xref: /freebsd/sys/kern/vfs_vnops.c (revision 1603881667360c015f6685131f2f25474fa67a72)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/prng.h>
74 #include <sys/sx.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
78 #include <sys/conf.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
81 #include <sys/user.h>
82 #include <sys/ktrace.h>
83 
84 #include <security/audit/audit.h>
85 #include <security/mac/mac_framework.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_extern.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pager.h>
94 
95 #ifdef HWPMC_HOOKS
96 #include <sys/pmckern.h>
97 #endif
98 
99 static fo_rdwr_t	vn_read;
100 static fo_rdwr_t	vn_write;
101 static fo_rdwr_t	vn_io_fault;
102 static fo_truncate_t	vn_truncate;
103 static fo_ioctl_t	vn_ioctl;
104 static fo_poll_t	vn_poll;
105 static fo_kqfilter_t	vn_kqfilter;
106 static fo_close_t	vn_closefile;
107 static fo_mmap_t	vn_mmap;
108 static fo_fallocate_t	vn_fallocate;
109 
110 struct 	fileops vnops = {
111 	.fo_read = vn_io_fault,
112 	.fo_write = vn_io_fault,
113 	.fo_truncate = vn_truncate,
114 	.fo_ioctl = vn_ioctl,
115 	.fo_poll = vn_poll,
116 	.fo_kqfilter = vn_kqfilter,
117 	.fo_stat = vn_statfile,
118 	.fo_close = vn_closefile,
119 	.fo_chmod = vn_chmod,
120 	.fo_chown = vn_chown,
121 	.fo_sendfile = vn_sendfile,
122 	.fo_seek = vn_seek,
123 	.fo_fill_kinfo = vn_fill_kinfo,
124 	.fo_mmap = vn_mmap,
125 	.fo_fallocate = vn_fallocate,
126 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
127 };
128 
129 const u_int io_hold_cnt = 16;
130 static int vn_io_fault_enable = 1;
131 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN,
132     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
133 static int vn_io_fault_prefault = 0;
134 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN,
135     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
136 static int vn_io_pgcache_read_enable = 1;
137 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN,
138     &vn_io_pgcache_read_enable, 0,
139     "Enable copying from page cache for reads, avoiding fs");
140 static u_long vn_io_faults_cnt;
141 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
142     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
143 
144 static int vfs_allow_read_dir = 0;
145 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW,
146     &vfs_allow_read_dir, 0,
147     "Enable read(2) of directory by root for filesystems that support it");
148 
149 /*
150  * Returns true if vn_io_fault mode of handling the i/o request should
151  * be used.
152  */
153 static bool
154 do_vn_io_fault(struct vnode *vp, struct uio *uio)
155 {
156 	struct mount *mp;
157 
158 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
159 	    (mp = vp->v_mount) != NULL &&
160 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
161 }
162 
163 /*
164  * Structure used to pass arguments to vn_io_fault1(), to do either
165  * file- or vnode-based I/O calls.
166  */
167 struct vn_io_fault_args {
168 	enum {
169 		VN_IO_FAULT_FOP,
170 		VN_IO_FAULT_VOP
171 	} kind;
172 	struct ucred *cred;
173 	int flags;
174 	union {
175 		struct fop_args_tag {
176 			struct file *fp;
177 			fo_rdwr_t *doio;
178 		} fop_args;
179 		struct vop_args_tag {
180 			struct vnode *vp;
181 		} vop_args;
182 	} args;
183 };
184 
185 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
186     struct vn_io_fault_args *args, struct thread *td);
187 
188 int
189 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp)
190 {
191 	struct thread *td = ndp->ni_cnd.cn_thread;
192 
193 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
194 }
195 
196 static uint64_t
197 open2nameif(int fmode, u_int vn_open_flags)
198 {
199 	uint64_t res;
200 
201 	res = ISOPEN | LOCKLEAF;
202 	if ((fmode & O_RESOLVE_BENEATH) != 0)
203 		res |= RBENEATH;
204 	if ((fmode & O_EMPTY_PATH) != 0)
205 		res |= EMPTYPATH;
206 	if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0)
207 		res |= AUDITVNODE1;
208 	if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0)
209 		res |= NOCAPCHECK;
210 	return (res);
211 }
212 
213 /*
214  * Common code for vnode open operations via a name lookup.
215  * Lookup the vnode and invoke VOP_CREATE if needed.
216  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
217  *
218  * Note that this does NOT free nameidata for the successful case,
219  * due to the NDINIT being done elsewhere.
220  */
221 int
222 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
223     struct ucred *cred, struct file *fp)
224 {
225 	struct vnode *vp;
226 	struct mount *mp;
227 	struct thread *td = ndp->ni_cnd.cn_thread;
228 	struct vattr vat;
229 	struct vattr *vap = &vat;
230 	int fmode, error;
231 	bool first_open;
232 
233 restart:
234 	first_open = false;
235 	fmode = *flagp;
236 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
237 	    O_EXCL | O_DIRECTORY) ||
238 	    (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH))
239 		return (EINVAL);
240 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
241 		ndp->ni_cnd.cn_nameiop = CREATE;
242 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
243 		/*
244 		 * Set NOCACHE to avoid flushing the cache when
245 		 * rolling in many files at once.
246 		 *
247 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
248 		 * exist despite NOCACHE.
249 		 */
250 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
251 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
252 			ndp->ni_cnd.cn_flags |= FOLLOW;
253 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
254 			bwillwrite();
255 		if ((error = namei(ndp)) != 0)
256 			return (error);
257 		if (ndp->ni_vp == NULL) {
258 			VATTR_NULL(vap);
259 			vap->va_type = VREG;
260 			vap->va_mode = cmode;
261 			if (fmode & O_EXCL)
262 				vap->va_vaflags |= VA_EXCLUSIVE;
263 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
264 				NDFREE(ndp, NDF_ONLY_PNBUF);
265 				vput(ndp->ni_dvp);
266 				if ((error = vn_start_write(NULL, &mp,
267 				    V_XSLEEP | PCATCH)) != 0)
268 					return (error);
269 				NDREINIT(ndp);
270 				goto restart;
271 			}
272 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
273 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
274 #ifdef MAC
275 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
276 			    &ndp->ni_cnd, vap);
277 			if (error == 0)
278 #endif
279 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
280 				    &ndp->ni_cnd, vap);
281 			vp = ndp->ni_vp;
282 			if (error == 0 && (fmode & O_EXCL) != 0 &&
283 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
284 				VI_LOCK(vp);
285 				vp->v_iflag |= VI_FOPENING;
286 				VI_UNLOCK(vp);
287 				first_open = true;
288 			}
289 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
290 			    false);
291 			vn_finished_write(mp);
292 			if (error) {
293 				NDFREE(ndp, NDF_ONLY_PNBUF);
294 				if (error == ERELOOKUP) {
295 					NDREINIT(ndp);
296 					goto restart;
297 				}
298 				return (error);
299 			}
300 			fmode &= ~O_TRUNC;
301 		} else {
302 			if (ndp->ni_dvp == ndp->ni_vp)
303 				vrele(ndp->ni_dvp);
304 			else
305 				vput(ndp->ni_dvp);
306 			ndp->ni_dvp = NULL;
307 			vp = ndp->ni_vp;
308 			if (fmode & O_EXCL) {
309 				error = EEXIST;
310 				goto bad;
311 			}
312 			if (vp->v_type == VDIR) {
313 				error = EISDIR;
314 				goto bad;
315 			}
316 			fmode &= ~O_CREAT;
317 		}
318 	} else {
319 		ndp->ni_cnd.cn_nameiop = LOOKUP;
320 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
321 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
322 		    FOLLOW;
323 		if ((fmode & FWRITE) == 0)
324 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
325 		if ((error = namei(ndp)) != 0)
326 			return (error);
327 		vp = ndp->ni_vp;
328 	}
329 	error = vn_open_vnode(vp, fmode, cred, td, fp);
330 	if (first_open) {
331 		VI_LOCK(vp);
332 		vp->v_iflag &= ~VI_FOPENING;
333 		wakeup(vp);
334 		VI_UNLOCK(vp);
335 	}
336 	if (error)
337 		goto bad;
338 	*flagp = fmode;
339 	return (0);
340 bad:
341 	NDFREE(ndp, NDF_ONLY_PNBUF);
342 	vput(vp);
343 	*flagp = fmode;
344 	ndp->ni_vp = NULL;
345 	return (error);
346 }
347 
348 static int
349 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
350 {
351 	struct flock lf;
352 	int error, lock_flags, type;
353 
354 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
355 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
356 		return (0);
357 	KASSERT(fp != NULL, ("open with flock requires fp"));
358 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
359 		return (EOPNOTSUPP);
360 
361 	lock_flags = VOP_ISLOCKED(vp);
362 	VOP_UNLOCK(vp);
363 
364 	lf.l_whence = SEEK_SET;
365 	lf.l_start = 0;
366 	lf.l_len = 0;
367 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
368 	type = F_FLOCK;
369 	if ((fmode & FNONBLOCK) == 0)
370 		type |= F_WAIT;
371 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
372 		type |= F_FIRSTOPEN;
373 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
374 	if (error == 0)
375 		fp->f_flag |= FHASLOCK;
376 
377 	vn_lock(vp, lock_flags | LK_RETRY);
378 	return (error);
379 }
380 
381 /*
382  * Common code for vnode open operations once a vnode is located.
383  * Check permissions, and call the VOP_OPEN routine.
384  */
385 int
386 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
387     struct thread *td, struct file *fp)
388 {
389 	accmode_t accmode;
390 	int error;
391 
392 	if (vp->v_type == VLNK) {
393 		if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0)
394 			return (EMLINK);
395 	}
396 	if (vp->v_type == VSOCK)
397 		return (EOPNOTSUPP);
398 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
399 		return (ENOTDIR);
400 
401 	accmode = 0;
402 	if ((fmode & O_PATH) == 0) {
403 		if ((fmode & (FWRITE | O_TRUNC)) != 0) {
404 			if (vp->v_type == VDIR)
405 				return (EISDIR);
406 			accmode |= VWRITE;
407 		}
408 		if ((fmode & FREAD) != 0)
409 			accmode |= VREAD;
410 		if ((fmode & O_APPEND) && (fmode & FWRITE))
411 			accmode |= VAPPEND;
412 #ifdef MAC
413 		if ((fmode & O_CREAT) != 0)
414 			accmode |= VCREAT;
415 #endif
416 	}
417 	if ((fmode & FEXEC) != 0)
418 		accmode |= VEXEC;
419 #ifdef MAC
420 	if ((fmode & O_VERIFY) != 0)
421 		accmode |= VVERIFY;
422 	error = mac_vnode_check_open(cred, vp, accmode);
423 	if (error != 0)
424 		return (error);
425 
426 	accmode &= ~(VCREAT | VVERIFY);
427 #endif
428 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
429 		error = VOP_ACCESS(vp, accmode, cred, td);
430 		if (error != 0)
431 			return (error);
432 	}
433 	if ((fmode & O_PATH) != 0) {
434 		if (vp->v_type == VFIFO)
435 			error = EPIPE;
436 		else
437 			error = VOP_ACCESS(vp, VREAD, cred, td);
438 		if (error == 0)
439 			fp->f_flag |= FKQALLOWED;
440 		return (0);
441 	}
442 
443 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
444 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
445 	error = VOP_OPEN(vp, fmode, cred, td, fp);
446 	if (error != 0)
447 		return (error);
448 
449 	error = vn_open_vnode_advlock(vp, fmode, fp);
450 	if (error == 0 && (fmode & FWRITE) != 0) {
451 		error = VOP_ADD_WRITECOUNT(vp, 1);
452 		if (error == 0) {
453 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
454 			     __func__, vp, vp->v_writecount);
455 		}
456 	}
457 
458 	/*
459 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
460 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
461 	 */
462 	if (error != 0) {
463 		if (fp != NULL) {
464 			/*
465 			 * Arrange the call by having fdrop() to use
466 			 * vn_closefile().  This is to satisfy
467 			 * filesystems like devfs or tmpfs, which
468 			 * override fo_close().
469 			 */
470 			fp->f_flag |= FOPENFAILED;
471 			fp->f_vnode = vp;
472 			if (fp->f_ops == &badfileops) {
473 				fp->f_type = DTYPE_VNODE;
474 				fp->f_ops = &vnops;
475 			}
476 			vref(vp);
477 		} else {
478 			/*
479 			 * If there is no fp, due to kernel-mode open,
480 			 * we can call VOP_CLOSE() now.
481 			 */
482 			if (vp->v_type != VFIFO && (fmode & FWRITE) != 0 &&
483 			    !MNT_EXTENDED_SHARED(vp->v_mount) &&
484 			    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
485 				vn_lock(vp, LK_UPGRADE | LK_RETRY);
486 			(void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC),
487 			    cred, td);
488 		}
489 	}
490 
491 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
492 	return (error);
493 
494 }
495 
496 /*
497  * Check for write permissions on the specified vnode.
498  * Prototype text segments cannot be written.
499  * It is racy.
500  */
501 int
502 vn_writechk(struct vnode *vp)
503 {
504 
505 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
506 	/*
507 	 * If there's shared text associated with
508 	 * the vnode, try to free it up once.  If
509 	 * we fail, we can't allow writing.
