xref: /freebsd/sys/kern/vfs_vnops.c (revision d485c77f203fb0f4cdc08dea5ff81631b51d8809)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
13  * Copyright (c) 2013, 2014 The FreeBSD Foundation
14  *
15  * Portions of this software were developed by Konstantin Belousov
16  * under sponsorship from the FreeBSD Foundation.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_hwpmc_hooks.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/disk.h>
53 #include <sys/fail.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/kdb.h>
57 #include <sys/ktr.h>
58 #include <sys/stat.h>
59 #include <sys/priv.h>
60 #include <sys/proc.h>
61 #include <sys/limits.h>
62 #include <sys/lock.h>
63 #include <sys/mman.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/vnode.h>
68 #include <sys/bio.h>
69 #include <sys/buf.h>
70 #include <sys/filio.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/prng.h>
74 #include <sys/sx.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/sysctl.h>
77 #include <sys/ttycom.h>
78 #include <sys/conf.h>
79 #include <sys/syslog.h>
80 #include <sys/unistd.h>
81 #include <sys/user.h>
82 
83 #include <security/audit/audit.h>
84 #include <security/mac/mac_framework.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_extern.h>
88 #include <vm/pmap.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pager.h>
93 
94 #ifdef HWPMC_HOOKS
95 #include <sys/pmckern.h>
96 #endif
97 
98 static fo_rdwr_t	vn_read;
99 static fo_rdwr_t	vn_write;
100 static fo_rdwr_t	vn_io_fault;
101 static fo_truncate_t	vn_truncate;
102 static fo_ioctl_t	vn_ioctl;
103 static fo_poll_t	vn_poll;
104 static fo_kqfilter_t	vn_kqfilter;
105 static fo_stat_t	vn_statfile;
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_BENEATH) != 0)
203 		res |= BENEATH;
204 	if ((fmode & O_RESOLVE_BENEATH) != 0)
205 		res |= RBENEATH;
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 		return (EINVAL);
239 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
240 		ndp->ni_cnd.cn_nameiop = CREATE;
241 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
242 		/*
243 		 * Set NOCACHE to avoid flushing the cache when
244 		 * rolling in many files at once.
245 		 *
246 		 * Set NC_KEEPPOSENTRY to keep positive entries if they already
247 		 * exist despite NOCACHE.
248 		 */
249 		ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY;
250 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
251 			ndp->ni_cnd.cn_flags |= FOLLOW;
252 		if ((vn_open_flags & VN_OPEN_INVFS) == 0)
253 			bwillwrite();
254 		if ((error = namei(ndp)) != 0)
255 			return (error);
256 		if (ndp->ni_vp == NULL) {
257 			VATTR_NULL(vap);
258 			vap->va_type = VREG;
259 			vap->va_mode = cmode;
260 			if (fmode & O_EXCL)
261 				vap->va_vaflags |= VA_EXCLUSIVE;
262 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
263 				NDFREE(ndp, NDF_ONLY_PNBUF);
264 				vput(ndp->ni_dvp);
265 				if ((error = vn_start_write(NULL, &mp,
266 				    V_XSLEEP | PCATCH)) != 0)
267 					return (error);
268 				NDREINIT(ndp);
269 				goto restart;
270 			}
271 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
272 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
273 #ifdef MAC
274 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
275 			    &ndp->ni_cnd, vap);
276 			if (error == 0)
277 #endif
278 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
279 				    &ndp->ni_cnd, vap);
280 			vp = ndp->ni_vp;
281 			if (error == 0 && (fmode & O_EXCL) != 0 &&
282 			    (fmode & (O_EXLOCK | O_SHLOCK)) != 0) {
283 				VI_LOCK(vp);
284 				vp->v_iflag |= VI_FOPENING;
285 				VI_UNLOCK(vp);
286 				first_open = true;
287 			}
288 			VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL,
289 			    false);
290 			vn_finished_write(mp);
291 			if (error) {
292 				NDFREE(ndp, NDF_ONLY_PNBUF);
293 				if (error == ERELOOKUP) {
294 					NDREINIT(ndp);
295 					goto restart;
296 				}
297 				return (error);
298 			}
299 			fmode &= ~O_TRUNC;
300 		} else {
301 			if (ndp->ni_dvp == ndp->ni_vp)
302 				vrele(ndp->ni_dvp);
303 			else
304 				vput(ndp->ni_dvp);
305 			ndp->ni_dvp = NULL;
306 			vp = ndp->ni_vp;
307 			if (fmode & O_EXCL) {
308 				error = EEXIST;
309 				goto bad;
