size && (xfer == len || holetoeof)) {
3272 					/* Grow output file (hole at end). */
3273 					error = vn_write_outvp(outvp, dat,
3274 					    *outoffp, xfer, blksize, true,
3275 					    false, outcred);
3276 				}
3277 				if (error == 0) {
3278 					*inoffp += xfer;
3279 					*outoffp += xfer;
3280 					len -= xfer;
3281 					if (len < savlen)
3282 						interrupted = sig_intr();
3283 				}
3284 			}
3285 			copylen = MIN(len, endoff - startoff);
3286 			cantseek = false;
3287 		} else {
3288 			cantseek = true;
3289 			startoff = *inoffp;
3290 			copylen = len;
3291 			error = 0;
3292 		}
3293 
3294 		xfer = blksize;
3295 		if (cantseek) {
3296 			/*
3297 			 * Set first xfer to end at a block boundary, so that
3298 			 * holes are more likely detected in the loop below via
3299 			 * the for all bytes 0 method.
3300 			 */
3301 			xfer -= (*inoffp % blksize);
3302 		}
3303 		/* Loop copying the data block. */
3304 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3305 			if (copylen < xfer)
3306 				xfer = copylen;
3307 			error = vn_lock(invp, LK_SHARED);
3308 			if (error != 0)
3309 				goto out;
3310 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3311 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3312 			    curthread->td_ucred, incred, &aresid,
3313 			    curthread);
3314 			VOP_UNLOCK(invp);
3315 			lastblock = false;
3316 			if (error == 0 && aresid > 0) {
3317 				/* Stop the copy at EOF on the input file. */
3318 				xfer -= aresid;
3319 				eof = true;
3320 				lastblock = true;
3321 			}
3322 			if (error == 0) {
3323 				/*
3324 				 * Skip the write for holes past the initial EOF
3325 				 * of the output file, unless this is the last
3326 				 * write of the output file at EOF.
3327 				 */
3328 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3329 				    false;
3330 				if (xfer == len)
3331 					lastblock = true;
3332 				if (!cantseek || *outoffp < va.va_size ||
3333 				    lastblock || !readzeros)
3334 					error = vn_write_outvp(outvp, dat,
3335 					    *outoffp, xfer, blksize,
3336 					    readzeros && lastblock &&
3337 					    *outoffp >= va.va_size, false,
3338 					    outcred);
3339 				if (error == 0) {
3340 					*inoffp += xfer;
3341 					startoff += xfer;
3342 					*outoffp += xfer;
3343 					copylen -= xfer;
3344 					len -= xfer;
3345 					if (len < savlen)
3346 						interrupted = sig_intr();
3347 				}
3348 			}
3349 			xfer = blksize;
3350 		}
3351 	}
3352 out:
3353 	*lenp = savlen - len;
3354 	free(dat, M_TEMP);
3355 	return (error);
3356 }
3357 
3358 static int
3359 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3360 {
3361 	struct mount *mp;
3362 	struct vnode *vp;
3363 	off_t olen, ooffset;
3364 	int error;
3365 #ifdef AUDIT
3366 	int audited_vnode1 = 0;
3367 #endif
3368 
3369 	vp = fp->f_vnode;
3370 	if (vp->v_type != VREG)
3371 		return (ENODEV);
3372 
3373 	/* Allocating blocks may take a long time, so iterate. */
3374 	for (;;) {
3375 		olen = len;
3376 		ooffset = offset;
3377 
3378 		bwillwrite();
3379 		mp = NULL;
3380 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3381 		if (error != 0)
3382 			break;
3383 		error = vn_lock(vp, LK_EXCLUSIVE);
3384 		if (error != 0) {
3385 			vn_finished_write(mp);
3386 			break;
3387 		}
3388 #ifdef AUDIT
3389 		if (!audited_vnode1) {
3390 			AUDIT_ARG_VNODE1(vp);
3391 			audited_vnode1 = 1;
3392 		}
3393 #endif
3394 #ifdef MAC
3395 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3396 		if (error == 0)
3397 #endif
3398 			error = VOP_ALLOCATE(vp, &offset, &len);
3399 		VOP_UNLOCK(vp);
3400 		vn_finished_write(mp);
3401 
3402 		if (olen + ooffset != offset + len) {
3403 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3404 			    ooffset, olen, offset, len);
3405 		}
3406 		if (error != 0 || len == 0)
3407 			break;
3408 		KASSERT(olen > len, ("Iteration did not make progress?"));
3409 		maybe_yield();
3410 	}
3411 
3412 	return (error);
3413 }
3414 
3415 static u_long vn_lock_pair_pause_cnt;
3416 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3417     &vn_lock_pair_pause_cnt, 0,
3418     "Count of vn_lock_pair deadlocks");
3419 
3420 u_int vn_lock_pair_pause_max;
3421 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3422     &vn_lock_pair_pause_max, 0,
3423     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3424 
3425 static void
3426 vn_lock_pair_pause(const char *wmesg)
3427 {
3428 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3429 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3430 }
3431 
3432 /*
3433  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3434  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3435  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3436  * can be NULL.
3437  *
3438  * The function returns with both vnodes exclusively locked, and
3439  * guarantees that it does not create lock order reversal with other
3440  * threads during its execution.  Both vnodes could be unlocked
3441  * temporary (and reclaimed).
3442  */
3443 void
3444 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3445     bool vp2_locked)
3446 {
3447 	int error;
3448 
3449 	if (vp1 == NULL && vp2 == NULL)
3450 		return;
3451 	if (vp1 != NULL) {
3452 		if (vp1_locked)
3453 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3454 		else
3455 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3456 	} else {
3457 		vp1_locked = true;
3458 	}
3459 	if (vp2 != NULL) {
3460 		if (vp2_locked)
3461 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3462 		else
3463 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3464 	} else {
3465 		vp2_locked = true;
3466 	}
3467 	if (!vp1_locked && !vp2_locked) {
3468 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3469 		vp1_locked = true;
3470 	}
3471 
3472 	for (;;) {
3473 		if (vp1_locked && vp2_locked)
3474 			break;
3475 		if (vp1_locked && vp2 != NULL) {
3476 			if (vp1 != NULL) {
3477 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3478 				    __FILE__, __LINE__);
3479 				if (error == 0)
3480 					break;
3481 				VOP_UNLOCK(vp1);
3482 				vp1_locked = false;
3483 				vn_lock_pair_pause("vlp1");
3484 			}
3485 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3486 			vp2_locked = true;
3487 		}
3488 		if (vp2_locked && vp1 != NULL) {
3489 			if (vp2 != NULL) {
3490 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3491 				    __FILE__, __LINE__);
3492 				if (error == 0)
3493 					break;
3494 				VOP_UNLOCK(vp2);
3495 				vp2_locked = false;
3496 				vn_lock_pair_pause("vlp2");
3497 			}
3498 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3499 			vp1_locked = true;
3500 		}
3501 	}
3502 	if (vp1 != NULL)
3503 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3504 	if (vp2 != NULL)
3505 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3506 }
3507