510 	 */
511 	if (VOP_IS_TEXT(vp))
512 		return (ETXTBSY);
513 
514 	return (0);
515 }
516 
517 /*
518  * Vnode close call
519  */
520 static int
521 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
522     struct thread *td, bool keep_ref)
523 {
524 	struct mount *mp;
525 	int error, lock_flags;
526 
527 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
528 	    MNT_EXTENDED_SHARED(vp->v_mount))
529 		lock_flags = LK_SHARED;
530 	else
531 		lock_flags = LK_EXCLUSIVE;
532 
533 	vn_start_write(vp, &mp, V_WAIT);
534 	vn_lock(vp, lock_flags | LK_RETRY);
535 	AUDIT_ARG_VNODE1(vp);
536 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
537 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
538 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
539 		    __func__, vp, vp->v_writecount);
540 	}
541 	error = VOP_CLOSE(vp, flags, file_cred, td);
542 	if (keep_ref)
543 		VOP_UNLOCK(vp);
544 	else
545 		vput(vp);
546 	vn_finished_write(mp);
547 	return (error);
548 }
549 
550 int
551 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
552     struct thread *td)
553 {
554 
555 	return (vn_close1(vp, flags, file_cred, td, false));
556 }
557 
558 /*
559  * Heuristic to detect sequential operation.
560  */
561 static int
562 sequential_heuristic(struct uio *uio, struct file *fp)
563 {
564 	enum uio_rw rw;
565 
566 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
567 
568 	rw = uio->uio_rw;
569 	if (fp->f_flag & FRDAHEAD)
570 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
571 
572 	/*
573 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
574 	 * that the first I/O is normally considered to be slightly
575 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
576 	 * unless previous seeks have reduced f_seqcount to 0, in which
577 	 * case offset 0 is not special.
578 	 */
579 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
580 	    uio->uio_offset == fp->f_nextoff[rw]) {
581 		/*
582 		 * f_seqcount is in units of fixed-size blocks so that it
583 		 * depends mainly on the amount of sequential I/O and not
584 		 * much on the number of sequential I/O's.  The fixed size
585 		 * of 16384 is hard-coded here since it is (not quite) just
586 		 * a magic size that works well here.  This size is more
587 		 * closely related to the best I/O size for real disks than
588 		 * to any block size used by software.
589 		 */
590 		if (uio->uio_resid >= IO_SEQMAX * 16384)
591 			fp->f_seqcount[rw] = IO_SEQMAX;
592 		else {
593 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
594 			if (fp->f_seqcount[rw] > IO_SEQMAX)
595 				fp->f_seqcount[rw] = IO_SEQMAX;
596 		}
597 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
598 	}
599 
600 	/* Not sequential.  Quickly draw-down sequentiality. */
601 	if (fp->f_seqcount[rw] > 1)
602 		fp->f_seqcount[rw] = 1;
603 	else
604 		fp->f_seqcount[rw] = 0;
605 	return (0);
606 }
607 
608 /*
609  * Package up an I/O request on a vnode into a uio and do it.
610  */
611 int
612 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
613     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
614     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
615 {
616 	struct uio auio;
617 	struct iovec aiov;
618 	struct mount *mp;
619 	struct ucred *cred;
620 	void *rl_cookie;
621 	struct vn_io_fault_args args;
622 	int error, lock_flags;
623 
624 	if (offset < 0 && vp->v_type != VCHR)
625 		return (EINVAL);
626 	auio.uio_iov = &aiov;
627 	auio.uio_iovcnt = 1;
628 	aiov.iov_base = base;
629 	aiov.iov_len = len;
630 	auio.uio_resid = len;
631 	auio.uio_offset = offset;
632 	auio.uio_segflg = segflg;
633 	auio.uio_rw = rw;
634 	auio.uio_td = td;
635 	error = 0;
636 
637 	if ((ioflg & IO_NODELOCKED) == 0) {
638 		if ((ioflg & IO_RANGELOCKED) == 0) {
639 			if (rw == UIO_READ) {
640 				rl_cookie = vn_rangelock_rlock(vp, offset,
641 				    offset + len);
642 			} else if ((ioflg & IO_APPEND) != 0) {
643 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
644 			} else {
645 				rl_cookie = vn_rangelock_wlock(vp, offset,
646 				    offset + len);
647 			}
648 		} else
649 			rl_cookie = NULL;
650 		mp = NULL;
651 		if (rw == UIO_WRITE) {
652 			if (vp->v_type != VCHR &&
653 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
654 			    != 0)
655 				goto out;
656 			if (MNT_SHARED_WRITES(mp) ||
657 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
658 				lock_flags = LK_SHARED;
659 			else
660 				lock_flags = LK_EXCLUSIVE;
661 		} else
662 			lock_flags = LK_SHARED;
663 		vn_lock(vp, lock_flags | LK_RETRY);
664 	} else
665 		rl_cookie = NULL;
666 
667 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
668 #ifdef MAC
669 	if ((ioflg & IO_NOMACCHECK) == 0) {
670 		if (rw == UIO_READ)
671 			error = mac_vnode_check_read(active_cred, file_cred,
672 			    vp);
673 		else
674 			error = mac_vnode_check_write(active_cred, file_cred,
675 			    vp);
676 	}
677 #endif
678 	if (error == 0) {
679 		if (file_cred != NULL)
680 			cred = file_cred;
681 		else
682 			cred = active_cred;
683 		if (do_vn_io_fault(vp, &auio)) {
684 			args.kind = VN_IO_FAULT_VOP;
685 			args.cred = cred;
686 			args.flags = ioflg;
687 			args.args.vop_args.vp = vp;
688 			error = vn_io_fault1(vp, &auio, &args, td);
689 		} else if (rw == UIO_READ) {
690 			error = VOP_READ(vp, &auio, ioflg, cred);
691 		} else /* if (rw == UIO_WRITE) */ {
692 			error = VOP_WRITE(vp, &auio, ioflg, cred);
693 		}
694 	}
695 	if (aresid)
696 		*aresid = auio.uio_resid;
697 	else
698 		if (auio.uio_resid && error == 0)
699 			error = EIO;
700 	if ((ioflg & IO_NODELOCKED) == 0) {
701 		VOP_UNLOCK(vp);
702 		if (mp != NULL)
703 			vn_finished_write(mp);
704 	}
705  out:
706 	if (rl_cookie != NULL)
707 		vn_rangelock_unlock(vp, rl_cookie);
708 	return (error);
709 }
710 
711 /*
712  * Package up an I/O request on a vnode into a uio and do it.  The I/O
713  * request is split up into smaller chunks and we try to avoid saturating
714  * the buffer cache while potentially holding a vnode locked, so we
715  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
716  * to give other processes a chance to lock the vnode (either other processes
717  * core'ing the same binary, or unrelated processes scanning the directory).
718  */
719 int
720 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
721     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
722     struct ucred *file_cred, size_t *aresid, struct thread *td)
723 {
724 	int error = 0;
725 	ssize_t iaresid;
726 
727 	do {
728 		int chunk;
729 
730 		/*
731 		 * Force `offset' to a multiple of MAXBSIZE except possibly
732 		 * for the first chunk, so that filesystems only need to
733 		 * write full blocks except possibly for the first and last
734 		 * chunks.
735 		 */
736 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
737 
738 		if (chunk > len)
739 			chunk = len;
740 		if (rw != UIO_READ && vp->v_type == VREG)
741 			bwillwrite();
742 		iaresid = 0;
743 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
744 		    ioflg, active_cred, file_cred, &iaresid, td);
745 		len -= chunk;	/* aresid calc already includes length */
746 		if (error)
747 			break;
748 		offset += chunk;
749 		base = (char *)base + chunk;
750 		kern_yield(PRI_USER);
751 	} while (len);
752 	if (aresid)
753 		*aresid = len + iaresid;
754 	return (error);
755 }
756 
757 #if OFF_MAX <= LONG_MAX
758 off_t
759 foffset_lock(struct file *fp, int flags)
760 {
761 	volatile short *flagsp;
762 	off_t res;
763 	short state;
764 
765 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
766 
767 	if ((flags & FOF_NOLOCK) != 0)
768 		return (atomic_load_long(&fp->f_offset));
769 
770 	/*
771 	 * According to McKusick the vn lock was protecting f_offset here.
772 	 * It is now protected by the FOFFSET_LOCKED flag.
773 	 */
774 	flagsp = &fp->f_vnread_flags;
775 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
776 		return (atomic_load_long(&fp->f_offset));
777 
778 	sleepq_lock(&fp->f_vnread_flags);
779 	state = atomic_load_16(flagsp);
780 	for (;;) {
781 		if ((state & FOFFSET_LOCKED) == 0) {
782 			if (!atomic_fcmpset_acq_16(flagsp, &state,
783 			    FOFFSET_LOCKED))
784 				continue;
785 			break;
786 		}
787 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
788 			if (!atomic_fcmpset_acq_16(flagsp, &state,
789 			    state | FOFFSET_LOCK_WAITING))
790 				continue;
791 		}
792 		DROP_GIANT();
793 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
794 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
795 		PICKUP_GIANT();
796 		sleepq_lock(&fp->f_vnread_flags);
797 		state = atomic_load_16(flagsp);
798 	}
799 	res = atomic_load_long(&fp->f_offset);
800 	sleepq_release(&fp->f_vnread_flags);
801 	return (res);
802 }
803 
804 void
805 foffset_unlock(struct file *fp, off_t val, int flags)
806 {
807 	volatile short *flagsp;
808 	short state;
809 
810 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
811 
812 	if ((flags & FOF_NOUPDATE) == 0)
813 		atomic_store_long(&fp->f_offset, val);
814 	if ((flags & FOF_NEXTOFF_R) != 0)
815 		fp->f_nextoff[UIO_READ] = val;
816 	if ((flags & FOF_NEXTOFF_W) != 0)
817 		fp->f_nextoff[UIO_WRITE] = val;
818 
819 	if ((flags & FOF_NOLOCK) != 0)
820 		return;
821 
822 	flagsp = &fp->f_vnread_flags;
823 	state = atomic_load_16(flagsp);
824 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
825 	    atomic_cmpset_rel_16(flagsp, state, 0))
826 		return;
827 
828 	sleepq_lock(&fp->f_vnread_flags);
829 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
830 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
831 	fp->f_vnread_flags = 0;
832 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
833 	sleepq_release(&fp->f_vnread_flags);
834 }
835 #else
836 off_t
837 foffset_lock(struct file *fp, int flags)
838 {
839 	struct mtx *mtxp;
840 	off_t res;
841 
842 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
843 
844 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
845 	mtx_lock(mtxp);
846 	if ((flags & FOF_NOLOCK) == 0) {
847 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
848 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
849 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
850 			    "vofflock", 0);
851 		}
852 		fp->f_vnread_flags |= FOFFSET_LOCKED;
853 	}
854 	res = fp->f_offset;
855 	mtx_unlock(mtxp);
856 	return (res);
857 }
858 
859 void
860 foffset_unlock(struct file *fp, off_t val, int flags)
861 {
862 	struct mtx *mtxp;
863 
864 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
865 
866 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
867 	mtx_lock(mtxp);
868 	if ((flags & FOF_NOUPDATE) == 0)
869 		fp->f_offset = val;
870 	if ((flags & FOF_NEXTOFF_R) != 0)
871 		fp->f_nextoff[UIO_READ] = val;
872 	if ((flags & FOF_NEXTOFF_W) != 0)
873 		fp->f_nextoff[UIO_WRITE] = val;
874 	if ((flags & FOF_NOLOCK) == 0) {
875 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
876 		    ("Lost FOFFSET_LOCKED"));
877 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
878 			wakeup(&fp->f_vnread_flags);
879 		fp->f_vnread_flags = 0;
880 	}
881 	mtx_unlock(mtxp);
882 }
883 #endif
884 
885 void
886 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
887 {
888 
889 	if ((flags & FOF_OFFSET) == 0)
890 		uio->uio_offset = foffset_lock(fp, flags);
891 }
892 
893 void
894 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
895 {
896 
897 	if ((flags & FOF_OFFSET) == 0)
898 		foffset_unlock(fp, uio->uio_offset, flags);
899 }
900 
901 static int
902 get_advice(struct file *fp, struct uio *uio)
903 {
904 	struct mtx *mtxp;
905 	int ret;
906 
907 	ret = POSIX_FADV_NORMAL;
908 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
909 		return (ret);
910 
911 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
912 	mtx_lock(mtxp);
913 	if (fp->f_advice != NULL &&
914 	    uio->uio_offset >= fp->f_advice->fa_start &&
915 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
916 		ret = fp->f_advice->fa_advice;
917 	mtx_unlock(mtxp);
918 	return (ret);
919 }
920 
921 int
922 vn_read_from_obj(struct vnode *vp, struct uio *uio)
923 {
924 	vm_object_t obj;
925 	vm_page_t ma[io_hold_cnt + 2];
926 	off_t off, vsz;
927 	ssize_t resid;
928 	int error, i, j;
929 
930 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
931 	obj = atomic_load_ptr(&vp->v_object);
932 	if (obj == NULL)
933 		return (EJUSTRETURN);
934 
935 	/*
936 	 * Depends on type stability of vm_objects.
937 	 */
938 	vm_object_pip_add(obj, 1);
939 	if ((obj->flags & OBJ_DEAD) != 0) {
940 		/*
941 		 * Note that object might be already reused from the
942 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
943 		 * we recheck for DOOMED vnode state after all pages
944 		 * are busied, and retract then.