310 			}
311 			if (vp->v_type == VDIR) {
312 				error = EISDIR;
313 				goto bad;
314 			}
315 			fmode &= ~O_CREAT;
316 		}
317 	} else {
318 		ndp->ni_cnd.cn_nameiop = LOOKUP;
319 		ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags);
320 		ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW :
321 		    FOLLOW;
322 		if ((fmode & FWRITE) == 0)
323 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
324 		if ((error = namei(ndp)) != 0)
325 			return (error);
326 		vp = ndp->ni_vp;
327 	}
328 	error = vn_open_vnode(vp, fmode, cred, td, fp);
329 	if (first_open) {
330 		VI_LOCK(vp);
331 		vp->v_iflag &= ~VI_FOPENING;
332 		wakeup(vp);
333 		VI_UNLOCK(vp);
334 	}
335 	if (error)
336 		goto bad;
337 	*flagp = fmode;
338 	return (0);
339 bad:
340 	NDFREE(ndp, NDF_ONLY_PNBUF);
341 	vput(vp);
342 	*flagp = fmode;
343 	ndp->ni_vp = NULL;
344 	return (error);
345 }
346 
347 static int
348 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp)
349 {
350 	struct flock lf;
351 	int error, lock_flags, type;
352 
353 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock");
354 	if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0)
355 		return (0);
356 	KASSERT(fp != NULL, ("open with flock requires fp"));
357 	if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE)
358 		return (EOPNOTSUPP);
359 
360 	lock_flags = VOP_ISLOCKED(vp);
361 	VOP_UNLOCK(vp);
362 
363 	lf.l_whence = SEEK_SET;
364 	lf.l_start = 0;
365 	lf.l_len = 0;
366 	lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK;
367 	type = F_FLOCK;
368 	if ((fmode & FNONBLOCK) == 0)
369 		type |= F_WAIT;
370 	if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
371 		type |= F_FIRSTOPEN;
372 	error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
373 	if (error == 0)
374 		fp->f_flag |= FHASLOCK;
375 
376 	vn_lock(vp, lock_flags | LK_RETRY);
377 	return (error);
378 }
379 
380 /*
381  * Common code for vnode open operations once a vnode is located.
382  * Check permissions, and call the VOP_OPEN routine.
383  */
384 int
385 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
386     struct thread *td, struct file *fp)
387 {
388 	accmode_t accmode;
389 	int error;
390 
391 	if (vp->v_type == VLNK)
392 		return (EMLINK);
393 	if (vp->v_type == VSOCK)
394 		return (EOPNOTSUPP);
395 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
396 		return (ENOTDIR);
397 	accmode = 0;
398 	if (fmode & (FWRITE | O_TRUNC)) {
399 		if (vp->v_type == VDIR)
400 			return (EISDIR);
401 		accmode |= VWRITE;
402 	}
403 	if (fmode & FREAD)
404 		accmode |= VREAD;
405 	if (fmode & FEXEC)
406 		accmode |= VEXEC;
407 	if ((fmode & O_APPEND) && (fmode & FWRITE))
408 		accmode |= VAPPEND;
409 #ifdef MAC
410 	if (fmode & O_CREAT)
411 		accmode |= VCREAT;
412 	if (fmode & O_VERIFY)
413 		accmode |= VVERIFY;
414 	error = mac_vnode_check_open(cred, vp, accmode);
415 	if (error)
416 		return (error);
417 
418 	accmode &= ~(VCREAT | VVERIFY);
419 #endif
420 	if ((fmode & O_CREAT) == 0 && accmode != 0) {
421 		error = VOP_ACCESS(vp, accmode, cred, td);
422 		if (error != 0)
423 			return (error);
424 	}
425 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
426 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
427 	error = VOP_OPEN(vp, fmode, cred, td, fp);
428 	if (error != 0)
429 		return (error);
430 
431 	error = vn_open_vnode_advlock(vp, fmode, fp);
432 	if (error == 0 && (fmode & FWRITE) != 0) {
433 		error = VOP_ADD_WRITECOUNT(vp, 1);
434 		if (error == 0) {
435 			CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
436 			     __func__, vp, vp->v_writecount);
437 		}
438 	}
439 
440 	/*
441 	 * Error from advlock or VOP_ADD_WRITECOUNT() still requires
442 	 * calling VOP_CLOSE() to pair with earlier VOP_OPEN().
443 	 * Arrange for that by having fdrop() to use vn_closefile().
444 	 */
445 	if (error != 0) {
446 		fp->f_flag |= FOPENFAILED;
447 		fp->f_vnode = vp;
448 		if (fp->f_ops == &badfileops) {
449 			fp->f_type = DTYPE_VNODE;
450 			fp->f_ops = &vnops;
451 		}
452 		vref(vp);
453 	}
454 
455 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
456 	return (error);
457 
458 }
459 
460 /*
461  * Check for write permissions on the specified vnode.