945 		 *
946 		 * But we check for OBJ_DEAD to ensure that we do not
947 		 * busy pages while vm_object_terminate_pages()
948 		 * processes the queue.
949 		 */
950 		error = EJUSTRETURN;
951 		goto out_pip;
952 	}
953 
954 	resid = uio->uio_resid;
955 	off = uio->uio_offset;
956 	for (i = 0; resid > 0; i++) {
957 		MPASS(i < io_hold_cnt + 2);
958 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
959 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
960 		    VM_ALLOC_NOWAIT);
961 		if (ma[i] == NULL)
962 			break;
963 
964 		/*
965 		 * Skip invalid pages.  Valid mask can be partial only
966 		 * at EOF, and we clip later.
967 		 */
968 		if (vm_page_none_valid(ma[i])) {
969 			vm_page_sunbusy(ma[i]);
970 			break;
971 		}
972 
973 		resid -= PAGE_SIZE;
974 		off += PAGE_SIZE;
975 	}
976 	if (i == 0) {
977 		error = EJUSTRETURN;
978 		goto out_pip;
979 	}
980 
981 	/*
982 	 * Check VIRF_DOOMED after we busied our pages.  Since
983 	 * vgonel() terminates the vnode' vm_object, it cannot
984 	 * process past pages busied by us.
985 	 */
986 	if (VN_IS_DOOMED(vp)) {
987 		error = EJUSTRETURN;
988 		goto out;
989 	}
990 
991 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
992 	if (resid > uio->uio_resid)
993 		resid = uio->uio_resid;
994 
995 	/*
996 	 * Unlocked read of vnp_size is safe because truncation cannot
997 	 * pass busied page.  But we load vnp_size into a local
998 	 * variable so that possible concurrent extension does not
999 	 * break calculation.
1000 	 */
1001 #if defined(__powerpc__) && !defined(__powerpc64__)
1002 	vsz = obj->un_pager.vnp.vnp_size;
1003 #else
1004 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
1005 #endif
1006 	if (uio->uio_offset >= vsz) {
1007 		error = EJUSTRETURN;
1008 		goto out;
1009 	}
1010 	if (uio->uio_offset + resid > vsz)
1011 		resid = vsz - uio->uio_offset;
1012 
1013 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
1014 
1015 out:
1016 	for (j = 0; j < i; j++) {
1017 		if (error == 0)
1018 			vm_page_reference(ma[j]);
1019 		vm_page_sunbusy(ma[j]);
1020 	}
1021 out_pip:
1022 	vm_object_pip_wakeup(obj);
1023 	if (error != 0)
1024 		return (error);
1025 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
1026 }
1027 
1028 /*
1029  * File table vnode read routine.
1030  */
1031 static int
1032 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1033     struct thread *td)
1034 {
1035 	struct vnode *vp;
1036 	off_t orig_offset;
1037 	int error, ioflag;
1038 	int advice;
1039 
1040 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1041 	    uio->uio_td, td));
1042 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1043 	vp = fp->f_vnode;
1044 	ioflag = 0;
1045 	if (fp->f_flag & FNONBLOCK)
1046 		ioflag |= IO_NDELAY;
1047 	if (fp->f_flag & O_DIRECT)
1048 		ioflag |= IO_DIRECT;
1049 
1050 	/*
1051 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1052 	 * allows us to avoid unneeded work outright.
1053 	 */
1054 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1055 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1056 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1057 		if (error == 0) {
1058 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1059 			return (0);
1060 		}
1061 		if (error != EJUSTRETURN)
1062 			return (error);
1063 	}
1064 
1065 	advice = get_advice(fp, uio);
1066 	vn_lock(vp, LK_SHARED | LK_RETRY);
1067 
1068 	switch (advice) {
1069 	case POSIX_FADV_NORMAL:
1070 	case POSIX_FADV_SEQUENTIAL:
1071 	case POSIX_FADV_NOREUSE:
1072 		ioflag |= sequential_heuristic(uio, fp);
1073 		break;
1074 	case POSIX_FADV_RANDOM:
1075 		/* Disable read-ahead for random I/O. */
1076 		break;
1077 	}
1078 	orig_offset = uio->uio_offset;
1079 
1080 #ifdef MAC
1081 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1082 	if (error == 0)
1083 #endif
1084 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1085 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1086 	VOP_UNLOCK(vp);
1087 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1088 	    orig_offset != uio->uio_offset)
1089 		/*
1090 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1091 		 * for the backing file after a POSIX_FADV_NOREUSE
1092 		 * read(2).
1093 		 */
1094 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1095 		    POSIX_FADV_DONTNEED);
1096 	return (error);
1097 }
1098 
1099 /*
1100  * File table vnode write routine.
1101  */
1102 static int
1103 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1104     struct thread *td)
1105 {
1106 	struct vnode *vp;
1107 	struct mount *mp;
1108 	off_t orig_offset;
1109 	int error, ioflag, lock_flags;
1110 	int advice;
1111 	bool need_finished_write;
1112 
1113 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1114 	    uio->uio_td, td));
1115 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1116 	vp = fp->f_vnode;
1117 	if (vp->v_type == VREG)
1118 		bwillwrite();
1119 	ioflag = IO_UNIT;
1120 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1121 		ioflag |= IO_APPEND;
1122 	if (fp->f_flag & FNONBLOCK)
1123 		ioflag |= IO_NDELAY;
1124 	if (fp->f_flag & O_DIRECT)
1125 		ioflag |= IO_DIRECT;
1126 
1127 	mp = atomic_load_ptr(&vp->v_mount);
1128 	if ((fp->f_flag & O_FSYNC) ||
1129 	    (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS)))
1130 		ioflag |= IO_SYNC;
1131 
1132 	/*
1133 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1134 	 * implementations that don't understand IO_DATASYNC fall back to full
1135 	 * O_SYNC behavior.
1136 	 */
1137 	if (fp->f_flag & O_DSYNC)
1138 		ioflag |= IO_SYNC | IO_DATASYNC;
1139 	mp = NULL;
1140 	need_finished_write = false;
1141 	if (vp->v_type != VCHR) {
1142 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1143 		if (error != 0)
1144 			goto unlock;
1145 		need_finished_write = true;
1146 	}
1147 
1148 	advice = get_advice(fp, uio);
1149 
1150 	if (MNT_SHARED_WRITES(mp) ||
1151 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1152 		lock_flags = LK_SHARED;
1153 	} else {
1154 		lock_flags = LK_EXCLUSIVE;
1155 	}
1156 
1157 	vn_lock(vp, lock_flags | LK_RETRY);
1158 	switch (advice) {
1159 	case POSIX_FADV_NORMAL:
1160 	case POSIX_FADV_SEQUENTIAL:
1161 	case POSIX_FADV_NOREUSE:
1162 		ioflag |= sequential_heuristic(uio, fp);
1163 		break;
1164 	case POSIX_FADV_RANDOM:
1165 		/* XXX: Is this correct? */
1166 		break;
1167 	}
1168 	orig_offset = uio->uio_offset;
1169 
1170 #ifdef MAC
1171 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1172 	if (error == 0)
1173 #endif
1174 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1175 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1176 	VOP_UNLOCK(vp);
1177 	if (need_finished_write)
1178 		vn_finished_write(mp);
1179 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1180 	    orig_offset != uio->uio_offset)
1181 		/*
1182 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1183 		 * for the backing file after a POSIX_FADV_NOREUSE
1184 		 * write(2).
1185 		 */
1186 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1187 		    POSIX_FADV_DONTNEED);
1188 unlock:
1189 	return (error);
1190 }
1191 
1192 /*
1193  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1194  * prevent the following deadlock:
1195  *
1196  * Assume that the thread A reads from the vnode vp1 into userspace
1197  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1198  * currently not resident, then system ends up with the call chain
1199  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1200  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1201  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1202  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1203  * backed by the pages of vnode vp1, and some page in buf2 is not
1204  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1205  *
1206  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1207  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1208  * Instead, it first tries to do the whole range i/o with pagefaults
1209  * disabled. If all pages in the i/o buffer are resident and mapped,
1210  * VOP will succeed (ignoring the genuine filesystem errors).
1211  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1212  * i/o in chunks, with all pages in the chunk prefaulted and held
1213  * using vm_fault_quick_hold_pages().
1214  *
1215  * Filesystems using this deadlock avoidance scheme should use the
1216  * array of the held pages from uio, saved in the curthread->td_ma,
1217  * instead of doing uiomove().  A helper function
1218  * vn_io_fault_uiomove() converts uiomove request into
1219  * uiomove_fromphys() over td_ma array.
1220  *
1221  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1222  * make the current i/o request atomic with respect to other i/os and
1223  * truncations.
1224  */
1225 
1226 /*
1227  * Decode vn_io_fault_args and perform the corresponding i/o.
1228  */
1229 static int
1230 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1231     struct thread *td)
1232 {
1233 	int error, save;
1234 
1235 	error = 0;
1236 	save = vm_fault_disable_pagefaults();
1237 	switch (args->kind) {
1238 	case VN_IO_FAULT_FOP:
1239 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1240 		    uio, args->cred, args->flags, td);
1241 		break;
1242 	case VN_IO_FAULT_VOP:
1243 		if (uio->uio_rw == UIO_READ) {
1244 			error = VOP_READ(args->args.vop_args.vp, uio,
1245 			    args->flags, args->cred);
1246 		} else if (uio->uio_rw == UIO_WRITE) {
1247 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1248 			    args->flags, args->cred);
1249 		}
1250 		break;
1251 	default:
1252 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1253 		    args->kind, uio->uio_rw);
1254 	}
1255 	vm_fault_enable_pagefaults(save);
1256 	return (error);
1257 }
1258 
1259 static int
1260 vn_io_fault_touch(char *base, const struct uio *uio)
1261 {
1262 	int r;
1263 
1264 	r = fubyte(base);
1265 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1266 		return (EFAULT);
1267 	return (0);
1268 }
1269 
1270 static int
1271 vn_io_fault_prefault_user(const struct uio *uio)
1272 {
1273 	char *base;
1274 	const struct iovec *iov;
1275 	size_t len;
1276 	ssize_t resid;
1277 	int error, i;
1278 
1279 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1280 	    ("vn_io_fault_prefault userspace"));
1281 
1282 	error = i = 0;
1283 	iov = uio->uio_iov;
1284 	resid = uio->uio_resid;
1285 	base = iov->iov_base;
1286 	len = iov->iov_len;
1287 	while (resid > 0) {
1288 		error = vn_io_fault_touch(base, uio);
1289 		if (error != 0)
1290 			break;
1291 		if (len < PAGE_SIZE) {
1292 			if (len != 0) {
1293 				error = vn_io_fault_touch(base + len - 1, uio);
1294 				if (error != 0)
1295 					break;
1296 				resid -= len;
1297 			}
1298 			if (++i >= uio->uio_iovcnt)
1299 				break;
1300 			iov = uio->uio_iov + i;
1301 			base = iov->iov_base;
1302 			len = iov->iov_len;
1303 		} else {
1304 			len -= PAGE_SIZE;
1305 			base += PAGE_SIZE;
1306 			resid -= PAGE_SIZE;
1307 		}
1308 	}
1309 	return (error);
1310 }
1311 
1312 /*
1313  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1314  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1315  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1316  * into args and call vn_io_fault1() to handle faults during the user
1317  * mode buffer accesses.
1318  */
1319 static int
1320 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1321     struct thread *td)
1322 {
1323 	vm_page_t ma[io_hold_cnt + 2];
1324 	struct uio *uio_clone, short_uio;
1325 	struct iovec short_iovec[1];
1326 	vm_page_t *prev_td_ma;
1327 	vm_prot_t prot;
1328 	vm_offset_t addr, end;
1329 	size_t len, resid;
1330 	ssize_t adv;
1331 	int error, cnt, saveheld, prev_td_ma_cnt;
1332 
1333 	if (vn_io_fault_prefault) {
1334 		error = vn_io_fault_prefault_user(uio);
1335 		if (error != 0)
1336 			return (error); /* Or ignore ? */
1337 	}
1338 
1339 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1340 
1341 	/*
1342 	 * The UFS follows IO_UNIT directive and replays back both
1343 	 * uio_offset and uio_resid if an error is encountered during the
1344 	 * operation.  But, since the iovec may be already advanced,
1345 	 * uio is still in an inconsistent state.
1346 	 *
1347 	 * Cache a copy of the original uio, which is advanced to the redo
1348 	 * point using UIO_NOCOPY below.
1349 	 */
1350 	uio_clone = cloneuio(uio);
1351 	resid = uio->uio_resid;
1352 
1353 	short_uio.uio_segflg = UIO_USERSPACE;
1354 	short_uio.uio_rw = uio->uio_rw;
1355 	short_uio.uio_td = uio->uio_td;
1356 
1357 	error = vn_io_fault_doio(args, uio, td);
1358 	if (error != EFAULT)
1359 		goto out;
1360 
1361 	atomic_add_long(&vn_io_faults_cnt, 1);
1362 	uio_clone->uio_segflg = UIO_NOCOPY;
1363 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1364 	uio_clone->uio_segflg = uio->uio_segflg;
1365 
1366 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1367 	prev_td_ma = td->td_ma;
1368 	prev_td_ma_cnt = td->td_ma_cnt;
1369 
1370 	while (uio_clone->uio_resid != 0) {
1371 		len = uio_clone->uio_iov->iov_len;
1372 		if (len == 0) {
1373 			KASSERT(uio_clone->uio_iovcnt >= 1,
1374 			    ("iovcnt underflow"));
1375 			uio_clone->uio_iov++;
1376 			uio_clone->uio_iovcnt--;
1377 			continue;
1378 		}
1379 		if (len > ptoa(io_hold_cnt))
1380 			len = ptoa(io_hold_cnt);
1381 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1382 		end = round_page(addr + len);
1383 		if (end < addr) {
1384 			error = EFAULT;
1385 			break;
1386 		}
1387 		cnt = atop(end - trunc_page(addr));
1388 		/*
1389 		 * A perfectly misaligned address and length could cause
1390 		 * both the start and the end of the chunk to use partial
1391 		 * page.  +2 accounts for such a situation.
1392 		 */
1393 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1394 		    addr, len, prot, ma, io_hold_cnt + 2);
1395 		if (cnt == -1) {
1396 			error = EFAULT;
1397 			break;
1398 		}
1399 		short_uio.uio_iov = &short_iovec[0];
1400 		short_iovec[0].iov_base = (void *)addr;
1401 		short_uio.uio_iovcnt = 1;
1402 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1403 		short_uio.uio_offset = uio_clone->uio_offset;
1404 		td->td_ma = ma;
1405 		td->td_ma_cnt = cnt;
1406 
1407 		error = vn_io_fault_doio(args, &short_uio, td);
1408 		vm_page_unhold_pages(ma, cnt);
1409 		adv = len - short_uio.uio_resid;
1410 
1411 		uio_clone->uio_iov->iov_base =
1412 		    (char *)uio_clone->uio_iov->iov_base + adv;
1413 		uio_clone->uio_iov->iov_len -= adv;
1414 		uio_clone->uio_resid -= adv;
1415 		uio_clone->uio_offset += adv;
1416 
1417 		uio->uio_resid -= adv;
1418 		uio->uio_offset += adv;
1419 
1420 		if (error != 0 || adv == 0)
1421 			break;
1422 	}
1423 	td->td_ma = prev_td_ma;
1424 	td->td_ma_cnt = prev_td_ma_cnt;
1425 	curthread_pflags_restore(saveheld);
1426 out:
1427 	free(uio_clone, M_IOV);
1428 	return (error);
1429 }
1430 
1431 static int
1432 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1433     int flags, struct thread *td)
1434 {
1435 	fo_rdwr_t *doio;
1436 	struct vnode *vp;
1437 	void *rl_cookie;
1438 	struct vn_io_fault_args args;
1439 	int error;
1440 
1441 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1442 	vp = fp->f_vnode;
1443 
1444 	/*
1445 	 * The ability to read(2) on a directory has historically been
1446 	 * allowed for all users, but this can and has been the source of
1447 	 * at least one security issue in the past.  As such, it is now hidden
1448 	 * away behind a sysctl for those that actually need it to use it, and
1449 	 * restricted to root when it's turned on to make it relatively safe to
1450 	 * leave on for longer sessions of need.
1451 	 */
1452 	if (vp->v_type == VDIR) {
1453 		KASSERT(uio->uio_rw == UIO_READ,
1454 		    ("illegal write attempted on a directory"));
1455 		if (!vfs_allow_read_dir)
1456 			return (EISDIR);
1457 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1458 			return (EISDIR);
1459 	}
1460 
1461 	foffset_lock_uio(fp, uio, flags);
1462 	if (do_vn_io_fault(vp, uio)) {
1463 		args.kind = VN_IO_FAULT_FOP;
1464 		args.args.fop_args.fp = fp;
1465 		args.args.fop_args.doio = doio;
1466 		args.cred = active_cred;
1467 		args.flags = flags | FOF_OFFSET;
1468 		if (uio->uio_rw == UIO_READ) {
1469 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1470 			    uio->uio_offset + uio->uio_resid);
1471 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1472 		    (flags & FOF_OFFSET) == 0) {
1473 			/* For appenders, punt and lock the whole range. */
1474 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1475 		} else {
1476 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1477 			    uio->uio_offset + uio->uio_resid);
1478 		}
1479 		error = vn_io_fault1(vp, uio, &args, td);
1480 		vn_rangelock_unlock(vp, rl_cookie);
1481 	} else {
1482 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1483 	}
1484 	foffset_unlock_uio(fp, uio, flags);
1485 	return (error);
1486 }
1487 
1488 /*
1489  * Helper function to perform the requested uiomove operation using
1490  * the held pages for io->uio_iov[0].iov_base buffer instead of
1491  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1492  * instead of iov_base prevents page faults that could occur due to
1493  * pmap_collect() invalidating the mapping created by
1494  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1495  * object cleanup revoking the write access from page mappings.
1496  *
1497  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1498  * instead of plain uiomove().
1499  */
1500 int
1501 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1502 {
1503 	struct uio transp_uio;
1504 	struct iovec transp_iov[1];
1505 	struct thread *td;
1506 	size_t adv;
1507 	int error, pgadv;
1508 
1509 	td = curthread;
1510 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1511 	    uio->uio_segflg != UIO_USERSPACE)
1512 		return (uiomove(data, xfersize, uio));
1513 
1514 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1515 	transp_iov[0].iov_base = data;
1516 	transp_uio.uio_iov = &transp_iov[0];
1517 	transp_uio.uio_iovcnt = 1;
1518 	if (xfersize > uio->uio_resid)
1519 		xfersize = uio->uio_resid;
1520 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1521 	transp_uio.uio_offset = 0;
1522 	transp_uio.uio_segflg = UIO_SYSSPACE;
1523 	/*
1524 	 * Since transp_iov points to data, and td_ma page array
1525 	 * corresponds to original uio->uio_iov, we need to invert the
1526 	 * direction of the i/o operation as passed to
1527 	 * uiomove_fromphys().
1528 	 */
1529 	switch (uio->uio_rw) {
1530 	case UIO_WRITE:
1531 		transp_uio.uio_rw = UIO_READ;
1532 		break;
1533 	case UIO_READ:
1534 		transp_uio.uio_rw = UIO_WRITE;
1535 		break;
1536 	}
1537 	transp_uio.uio_td = uio->uio_td;
1538 	error = uiomove_fromphys(td->td_ma,
1539 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1540 	    xfersize, &transp_uio);
1541 	adv = xfersize - transp_uio.uio_resid;
1542 	pgadv =
1543 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1544 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1545 	td->td_ma += pgadv;
1546 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1547 	    pgadv));
1548 	td->td_ma_cnt -= pgadv;
1549 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1550 	uio->uio_iov->iov_len -= adv;
1551 	uio->uio_resid -= adv;
1552 	uio->uio_offset += adv;
1553 	return (error);
1554 }
1555 
1556 int
1557 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1558     struct uio *uio)
1559 {
1560 	struct thread *td;
1561 	vm_offset_t iov_base;
1562 	int cnt, pgadv;
1563 
1564 	td = curthread;
1565 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1566 	    uio->uio_segflg != UIO_USERSPACE)
1567 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1568 
1569 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1570 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1571 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1572 	switch (uio->uio_rw) {
1573 	case UIO_WRITE:
1574 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1575 		    offset, cnt);
1576 		break;
1577 	case UIO_READ:
1578 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1579 		    cnt);
1580 		break;
1581 	}
1582 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1583 	td->td_ma += pgadv;
1584 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1585 	    pgadv));
1586 	td->td_ma_cnt -= pgadv;
1587 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1588 	uio->uio_iov->iov_len -= cnt;
1589 	uio->uio_resid -= cnt;
1590 	uio->uio_offset += cnt;
1591 	return (0);
1592 }
1593 
1594 /*
1595  * File table truncate routine.
1596  */
1597 static int
1598 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1599     struct thread *td)
1600 {
1601 	struct mount *mp;
1602 	struct vnode *vp;
1603 	void *rl_cookie;
1604 	int error;
1605 
1606 	vp = fp->f_vnode;
1607 
1608 retry:
1609 	/*
1610 	 * Lock the whole range for truncation.  Otherwise split i/o
1611 	 * might happen partly before and partly after the truncation.
1612 	 */
1613 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1614 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1615 	if (error)
1616 		goto out1;
1617 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1618 	AUDIT_ARG_VNODE1(vp);
1619 	if (vp->v_type == VDIR) {
1620 		error = EISDIR;
1621 		goto out;
1622 	}
1623 #ifdef MAC
1624 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1625 	if (error)
1626 		goto out;
1627 #endif
1628 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1629 	    fp->f_cred);
1630 out:
1631 	VOP_UNLOCK(vp);
1632 	vn_finished_write(mp);
1633 out1:
1634 	vn_rangelock_unlock(vp, rl_cookie);
1635 	if (error == ERELOOKUP)
1636 		goto retry;
1637 	return (error);
1638 }
1639 
1640 /*
1641  * Truncate a file that is already locked.
1642  */
1643 int
1644 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1645     struct ucred *cred)
1646 {
1647 	struct vattr vattr;
1648 	int error;
1649 
1650 	error = VOP_ADD_WRITECOUNT(vp, 1);
1651 	if (error == 0) {
1652 		VATTR_NULL(&vattr);
1653 		vattr.va_size = length;
1654 		if (sync)
1655 			vattr.va_vaflags |= VA_SYNC;
1656 		error = VOP_SETATTR(vp, &vattr, cred);
1657 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1658 	}
1659 	return (error);
1660 }
1661 
1662 /*
1663  * File table vnode stat routine.
1664  */
1665 int
1666 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1667     struct thread *td)
1668 {
1669 	struct vnode *vp = fp->f_vnode;
1670 	int error;
1671 
1672 	vn_lock(vp, LK_SHARED | LK_RETRY);
1673 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1674 	VOP_UNLOCK(vp);
1675 
1676 	return (error);
1677 }
1678 
1679 /*
1680  * File table vnode ioctl routine.
1681  */
1682 static int
1683 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1684     struct thread *td)
1685 {
1686 	struct vattr vattr;
1687 	struct vnode *vp;
1688 	struct fiobmap2_arg *bmarg;
1689 	int error;
1690 
1691 	vp = fp->f_vnode;
1692 	switch (vp->v_type) {
1693 	case VDIR:
1694 	case VREG:
1695 		switch (com) {
1696 		case FIONREAD:
1697 			vn_lock(vp, LK_SHARED | LK_RETRY);
1698 			error = VOP_GETATTR(vp, &vattr, active_cred);
1699 			VOP_UNLOCK(vp);
1700 			if (error == 0)
1701 				*(int *)data = vattr.va_size - fp->f_offset;
1702 			return (error);
1703 		case FIOBMAP2:
1704 			bmarg = (struct fiobmap2_arg *)data;
1705 			vn_lock(vp, LK_SHARED | LK_RETRY);
1706 #ifdef MAC
1707 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1708 			    vp);
1709 			if (error == 0)
1710 #endif
1711 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1712 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1713 			VOP_UNLOCK(vp);
1714 			return (error);
1715 		case FIONBIO:
1716 		case FIOASYNC:
1717 			return (0);
1718 		default:
1719 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1720 			    active_cred, td));
1721 		}
1722 		break;
1723 	case VCHR:
1724 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1725 		    active_cred, td));
1726 	default:
1727 		return (ENOTTY);
1728 	}
1729 }
1730 
1731 /*
1732  * File table vnode poll routine.
1733  */
1734 static int
1735 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1736     struct thread *td)
1737 {
1738 	struct vnode *vp;
1739 	int error;
1740 
1741 	vp = fp->f_vnode;
1742 #if defined(MAC) || defined(AUDIT)
1743 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1744 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1745 		AUDIT_ARG_VNODE1(vp);
1746 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1747 		VOP_UNLOCK(vp);
1748 		if (error != 0)
1749 			return (error);
1750 	}
1751 #endif
1752 	error = VOP_POLL(vp, events, fp->f_cred, td);
1753 	return (error);
1754 }
1755 
1756 /*
1757  * Acquire the requested lock and then check for validity.  LK_RETRY
1758  * permits vn_lock to return doomed vnodes.
1759  */
1760 static int __noinline
1761 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1762     int error)
1763 {
1764 
1765 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1766 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1767 
1768 	if (error == 0)
1769 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1770 
1771 	if ((flags & LK_RETRY) == 0) {
1772 		if (error == 0) {
1773 			VOP_UNLOCK(vp);
1774 			error = ENOENT;
1775 		}
1776 		return (error);
1777 	}
1778 
1779 	/*
1780 	 * LK_RETRY case.
1781 	 *
1782 	 * Nothing to do if we got the lock.
1783 	 */
1784 	if (error == 0)
1785 		return (0);
1786 
1787 	/*
1788 	 * Interlock was dropped by the call in _vn_lock.
1789 	 */
1790 	flags &= ~LK_INTERLOCK;
1791 	do {
1792 		error = VOP_LOCK1(vp, flags, file, line);
1793 	} while (error != 0);
1794 	return (0);
1795 }
1796 
1797 int
1798 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1799 {
1800 	int error;
1801 
1802 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1803 	    ("vn_lock: no locktype (%d passed)", flags));
1804 	VNPASS(vp->v_holdcnt > 0, vp);
1805 	error = VOP_LOCK1(vp, flags, file, line);
1806 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1807 		return (_vn_lock_fallback(vp, flags, file, line, error));
1808 	return (0);
1809 }
1810 
1811 /*
1812  * File table vnode close routine.
1813  */
1814 static int
1815 vn_closefile(struct file *fp, struct thread *td)
1816 {
1817 	struct vnode *vp;
1818 	struct flock lf;
1819 	int error;
1820 	bool ref;
1821 
1822 	vp = fp->f_vnode;
1823 	fp->f_ops = &badfileops;
1824 	ref = (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1825 
1826 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1827 
1828 	if (__predict_false(ref)) {
1829 		lf.l_whence = SEEK_SET;
1830 		lf.l_start = 0;
1831 		lf.l_len = 0;
1832 		lf.l_type = F_UNLCK;
1833 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1834 		vrele(vp);
1835 	}
1836 	return (error);
1837 }
1838 
1839 /*
1840  * Preparing to start a filesystem write operation. If the operation is
1841  * permitted, then we bump the count of operations in progress and
1842  * proceed. If a suspend request is in progress, we wait until the
1843  * suspension is over, and then proceed.
1844  */
1845 static int
1846 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1847 {
1848 	struct mount_pcpu *mpcpu;
1849 	int error, mflags;
1850 
1851 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1852 	    vfs_op_thread_enter(mp, mpcpu)) {
1853 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1854 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1855 		vfs_op_thread_exit(mp, mpcpu);
1856 		return (0);
1857 	}
1858 
1859 	if (mplocked)
1860 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1861 	else
1862 		MNT_ILOCK(mp);
1863 
1864 	error = 0;
1865 
1866 	/*
1867 	 * Check on status of suspension.
1868 	 */
1869 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1870 	    mp->mnt_susp_owner != curthread) {
1871 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1872 		    (flags & PCATCH) : 0) | (PUSER - 1);
1873 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1874 			if (flags & V_NOWAIT) {
1875 				error = EWOULDBLOCK;
1876 				goto unlock;
1877 			}
1878 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1879 			    "suspfs", 0);
1880 			if (error)
1881 				goto unlock;
1882 		}
1883 	}
1884 	if (flags & V_XSLEEP)
1885 		goto unlock;
1886 	mp->mnt_writeopcount++;
1887 unlock:
1888 	if (error != 0 || (flags & V_XSLEEP) != 0)
1889 		MNT_REL(mp);
1890 	MNT_IUNLOCK(mp);
1891 	return (error);
1892 }
1893 
1894 int
1895 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1896 {
1897 	struct mount *mp;
1898 	int error;
1899 
1900 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1901 	    ("V_MNTREF requires mp"));
1902 
1903 	error = 0;
1904 	/*
1905 	 * If a vnode is provided, get and return the mount point that
1906 	 * to which it will write.
1907 	 */
1908 	if (vp != NULL) {
1909 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1910 			*mpp = NULL;
1911 			if (error != EOPNOTSUPP)
1912 				return (error);
1913 			return (0);
1914 		}
1915 	}
1916 	if ((mp = *mpp) == NULL)
1917 		return (0);
1918 
1919 	/*
1920 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1921 	 * a vfs_ref().
1922 	 * As long as a vnode is not provided we need to acquire a
1923 	 * refcount for the provided mountpoint too, in order to
1924 	 * emulate a vfs_ref().
1925 	 */
1926 	if (vp == NULL && (flags & V_MNTREF) == 0)
1927 		vfs_ref(mp);
1928 
1929 	return (vn_start_write_refed(mp, flags, false));
1930 }
1931 
1932 /*
1933  * Secondary suspension. Used by operations such as vop_inactive
1934  * routines that are needed by the higher level functions. These
1935  * are allowed to proceed until all the higher level functions have
1936  * completed (indicated by mnt_writeopcount dropping to zero). At that
1937  * time, these operations are halted until the suspension is over.
1938  */
1939 int
1940 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1941 {
1942 	struct mount *mp;
1943 	int error;
1944 
1945 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1946 	    ("V_MNTREF requires mp"));
1947 
1948  retry:
1949 	if (vp != NULL) {
1950 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1951 			*mpp = NULL;
1952 			if (error != EOPNOTSUPP)
1953 				return (error);
1954 			return (0);
1955 		}
1956 	}
1957 	/*
1958 	 * If we are not suspended or have not yet reached suspended
1959 	 * mode, then let the operation proceed.
1960 	 */
1961 	if ((mp = *mpp) == NULL)
1962 		return (0);
1963 
1964 	/*
1965 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1966 	 * a vfs_ref().
1967 	 * As long as a vnode is not provided we need to acquire a
1968 	 * refcount for the provided mountpoint too, in order to
1969 	 * emulate a vfs_ref().
1970 	 */
1971 	MNT_ILOCK(mp);
1972 	if (vp == NULL && (flags & V_MNTREF) == 0)
1973 		MNT_REF(mp);
1974 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1975 		mp->mnt_secondary_writes++;
1976 		mp->mnt_secondary_accwrites++;
1977 		MNT_IUNLOCK(mp);
1978 		return (0);
1979 	}
1980 	if (flags & V_NOWAIT) {
1981 		MNT_REL(mp);
1982 		MNT_IUNLOCK(mp);
1983 		return (EWOULDBLOCK);
1984 	}
1985 	/*
1986 	 * Wait for the suspension to finish.
1987 	 */
1988 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1989 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1990 	    "suspfs", 0);
1991 	vfs_rel(mp);
1992 	if (error == 0)
1993 		goto retry;
1994 	return (error);
1995 }
1996 
1997 /*
1998  * Filesystem write operation has completed. If we are suspending and this
1999  * operation is the last one, notify the suspender that the suspension is
2000  * now in effect.
2001  */
2002 void
2003 vn_finished_write(struct mount *mp)
2004 {
2005 	struct mount_pcpu *mpcpu;
2006 	int c;
2007 
2008 	if (mp == NULL)
2009 		return;
2010 
2011 	if (vfs_op_thread_enter(mp, mpcpu)) {
2012 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
2013 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
2014 		vfs_op_thread_exit(mp, mpcpu);
2015 		return;
2016 	}
2017 
2018 	MNT_ILOCK(mp);
2019 	vfs_assert_mount_counters(mp);
2020 	MNT_REL(mp);
2021 	c = --mp->mnt_writeopcount;
2022 	if (mp->mnt_vfs_ops == 0) {
2023 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
2024 		MNT_IUNLOCK(mp);
2025 		return;
2026 	}
2027 	if (c < 0)
2028 		vfs_dump_mount_counters(mp);
2029 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
2030 		wakeup(&mp->mnt_writeopcount);
2031 	MNT_IUNLOCK(mp);
2032 }
2033 
2034 /*
2035  * Filesystem secondary write operation has completed. If we are
2036  * suspending and this operation is the last one, notify the suspender
2037  * that the suspension is now in effect.
2038  */
2039 void
2040 vn_finished_secondary_write(struct mount *mp)
2041 {
2042 	if (mp == NULL)
2043 		return;
2044 	MNT_ILOCK(mp);
2045 	MNT_REL(mp);
2046 	mp->mnt_secondary_writes--;
2047 	if (mp->mnt_secondary_writes < 0)
2048 		panic("vn_finished_secondary_write: neg cnt");
2049 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2050 	    mp->mnt_secondary_writes <= 0)
2051 		wakeup(&mp->mnt_secondary_writes);
2052 	MNT_IUNLOCK(mp);
2053 }
2054 
2055 /*
2056  * Request a filesystem to suspend write operations.
2057  */
2058 int
2059 vfs_write_suspend(struct mount *mp, int flags)
2060 {
2061 	int error;
2062 
2063 	vfs_op_enter(mp);
2064 
2065 	MNT_ILOCK(mp);
2066 	vfs_assert_mount_counters(mp);
2067 	if (mp->mnt_susp_owner == curthread) {
2068 		vfs_op_exit_locked(mp);
2069 		MNT_IUNLOCK(mp);
2070 		return (EALREADY);
2071 	}
2072 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2073 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2074 
2075 	/*
2076 	 * Unmount holds a write reference on the mount point.  If we
2077 	 * own busy reference and drain for writers, we deadlock with
2078 	 * the reference draining in the unmount path.  Callers of
2079 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2080 	 * vfs_busy() reference is owned and caller is not in the
2081 	 * unmount context.
2082 	 */
2083 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2084 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2085 		vfs_op_exit_locked(mp);
2086 		MNT_IUNLOCK(mp);
2087 		return (EBUSY);
2088 	}
2089 
2090 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2091 	mp->mnt_susp_owner = curthread;
2092 	if (mp->mnt_writeopcount > 0)
2093 		(void) msleep(&mp->mnt_writeopcount,
2094 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2095 	else
2096 		MNT_IUNLOCK(mp);
2097 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2098 		vfs_write_resume(mp, 0);
2099 		/* vfs_write_resume does vfs_op_exit() for us */
2100 	}
2101 	return (error);
2102 }
2103 
2104 /*
2105  * Request a filesystem to resume write operations.
2106  */
2107 void
2108 vfs_write_resume(struct mount *mp, int flags)
2109 {
2110 
2111 	MNT_ILOCK(mp);
2112 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2113 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2114 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2115 				       MNTK_SUSPENDED);
2116 		mp->mnt_susp_owner = NULL;
2117 		wakeup(&mp->mnt_writeopcount);
2118 		wakeup(&mp->mnt_flag);
2119 		curthread->td_pflags &= ~TDP_IGNSUSP;
2120 		if ((flags & VR_START_WRITE) != 0) {
2121 			MNT_REF(mp);
2122 			mp->mnt_writeopcount++;
2123 		}
2124 		MNT_IUNLOCK(mp);
2125 		if ((flags & VR_NO_SUSPCLR) == 0)
2126 			VFS_SUSP_CLEAN(mp);
2127 		vfs_op_exit(mp);
2128 	} else if ((flags & VR_START_WRITE) != 0) {
2129 		MNT_REF(mp);
2130 		vn_start_write_refed(mp, 0, true);
2131 	} else {
2132 		MNT_IUNLOCK(mp);
2133 	}
2134 }
2135 
2136 /*
2137  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2138  * methods.
2139  */
2140 int
2141 vfs_write_suspend_umnt(struct mount *mp)
2142 {
2143 	int error;
2144 
2145 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2146 	    ("vfs_write_suspend_umnt: recursed"));
2147 
2148 	/* dounmount() already called vn_start_write(). */
2149 	for (;;) {
2150 		vn_finished_write(mp);
2151 		error = vfs_write_suspend(mp, 0);
2152 		if (error != 0) {
2153 			vn_start_write(NULL, &mp, V_WAIT);
2154 			return (error);
2155 		}
2156 		MNT_ILOCK(mp);
2157 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2158 			break;
2159 		MNT_IUNLOCK(mp);
2160 		vn_start_write(NULL, &mp, V_WAIT);
2161 	}
2162 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2163 	wakeup(&mp->mnt_flag);
2164 	MNT_IUNLOCK(mp);
2165 	curthread->td_pflags |= TDP_IGNSUSP;
2166 	return (0);
2167 }
2168 
2169 /*
2170  * Implement kqueues for files by translating it to vnode operation.
2171  */
2172 static int
2173 vn_kqfilter(struct file *fp, struct knote *kn)
2174 {
2175 
2176 	return (VOP_KQFILTER(fp->f_vnode, kn));
2177 }
2178 
2179 int
2180 vn_kqfilter_opath(struct file *fp, struct knote *kn)
2181 {
2182 	if ((fp->f_flag & FKQALLOWED) == 0)
2183 		return (EBADF);
2184 	return (vn_kqfilter(fp, kn));
2185 }
2186 
2187 /*
2188  * Simplified in-kernel wrapper calls for extended attribute access.
2189  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2190  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2191  */
2192 int
2193 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2194     const char *attrname, int *buflen, char *buf, struct thread *td)
2195 {
2196 	struct uio	auio;
2197 	struct iovec	iov;
2198 	int	error;
2199 
2200 	iov.iov_len = *buflen;
2201 	iov.iov_base = buf;
2202 
2203 	auio.uio_iov = &iov;
2204 	auio.uio_iovcnt = 1;
2205 	auio.uio_rw = UIO_READ;
2206 	auio.uio_segflg = UIO_SYSSPACE;
2207 	auio.uio_td = td;
2208 	auio.uio_offset = 0;
2209 	auio.uio_resid = *buflen;
2210 
2211 	if ((ioflg & IO_NODELOCKED) == 0)
2212 		vn_lock(vp, LK_SHARED | LK_RETRY);
2213 
2214 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2215 
2216 	/* authorize attribute retrieval as kernel */
2217 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2218 	    td);
2219 
2220 	if ((ioflg & IO_NODELOCKED) == 0)
2221 		VOP_UNLOCK(vp);
2222 
2223 	if (error == 0) {
2224 		*buflen = *buflen - auio.uio_resid;
2225 	}
2226 
2227 	return (error);
2228 }
2229 
2230 /*
2231  * XXX failure mode if partially written?
2232  */
2233 int
2234 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2235     const char *attrname, int buflen, char *buf, struct thread *td)
2236 {
2237 	struct uio	auio;
2238 	struct iovec	iov;
2239 	struct mount	*mp;
2240 	int	error;
2241 
2242 	iov.iov_len = buflen;
2243 	iov.iov_base = buf;
2244 
2245 	auio.uio_iov = &iov;
2246 	auio.uio_iovcnt = 1;
2247 	auio.uio_rw = UIO_WRITE;
2248 	auio.uio_segflg = UIO_SYSSPACE;
2249 	auio.uio_td = td;
2250 	auio.uio_offset = 0;
2251 	auio.uio_resid = buflen;
2252 
2253 	if ((ioflg & IO_NODELOCKED) == 0) {
2254 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2255 			return (error);
2256 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2257 	}
2258 
2259 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2260 
2261 	/* authorize attribute setting as kernel */
2262 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2263 
2264 	if ((ioflg & IO_NODELOCKED) == 0) {
2265 		vn_finished_write(mp);
2266 		VOP_UNLOCK(vp);
2267 	}
2268 
2269 	return (error);
2270 }
2271 
2272 int
2273 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2274     const char *attrname, struct thread *td)
2275 {
2276 	struct mount	*mp;
2277 	int	error;
2278 
2279 	if ((ioflg & IO_NODELOCKED) == 0) {
2280 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2281 			return (error);
2282 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2283 	}
2284 
2285 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2286 
2287 	/* authorize attribute removal as kernel */
2288 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2289 	if (error == EOPNOTSUPP)
2290 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2291 		    NULL, td);
2292 
2293 	if ((ioflg & IO_NODELOCKED) == 0) {
2294 		vn_finished_write(mp);
2295 		VOP_UNLOCK(vp);
2296 	}
2297 
2298 	return (error);
2299 }
2300 
2301 static int
2302 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2303     struct vnode **rvp)
2304 {
2305 
2306 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2307 }
2308 
2309 int
2310 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2311 {
2312 
2313 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2314 	    lkflags, rvp));
2315 }
2316 
2317 int
2318 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2319     int lkflags, struct vnode **rvp)
2320 {
2321 	struct mount *mp;
2322 	int ltype, error;
2323 
2324 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2325 	mp = vp->v_mount;
2326 	ltype = VOP_ISLOCKED(vp);
2327 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2328 	    ("vn_vget_ino: vp not locked"));
2329 	error = vfs_busy(mp, MBF_NOWAIT);
2330 	if (error != 0) {
2331 		vfs_ref(mp);
2332 		VOP_UNLOCK(vp);
2333 		error = vfs_busy(mp, 0);
2334 		vn_lock(vp, ltype | LK_RETRY);
2335 		vfs_rel(mp);
2336 		if (error != 0)
2337 			return (ENOENT);
2338 		if (VN_IS_DOOMED(vp)) {
2339 			vfs_unbusy(mp);
2340 			return (ENOENT);
2341 		}
2342 	}
2343 	VOP_UNLOCK(vp);
2344 	error = alloc(mp, alloc_arg, lkflags, rvp);
2345 	vfs_unbusy(mp);
2346 	if (error != 0 || *rvp != vp)
2347 		vn_lock(vp, ltype | LK_RETRY);
2348 	if (VN_IS_DOOMED(vp)) {
2349 		if (error == 0) {
2350 			if (*rvp == vp)
2351 				vunref(vp);
2352 			else
2353 				vput(*rvp);
2354 		}
2355 		error = ENOENT;
2356 	}
2357 	return (error);
2358 }
2359 
2360 int
2361 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2362     struct thread *td)
2363 {
2364 	off_t lim;
2365 	bool ktr_write;
2366 
2367 	if (td == NULL)
2368 		return (0);
2369 
2370 	/*
2371 	 * There are conditions where the limit is to be ignored.
2372 	 * However, since it is almost never reached, check it first.
2373 	 */
2374 	ktr_write = (td->td_pflags & TDP_INKTRACE) != 0;
2375 	lim = lim_cur(td, RLIMIT_FSIZE);
2376 	if (__predict_false(ktr_write))
2377 		lim = td->td_ktr_io_lim;
2378 	if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim))
2379 		return (0);
2380 
2381 	/*
2382 	 * The limit is reached.
2383 	 */
2384 	if (vp->v_type != VREG ||
2385 	    (td->td_pflags2 & TDP2_ACCT) != 0)
2386 		return (0);
2387 
2388 	if (!ktr_write || ktr_filesize_limit_signal) {
2389 		PROC_LOCK(td->td_proc);
2390 		kern_psignal(td->td_proc, SIGXFSZ);
2391 		PROC_UNLOCK(td->td_proc);
2392 	}
2393 	return (EFBIG);
2394 }
2395 
2396 int
2397 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2398     struct thread *td)
2399 {
2400 	struct vnode *vp;
2401 
2402 	vp = fp->f_vnode;
2403 #ifdef AUDIT
2404 	vn_lock(vp, LK_SHARED | LK_RETRY);
2405 	AUDIT_ARG_VNODE1(vp);
2406 	VOP_UNLOCK(vp);
2407 #endif
2408 	return (setfmode(td, active_cred, vp, mode));
2409 }
2410 
2411 int
2412 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2413     struct thread *td)
2414 {
2415 	struct vnode *vp;
2416 
2417 	vp = fp->f_vnode;
2418 #ifdef AUDIT
2419 	vn_lock(vp, LK_SHARED | LK_RETRY);
2420 	AUDIT_ARG_VNODE1(vp);
2421 	VOP_UNLOCK(vp);
2422 #endif
2423 	return (setfown(td, active_cred, vp, uid, gid));
2424 }
2425 
2426 void
2427 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2428 {
2429 	vm_object_t object;
2430 
2431 	if ((object = vp->v_object) == NULL)
2432 		return;
2433 	VM_OBJECT_WLOCK(object);
2434 	vm_object_page_remove(object, start, end, 0);
2435 	VM_OBJECT_WUNLOCK(object);
2436 }
2437 
2438 int
2439 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2440 {
2441 	struct vattr va;
2442 	daddr_t bn, bnp;
2443 	uint64_t bsize;
2444 	off_t noff;
2445 	int error;
2446 
2447 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2448 	    ("Wrong command %lu", cmd));
2449 
2450 	if (vn_lock(vp, LK_SHARED) != 0)
2451 		return (EBADF);
2452 	if (vp->v_type != VREG) {
2453 		error = ENOTTY;
2454 		goto unlock;
2455 	}
2456 	error = VOP_GETATTR(vp, &va, cred);
2457 	if (error != 0)
2458 		goto unlock;
2459 	noff = *off;
2460 	if (noff >= va.va_size) {
2461 		error = ENXIO;
2462 		goto unlock;
2463 	}
2464 	bsize = vp->v_mount->mnt_stat.f_iosize;
2465 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2466 	    noff % bsize) {
2467 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2468 		if (error == EOPNOTSUPP) {
2469 			error = ENOTTY;
2470 			goto unlock;
2471 		}
2472 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2473 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2474 			noff = bn * bsize;
2475 			if (noff < *off)
2476 				noff = *off;
2477 			goto unlock;
2478 		}
2479 	}
2480 	if (noff > va.va_size)
2481 		noff = va.va_size;
2482 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2483 	if (cmd == FIOSEEKDATA)
2484 		error = ENXIO;
2485 unlock:
2486 	VOP_UNLOCK(vp);
2487 	if (error == 0)
2488 		*off = noff;
2489 	return (error);
2490 }
2491 
2492 int
2493 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2494 {
2495 	struct ucred *cred;
2496 	struct vnode *vp;
2497 	struct vattr vattr;
2498 	off_t foffset, size;
2499 	int error, noneg;
2500 
2501 	cred = td->td_ucred;
2502 	vp = fp->f_vnode;
2503 	foffset = foffset_lock(fp, 0);
2504 	noneg = (vp->v_type != VCHR);
2505 	error = 0;
2506 	switch (whence) {
2507 	case L_INCR:
2508 		if (noneg &&
2509 		    (foffset < 0 ||
2510 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2511 			error = EOVERFLOW;
2512 			break;
2513 		}
2514 		offset += foffset;
2515 		break;
2516 	case L_XTND:
2517 		vn_lock(vp, LK_SHARED | LK_RETRY);
2518 		error = VOP_GETATTR(vp, &vattr, cred);
2519 		VOP_UNLOCK(vp);
2520 		if (error)
2521 			break;
2522 
2523 		/*
2524 		 * If the file references a disk device, then fetch
2525 		 * the media size and use that to determine the ending
2526 		 * offset.
2527 		 */
2528 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2529 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2530 			vattr.va_size = size;
2531 		if (noneg &&
2532 		    (vattr.va_size > OFF_MAX ||
2533 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2534 			error = EOVERFLOW;
2535 			break;
2536 		}
2537 		offset += vattr.va_size;
2538 		break;
2539 	case L_SET:
2540 		break;
2541 	case SEEK_DATA:
2542 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2543 		if (error == ENOTTY)
2544 			error = EINVAL;
2545 		break;
2546 	case SEEK_HOLE:
2547 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2548 		if (error == ENOTTY)
2549 			error = EINVAL;
2550 		break;
2551 	default:
2552 		error = EINVAL;
2553 	}
2554 	if (error == 0 && noneg && offset < 0)
2555 		error = EINVAL;
2556 	if (error != 0)
2557 		goto drop;
2558 	VFS_KNOTE_UNLOCKED(vp, 0);
2559 	td->td_uretoff.tdu_off = offset;
2560 drop:
2561 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2562 	return (error);
2563 }
2564 
2565 int
2566 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2567     struct thread *td)
2568 {
2569 	int error;
2570 
2571 	/*
2572 	 * Grant permission if the caller is the owner of the file, or
2573 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2574 	 * on the file.  If the time pointer is null, then write
2575 	 * permission on the file is also sufficient.
2576 	 *
2577 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2578 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2579 	 * will be allowed to set the times [..] to the current
2580 	 * server time.
2581 	 */
2582 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2583 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2584 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2585 	return (error);
2586 }
2587 
2588 int
2589 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2590 {
2591 	struct vnode *vp;
2592 	int error;
2593 
2594 	if (fp->f_type == DTYPE_FIFO)
2595 		kif->kf_type = KF_TYPE_FIFO;
2596 	else
2597 		kif->kf_type = KF_TYPE_VNODE;
2598 	vp = fp->f_vnode;
2599 	vref(vp);
2600 	FILEDESC_SUNLOCK(fdp);
2601 	error = vn_fill_kinfo_vnode(vp, kif);
2602 	vrele(vp);
2603 	FILEDESC_SLOCK(fdp);
2604 	return (error);
2605 }
2606 
2607 static inline void
2608 vn_fill_junk(struct kinfo_file *kif)
2609 {
2610 	size_t len, olen;
2611 
2612 	/*
2613 	 * Simulate vn_fullpath returning changing values for a given
2614 	 * vp during e.g. coredump.
2615 	 */
2616 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2617 	olen = strlen(kif->kf_path);
2618 	if (len < olen)
2619 		strcpy(&kif->kf_path[len - 1], "$");
2620 	else
2621 		for (; olen < len; olen++)
2622 			strcpy(&kif->kf_path[olen], "A");
2623 }
2624 
2625 int
2626 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2627 {
2628 	struct vattr va;
2629 	char *fullpath, *freepath;
2630 	int error;
2631 
2632 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2633 	freepath = NULL;
2634 	fullpath = "-";
2635 	error = vn_fullpath(vp, &fullpath, &freepath);
2636 	if (error == 0) {
2637 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2638 	}
2639 	if (freepath != NULL)
2640 		free(freepath, M_TEMP);
2641 
2642 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2643 		vn_fill_junk(kif);
2644 	);
2645 
2646 	/*
2647 	 * Retrieve vnode attributes.
2648 	 */
2649 	va.va_fsid = VNOVAL;
2650 	va.va_rdev = NODEV;
2651 	vn_lock(vp, LK_SHARED | LK_RETRY);
2652 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2653 	VOP_UNLOCK(vp);
2654 	if (error != 0)
2655 		return (error);
2656 	if (va.va_fsid != VNOVAL)
2657 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2658 	else
2659 		kif->kf_un.kf_file.kf_file_fsid =
2660 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2661 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2662 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2663 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2664 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2665 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2666 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2667 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2668 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2669 	return (0);
2670 }
2671 
2672 int
2673 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2674     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2675     struct thread *td)
2676 {
2677 #ifdef HWPMC_HOOKS
2678 	struct pmckern_map_in pkm;
2679 #endif
2680 	struct mount *mp;
2681 	struct vnode *vp;
2682 	vm_object_t object;
2683 	vm_prot_t maxprot;
2684 	boolean_t writecounted;
2685 	int error;
2686 
2687 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2688     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2689 	/*
2690 	 * POSIX shared-memory objects are defined to have
2691 	 * kernel persistence, and are not defined to support
2692 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2693 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2694 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2695 	 * flag to request this behavior.
2696 	 */
2697 	if ((fp->f_flag & FPOSIXSHM) != 0)
2698 		flags |= MAP_NOSYNC;
2699 #endif
2700 	vp = fp->f_vnode;
2701 
2702 	/*
2703 	 * Ensure that file and memory protections are
2704 	 * compatible.  Note that we only worry about
2705 	 * writability if mapping is shared; in this case,
2706 	 * current and max prot are dictated by the open file.
2707 	 * XXX use the vnode instead?  Problem is: what
2708 	 * credentials do we use for determination? What if
2709 	 * proc does a setuid?
2710 	 */
2711 	mp = vp->v_mount;
2712 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2713 		maxprot = VM_PROT_NONE;
2714 		if ((prot & VM_PROT_EXECUTE) != 0)
2715 			return (EACCES);
2716 	} else
2717 		maxprot = VM_PROT_EXECUTE;
2718 	if ((fp->f_flag & FREAD) != 0)
2719 		maxprot |= VM_PROT_READ;
2720 	else if ((prot & VM_PROT_READ) != 0)
2721 		return (EACCES);
2722 
2723 	/*
2724 	 * If we are sharing potential changes via MAP_SHARED and we
2725 	 * are trying to get write permission although we opened it
2726 	 * without asking for it, bail out.
2727 	 */
2728 	if ((flags & MAP_SHARED) != 0) {
2729 		if ((fp->f_flag & FWRITE) != 0)
2730 			maxprot |= VM_PROT_WRITE;
2731 		else if ((prot & VM_PROT_WRITE) != 0)
2732 			return (EACCES);
2733 	} else {
2734 		maxprot |= VM_PROT_WRITE;
2735 		cap_maxprot |= VM_PROT_WRITE;
2736 	}
2737 	maxprot &= cap_maxprot;
2738 
2739 	/*
2740 	 * For regular files and shared memory, POSIX requires that
2741 	 * the value of foff be a legitimate offset within the data
2742 	 * object.  In particular, negative offsets are invalid.
2743 	 * Blocking negative offsets and overflows here avoids
2744 	 * possible wraparound or user-level access into reserved
2745 	 * ranges of the data object later.  In contrast, POSIX does
2746 	 * not dictate how offsets are used by device drivers, so in
2747 	 * the case of a device mapping a negative offset is passed
2748 	 * on.
2749 	 */
2750 	if (
2751 #ifdef _LP64
2752 	    size > OFF_MAX ||
2753 #endif
2754 	    foff > OFF_MAX - size)
2755 		return (EINVAL);
2756 
2757 	writecounted = FALSE;
2758 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2759 	    &foff, &object, &writecounted);
2760 	if (error != 0)
2761 		return (error);
2762 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2763 	    foff, writecounted, td);
2764 	if (error != 0) {
2765 		/*
2766 		 * If this mapping was accounted for in the vnode's
2767 		 * writecount, then undo that now.
2768 		 */
2769 		if (writecounted)
2770 			vm_pager_release_writecount(object, 0, size);
2771 		vm_object_deallocate(object);
2772 	}
2773 #ifdef HWPMC_HOOKS
2774 	/* Inform hwpmc(4) if an executable is being mapped. */
2775 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2776 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2777 			pkm.pm_file = vp;
2778 			pkm.pm_address = (uintptr_t) *addr;
2779 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2780 		}
2781 	}
2782 #endif
2783 	return (error);
2784 }
2785 
2786 void
2787 vn_fsid(struct vnode *vp, struct vattr *va)
2788 {
2789 	fsid_t *f;
2790 
2791 	f = &vp->v_mount->mnt_stat.f_fsid;
2792 	va->va_fsid = (uint32_t)f->val[1];
2793 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2794 	va->va_fsid += (uint32_t)f->val[0];
2795 }
2796 
2797 int
2798 vn_fsync_buf(struct vnode *vp, int waitfor)
2799 {
2800 	struct buf *bp, *nbp;
2801 	struct bufobj *bo;
2802 	struct mount *mp;
2803 	int error, maxretry;
2804 
2805 	error = 0;
2806 	maxretry = 10000;     /* large, arbitrarily chosen */
2807 	mp = NULL;
2808 	if (vp->v_type == VCHR) {
2809 		VI_LOCK(vp);
2810 		mp = vp->v_rdev->si_mountpt;
2811 		VI_UNLOCK(vp);
2812 	}
2813 	bo = &vp->v_bufobj;
2814 	BO_LOCK(bo);
2815 loop1:
2816 	/*
2817 	 * MARK/SCAN initialization to avoid infinite loops.
2818 	 */
2819         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2820 		bp->b_vflags &= ~BV_SCANNED;
2821 		bp->b_error = 0;
2822 	}
2823 
2824 	/*
2825 	 * Flush all dirty buffers associated with a vnode.
2826 	 */
2827 loop2:
2828 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2829 		if ((bp->b_vflags & BV_SCANNED) != 0)
2830 			continue;
2831 		bp->b_vflags |= BV_SCANNED;
2832 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2833 			if (waitfor != MNT_WAIT)
2834 				continue;
2835 			if (BUF_LOCK(bp,
2836 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2837 			    BO_LOCKPTR(bo)) != 0) {
2838 				BO_LOCK(bo);
2839 				goto loop1;
2840 			}
2841 			BO_LOCK(bo);
2842 		}
2843 		BO_UNLOCK(bo);
2844 		KASSERT(bp->b_bufobj == bo,
2845 		    ("bp %p wrong b_bufobj %p should be %p",
2846 		    bp, bp->b_bufobj, bo));
2847 		if ((bp->b_flags & B_DELWRI) == 0)
2848 			panic("fsync: not dirty");
2849 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2850 			vfs_bio_awrite(bp);
2851 		} else {
2852 			bremfree(bp);
2853 			bawrite(bp);
2854 		}
2855 		if (maxretry < 1000)
2856 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2857 		BO_LOCK(bo);
2858 		goto loop2;
2859 	}
2860 
2861 	/*
2862 	 * If synchronous the caller expects us to completely resolve all
2863 	 * dirty buffers in the system.  Wait for in-progress I/O to
2864 	 * complete (which could include background bitmap writes), then
2865 	 * retry if dirty blocks still exist.
2866 	 */
2867 	if (waitfor == MNT_WAIT) {
2868 		bufobj_wwait(bo, 0, 0);
2869 		if (bo->bo_dirty.bv_cnt > 0) {
2870 			/*
2871 			 * If we are unable to write any of these buffers
2872 			 * then we fail now rather than trying endlessly
2873 			 * to write them out.
2874 			 */
2875 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2876 				if ((error = bp->b_error) != 0)
2877 					break;
2878 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2879 			    (error == 0 && --maxretry >= 0))
2880 				goto loop1;
2881 			if (error == 0)
2882 				error = EAGAIN;
2883 		}
2884 	}
2885 	BO_UNLOCK(bo);
2886 	if (error != 0)
2887 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2888 
2889 	return (error);
2890 }
2891 
2892 /*
2893  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2894  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2895  * to do the actual copy.
2896  * vn_generic_copy_file_range() is factored out, so it can be called
2897  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2898  * different file systems.
2899  */
2900 int
2901 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2902     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2903     struct ucred *outcred, struct thread *fsize_td)
2904 {
2905 	int error;
2906 	size_t len;
2907 	uint64_t uval;
2908 
2909 	len = *lenp;
2910 	*lenp = 0;		/* For error returns. */
2911 	error = 0;
2912 
2913 	/* Do some sanity checks on the arguments. */
2914 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2915 		error = EISDIR;
2916 	else if (*inoffp < 0 || *outoffp < 0 ||
2917 	    invp->v_type != VREG || outvp->v_type != VREG)
2918 		error = EINVAL;
2919 	if (error != 0)
2920 		goto out;
2921 
2922 	/* Ensure offset + len does not wrap around. */
2923 	uval = *inoffp;
2924 	uval += len;
2925 	if (uval > INT64_MAX)
2926 		len = INT64_MAX - *inoffp;
2927 	uval = *outoffp;
2928 	uval += len;
2929 	if (uval > INT64_MAX)
2930 		len = INT64_MAX - *outoffp;
2931 	if (len == 0)
2932 		goto out;
2933 
2934 	/*
2935 	 * If the two vnode are for the same file system, call
2936 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2937 	 * which can handle copies across multiple file systems.
2938 	 */
2939 	*lenp = len;
2940 	if (invp->v_mount == outvp->v_mount)
2941 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2942 		    lenp, flags, incred, outcred, fsize_td);
2943 	else
2944 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2945 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2946 out:
2947 	return (error);
2948 }
2949 
2950 /*
2951  * Test len bytes of data starting at dat for all bytes == 0.
2952  * Return true if all bytes are zero, false otherwise.
2953  * Expects dat to be well aligned.
2954  */
2955 static bool
2956 mem_iszero(void *dat, int len)
2957 {
2958 	int i;
2959 	const u_int *p;
2960 	const char *cp;
2961 
2962 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2963 		if (len >= sizeof(*p)) {
2964 			if (*p != 0)
2965 				return (false);
2966 		} else {
2967 			cp = (const char *)p;
2968 			for (i = 0; i < len; i++, cp++)
2969 				if (*cp != '\0')
2970 					return (false);
2971 		}
2972 	}
2973 	return (true);
2974 }
2975 
2976 /*
2977  * Look for a hole in the output file and, if found, adjust *outoffp
2978  * and *xferp to skip past the hole.
2979  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2980  * to be written as 0's upon return.
2981  */
2982 static off_t
2983 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2984     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2985 {
2986 	int error;
2987 	off_t delta;
2988 
2989 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2990 		*dataoffp = *outoffp;
2991 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2992 		    curthread);
2993 		if (error == 0) {
2994 			*holeoffp = *dataoffp;
2995 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2996 			    curthread);
2997 		}
2998 		if (error != 0 || *holeoffp == *dataoffp) {
2999 			/*
3000 			 * Since outvp is unlocked, it may be possible for
3001 			 * another thread to do a truncate(), lseek(), write()
3002 			 * creating a hole at startoff between the above
3003 			 * VOP_IOCTL() calls, if the other thread does not do
3004 			 * rangelocking.
3005 			 * If that happens, *holeoffp == *dataoffp and finding
3006 			 * the hole has failed, so disable vn_skip_hole().
3007 			 */
3008 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
3009 			return (xfer2);
3010 		}
3011 		KASSERT(*dataoffp >= *outoffp,
3012 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
3013 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
3014 		KASSERT(*holeoffp > *dataoffp,
3015 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
3016 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
3017 	}
3018 
3019 	/*
3020 	 * If there is a hole before the data starts, advance *outoffp and
3021 	 * *xferp past the hole.
3022 	 */
3023 	if (*dataoffp > *outoffp) {
3024 		delta = *dataoffp - *outoffp;
3025 		if (delta >= *xferp) {
3026 			/* Entire *xferp is a hole. */
3027 			*outoffp += *xferp;
3028 			*xferp = 0;
3029 			return (0);
3030 		}
3031 		*xferp -= delta;
3032 		*outoffp += delta;
3033 		xfer2 = MIN(xfer2, *xferp);
3034 	}
3035 
3036 	/*
3037 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
3038 	 * that the write ends at the start of the hole.
3039 	 * *holeoffp should always be greater than *outoffp, but for the
3040 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
3041 	 * value.
3042 	 */
3043 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
3044 		xfer2 = *holeoffp - *outoffp;
3045 	return (xfer2);
3046 }
3047 
3048 /*
3049  * Write an xfer sized chunk to outvp in blksize blocks from dat.
3050  * dat is a maximum of blksize in length and can be written repeatedly in
3051  * the chunk.
3052  * If growfile == true, just grow the file via vn_truncate_locked() instead
3053  * of doing actual writes.
3054  * If checkhole == true, a hole is being punched, so skip over any hole
3055  * already in the output file.
3056  */
3057 static int
3058 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
3059     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
3060 {
3061 	struct mount *mp;
3062 	off_t dataoff, holeoff, xfer2;
3063 	int error, lckf;
3064 
3065 	/*
3066 	 * Loop around doing writes of blksize until write has been completed.
3067 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
3068 	 * done for each iteration, since the xfer argument can be very
3069 	 * large if there is a large hole to punch in the output file.
3070 	 */
3071 	error = 0;
3072 	holeoff = 0;
3073 	do {
3074 		xfer2 = MIN(xfer, blksize);
3075 		if (checkhole) {
3076 			/*
3077 			 * Punching a hole.  Skip writing if there is
3078 			 * already a hole in the output file.
3079 			 */
3080 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3081 			    &dataoff, &holeoff, cred);
3082 			if (xfer == 0)
3083 				break;
3084 			if (holeoff < 0)
3085 				checkhole = false;
3086 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3087 			    (intmax_t)xfer2));
3088 		}
3089 		bwillwrite();
3090 		mp = NULL;
3091 		error = vn_start_write(outvp, &mp, V_WAIT);
3092 		if (error != 0)
3093 			break;
3094 		if (growfile) {
3095 			error = vn_lock(outvp, LK_EXCLUSIVE);
3096 			if (error == 0) {
3097 				error = vn_truncate_locked(outvp, outoff + xfer,
3098 				    false, cred);
3099 				VOP_UNLOCK(outvp);
3100 			}
3101 		} else {
3102 			if (MNT_SHARED_WRITES(mp))
3103 				lckf = LK_SHARED;
3104 			else
3105 				lckf = LK_EXCLUSIVE;
3106 			error = vn_lock(outvp, lckf);
3107 			if (error == 0) {
3108 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3109 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3110 				    curthread->td_ucred, cred, NULL, curthread);
3111 				outoff += xfer2;
3112 				xfer -= xfer2;
3113 				VOP_UNLOCK(outvp);
3114 			}
3115 		}
3116 		if (mp != NULL)
3117 			vn_finished_write(mp);
3118 	} while (!growfile && xfer > 0 && error == 0);
3119 	return (error);
3120 }
3121 
3122 /*
3123  * Copy a byte range of one file to another.  This function can handle the
3124  * case where invp and outvp are on different file systems.
3125  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3126  * is no better file system specific way to do it.
3127  */
3128 int
3129 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3130     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3131     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3132 {
3133 	struct vattr va, inva;
3134 	struct mount *mp;
3135 	struct uio io;
3136 	off_t startoff, endoff, xfer, xfer2;
3137 	u_long blksize;
3138 	int error, interrupted;
3139 	bool cantseek, readzeros, eof, lastblock, holetoeof;
3140 	ssize_t aresid;
3141 	size_t copylen, len, rem, savlen;
3142 	char *dat;
3143 	long holein, holeout;
3144 
3145 	holein = holeout = 0;
3146 	savlen = len = *lenp;
3147 	error = 0;
3148 	interrupted = 0;
3149 	dat = NULL;
3150 
3151 	error = vn_lock(invp, LK_SHARED);
3152 	if (error != 0)
3153 		goto out;
3154 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3155 		holein = 0;
3156 	if (holein > 0)
3157 		error = VOP_GETATTR(invp, &inva, incred);
3158 	VOP_UNLOCK(invp);
3159 	if (error != 0)
3160 		goto out;
3161 
3162 	mp = NULL;
3163 	error = vn_start_write(outvp, &mp, V_WAIT);
3164 	if (error == 0)
3165 		error = vn_lock(outvp, LK_EXCLUSIVE);
3166 	if (error == 0) {
3167 		/*
3168 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3169 		 * now that outvp is locked.
3170 		 */
3171 		if (fsize_td != NULL) {
3172 			io.uio_offset = *outoffp;
3173 			io.uio_resid = len;
3174 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3175 			if (error != 0)
3176 				error = EFBIG;
3177 		}
3178 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3179 			holeout = 0;
3180 		/*
3181 		 * Holes that are past EOF do not need to be written as a block
3182 		 * of zero bytes.  So, truncate the output file as far as
3183 		 * possible and then use va.va_size to decide if writing 0
3184 		 * bytes is necessary in the loop below.
3185 		 */
3186 		if (error == 0)
3187 			error = VOP_GETATTR(outvp, &va, outcred);
3188 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3189 		    *outoffp + len) {
3190 #ifdef MAC
3191 			error = mac_vnode_check_write(curthread->td_ucred,
3192 			    outcred, outvp);
3193 			if (error == 0)
3194 #endif
3195 				error = vn_truncate_locked(outvp, *outoffp,
3196 				    false, outcred);
3197 			if (error == 0)
3198 				va.va_size = *outoffp;
3199 		}
3200 		VOP_UNLOCK(outvp);
3201 	}
3202 	if (mp != NULL)
3203 		vn_finished_write(mp);
3204 	if (error != 0)
3205 		goto out;
3206 
3207 	/*
3208 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3209 	 * If hole sizes aren't available, set the blksize to the larger
3210 	 * f_iosize of invp and outvp.
3211 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3212 	 * This value is clipped at 4Kbytes and 1Mbyte.
3213 	 */
3214 	blksize = MAX(holein, holeout);
3215 
3216 	/* Clip len to end at an exact multiple of hole size. */
3217 	if (blksize > 1) {
3218 		rem = *inoffp % blksize;
3219 		if (rem > 0)
3220 			rem = blksize - rem;
3221 		if (len > rem && len - rem > blksize)
3222 			len = savlen = rounddown(len - rem, blksize) + rem;
3223 	}
3224 
3225 	if (blksize <= 1)
3226 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3227 		    outvp->v_mount->mnt_stat.f_iosize);
3228 	if (blksize < 4096)
3229 		blksize = 4096;
3230 	else if (blksize > 1024 * 1024)
3231 		blksize = 1024 * 1024;
3232 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3233 
3234 	/*
3235 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3236 	 * to find holes.  Otherwise, just scan the read block for all 0s
3237 	 * in the inner loop where the data copying is done.
3238 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3239 	 * support holes on the server, but do not support FIOSEEKHOLE.
3240 	 */
3241 	holetoeof = eof = false;
3242 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3243 		endoff = 0;			/* To shut up compilers. */
3244 		cantseek = true;
3245 		startoff = *inoffp;
3246 		copylen = len;
3247 
3248 		/*
3249 		 * Find the next data area.  If there is just a hole to EOF,
3250 		 * FIOSEEKDATA should fail with ENXIO.
3251 		 * (I do not know if any file system will report a hole to
3252 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3253 		 *  will fail for those file systems.)
3254 		 *
3255 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3256 		 * the code just falls through to the inner copy loop.
3257 		 */
3258 		error = EINVAL;
3259 		if (holein > 0) {
3260 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3261 			    incred, curthread);
3262 			if (error == ENXIO) {
3263 				startoff = endoff = inva.va_size;
3264 				eof = holetoeof = true;
3265 				error = 0;
3266 			}
3267 		}
3268 		if (error == 0 && !holetoeof) {
3269 			endoff = startoff;
3270 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3271 			    incred, curthread);
3272 			/*
3273 			 * Since invp is unlocked, it may be possible for
3274 			 * another thread to do a truncate(), lseek(), write()
3275 			 * creating a hole at startoff between the above
3276 			 * VOP_IOCTL() calls, if the other thread does not do
3277 			 * rangelocking.
3278 			 * If that happens, startoff == endoff and finding
3279 			 * the hole has failed, so set an error.
3280 			 */
3281 			if (error == 0 && startoff == endoff)
3282 				error = EINVAL; /* Any error. Reset to 0. */
3283 		}
3284 		if (error == 0) {
3285 			if (startoff > *inoffp) {
3286 				/* Found hole before data block. */
3287 				xfer = MIN(startoff - *inoffp, len);
3288 				if (*outoffp < va.va_size) {
3289 					/* Must write 0s to punch hole. */
3290 					xfer2 = MIN(va.va_size - *outoffp,
3291 					    xfer);
3292 					memset(dat, 0, MIN(xfer2, blksize));
3293 					error = vn_write_outvp(outvp, dat,
3294 					    *outoffp, xfer2, blksize, false,
3295 					    holeout > 0, outcred);
3296 				}
3297 
3298 				if (error == 0 && *outoffp + xfer >
3299 				    va.va_size && (xfer == len || holetoeof)) {
3300 					/* Grow output file (hole at end). */
3301 					error = vn_write_outvp(outvp, dat,
3302 					    *outoffp, xfer, blksize, true,
3303 					    false, outcred);
3304 				}
3305 				if (error == 0) {
3306 					*inoffp += xfer;
3307 					*outoffp += xfer;
3308 					len -= xfer;
3309 					if (len < savlen)
3310 						interrupted = sig_intr();
3311 				}
3312 			}
3313 			copylen = MIN(len, endoff - startoff);
3314 			cantseek = false;
3315 		} else {
3316 			cantseek = true;
3317 			startoff = *inoffp;
3318 			copylen = len;
3319 			error = 0;
3320 		}
3321 
3322 		xfer = blksize;
3323 		if (cantseek) {
3324 			/*
3325 			 * Set first xfer to end at a block boundary, so that
3326 			 * holes are more likely detected in the loop below via
3327 			 * the for all bytes 0 method.
3328 			 */
3329 			xfer -= (*inoffp % blksize);
3330 		}
3331 		/* Loop copying the data block. */
3332 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3333 			if (copylen < xfer)
3334 				xfer = copylen;
3335 			error = vn_lock(invp, LK_SHARED);
3336 			if (error != 0)
3337 				goto out;
3338 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3339 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3340 			    curthread->td_ucred, incred, &aresid,
3341 			    curthread);
3342 			VOP_UNLOCK(invp);
3343 			lastblock = false;
3344 			if (error == 0 && aresid > 0) {
3345 				/* Stop the copy at EOF on the input file. */
3346 				xfer -= aresid;
3347 				eof = true;
3348 				lastblock = true;
3349 			}
3350 			if (error == 0) {
3351 				/*
3352 				 * Skip the write for holes past the initial EOF
3353 				 * of the output file, unless this is the last
3354 				 * write of the output file at EOF.
3355 				 */
3356 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3357 				    false;
3358 				if (xfer == len)
3359 					lastblock = true;
3360 				if (!cantseek || *outoffp < va.va_size ||
3361 				    lastblock || !readzeros)
3362 					error = vn_write_outvp(outvp, dat,
3363 					    *outoffp, xfer, blksize,
3364 					    readzeros && lastblock &&
3365 					    *outoffp >= va.va_size, false,
3366 					    outcred);
3367 				if (error == 0) {
3368 					*inoffp += xfer;
3369 					startoff += xfer;
3370 					*outoffp += xfer;
3371 					copylen -= xfer;
3372 					len -= xfer;
3373 					if (len < savlen)
3374 						interrupted = sig_intr();
3375 				}
3376 			}
3377 			xfer = blksize;
3378 		}
3379 	}
3380 out:
3381 	*lenp = savlen - len;
3382 	free(dat, M_TEMP);
3383 	return (error);
3384 }
3385 
3386 static int
3387 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3388 {
3389 	struct mount *mp;
3390 	struct vnode *vp;
3391 	off_t olen, ooffset;
3392 	int error;
3393 #ifdef AUDIT
3394 	int audited_vnode1 = 0;
3395 #endif
3396 
3397 	vp = fp->f_vnode;
3398 	if (vp->v_type != VREG)
3399 		return (ENODEV);
3400 
3401 	/* Allocating blocks may take a long time, so iterate. */
3402 	for (;;) {
3403 		olen = len;
3404 		ooffset = offset;
3405 
3406 		bwillwrite();
3407 		mp = NULL;
3408 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3409 		if (error != 0)
3410 			break;
3411 		error = vn_lock(vp, LK_EXCLUSIVE);
3412 		if (error != 0) {
3413 			vn_finished_write(mp);
3414 			break;
3415 		}
3416 #ifdef AUDIT
3417 		if (!audited_vnode1) {
3418 			AUDIT_ARG_VNODE1(vp);
3419 			audited_vnode1 = 1;
3420 		}
3421 #endif
3422 #ifdef MAC
3423 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3424 		if (error == 0)
3425 #endif
3426 			error = VOP_ALLOCATE(vp, &offset, &len);
3427 		VOP_UNLOCK(vp);
3428 		vn_finished_write(mp);
3429 
3430 		if (olen + ooffset != offset + len) {
3431 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3432 			    ooffset, olen, offset, len);
3433 		}
3434 		if (error != 0 || len == 0)
3435 			break;
3436 		KASSERT(olen > len, ("Iteration did not make progress?"));
3437 		maybe_yield();
3438 	}
3439 
3440 	return (error);
3441 }
3442 
3443 static u_long vn_lock_pair_pause_cnt;
3444 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3445     &vn_lock_pair_pause_cnt, 0,
3446     "Count of vn_lock_pair deadlocks");
3447 
3448 u_int vn_lock_pair_pause_max;
3449 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3450     &vn_lock_pair_pause_max, 0,
3451     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3452 
3453 static void
3454 vn_lock_pair_pause(const char *wmesg)
3455 {
3456 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3457 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3458 }
3459 
3460 /*
3461  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3462  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3463  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3464  * can be NULL.
3465  *
3466  * The function returns with both vnodes exclusively locked, and
3467  * guarantees that it does not create lock order reversal with other
3468  * threads during its execution.  Both vnodes could be unlocked
3469  * temporary (and reclaimed).
3470  */
3471 void
3472 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3473     bool vp2_locked)
3474 {
3475 	int error;
3476 
3477 	if (vp1 == NULL && vp2 == NULL)
3478 		return;
3479 	if (vp1 != NULL) {
3480 		if (vp1_locked)
3481 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3482 		else
3483 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3484 	} else {
3485 		vp1_locked = true;
3486 	}
3487 	if (vp2 != NULL) {
3488 		if (vp2_locked)
3489 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3490 		else
3491 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3492 	} else {
3493 		vp2_locked = true;
3494 	}
3495 	if (!vp1_locked && !vp2_locked) {
3496 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3497 		vp1_locked = true;
3498 	}
3499 
3500 	for (;;) {
3501 		if (vp1_locked && vp2_locked)
3502 			break;
3503 		if (vp1_locked && vp2 != NULL) {
3504 			if (vp1 != NULL) {
3505 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3506 				    __FILE__, __LINE__);
3507 				if (error == 0)
3508 					break;
3509 				VOP_UNLOCK(vp1);
3510 				vp1_locked = false;
3511 				vn_lock_pair_pause("vlp1");
3512 			}
3513 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3514 			vp2_locked = true;
3515 		}
3516 		if (vp2_locked && vp1 != NULL) {
3517 			if (vp2 != NULL) {
3518 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3519 				    __FILE__, __LINE__);
3520 				if (error == 0)
3521 					break;
3522 				VOP_UNLOCK(vp2);
3523 				vp2_locked = false;
3524 				vn_lock_pair_pause("vlp2");
3525 			}
3526 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3527 			vp1_locked = true;
3528 		}
3529 	}
3530 	if (vp1 != NULL)
3531 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3532 	if (vp2 != NULL)
3533 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3534 }
3535