462  * Prototype text segments cannot be written.
463  * It is racy.
464  */
465 int
466 vn_writechk(struct vnode *vp)
467 {
468 
469 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
470 	/*
471 	 * If there's shared text associated with
472 	 * the vnode, try to free it up once.  If
473 	 * we fail, we can't allow writing.
474 	 */
475 	if (VOP_IS_TEXT(vp))
476 		return (ETXTBSY);
477 
478 	return (0);
479 }
480 
481 /*
482  * Vnode close call
483  */
484 static int
485 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred,
486     struct thread *td, bool keep_ref)
487 {
488 	struct mount *mp;
489 	int error, lock_flags;
490 
491 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
492 	    MNT_EXTENDED_SHARED(vp->v_mount))
493 		lock_flags = LK_SHARED;
494 	else
495 		lock_flags = LK_EXCLUSIVE;
496 
497 	vn_start_write(vp, &mp, V_WAIT);
498 	vn_lock(vp, lock_flags | LK_RETRY);
499 	AUDIT_ARG_VNODE1(vp);
500 	if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) {
501 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
502 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
503 		    __func__, vp, vp->v_writecount);
504 	}
505 	error = VOP_CLOSE(vp, flags, file_cred, td);
506 	if (keep_ref)
507 		VOP_UNLOCK(vp);
508 	else
509 		vput(vp);
510 	vn_finished_write(mp);
511 	return (error);
512 }
513 
514 int
515 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
516     struct thread *td)
517 {
518 
519 	return (vn_close1(vp, flags, file_cred, td, false));
520 }
521 
522 /*
523  * Heuristic to detect sequential operation.
524  */
525 static int
526 sequential_heuristic(struct uio *uio, struct file *fp)
527 {
528 	enum uio_rw rw;
529 
530 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
531 
532 	rw = uio->uio_rw;
533 	if (fp->f_flag & FRDAHEAD)
534 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
535 
536 	/*
537 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
538 	 * that the first I/O is normally considered to be slightly
539 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
540 	 * unless previous seeks have reduced f_seqcount to 0, in which
541 	 * case offset 0 is not special.
542 	 */
543 	if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) ||
544 	    uio->uio_offset == fp->f_nextoff[rw]) {
545 		/*
546 		 * f_seqcount is in units of fixed-size blocks so that it
547 		 * depends mainly on the amount of sequential I/O and not
548 		 * much on the number of sequential I/O's.  The fixed size
549 		 * of 16384 is hard-coded here since it is (not quite) just
550 		 * a magic size that works well here.  This size is more
551 		 * closely related to the best I/O size for real disks than
552 		 * to any block size used by software.
553 		 */
554 		if (uio->uio_resid >= IO_SEQMAX * 16384)
555 			fp->f_seqcount[rw] = IO_SEQMAX;
556 		else {
557 			fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384);
558 			if (fp->f_seqcount[rw] > IO_SEQMAX)
559 				fp->f_seqcount[rw] = IO_SEQMAX;
560 		}
561 		return (fp->f_seqcount[rw] << IO_SEQSHIFT);
562 	}
563 
564 	/* Not sequential.  Quickly draw-down sequentiality. */
565 	if (fp->f_seqcount[rw] > 1)
566 		fp->f_seqcount[rw] = 1;
567 	else
568 		fp->f_seqcount[rw] = 0;
569 	return (0);
570 }
571 
572 /*
573  * Package up an I/O request on a vnode into a uio and do it.
574  */
575 int
576 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
577     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
578     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
579 {
580 	struct uio auio;
581 	struct iovec aiov;
582 	struct mount *mp;
583 	struct ucred *cred;
584 	void *rl_cookie;
585 	struct vn_io_fault_args args;
586 	int error, lock_flags;
587 
588 	if (offset < 0 && vp->v_type != VCHR)
589 		return (EINVAL);
590 	auio.uio_iov = &aiov;
591 	auio.uio_iovcnt = 1;
592 	aiov.iov_base = base;
593 	aiov.iov_len = len;
594 	auio.uio_resid = len;
595 	auio.uio_offset = offset;
596 	auio.uio_segflg = segflg;
597 	auio.uio_rw = rw;
598 	auio.uio_td = td;
599 	error = 0;
600 
601 	if ((ioflg & IO_NODELOCKED) == 0) {
602 		if ((ioflg & IO_RANGELOCKED) == 0) {
603 			if (rw == UIO_READ) {
604 				rl_cookie = vn_rangelock_rlock(vp, offset,
605 				    offset + len);
606 			} else if ((ioflg & IO_APPEND) != 0) {
607 				rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
608 			} else {
609 				rl_cookie = vn_rangelock_wlock(vp, offset,
610 				    offset + len);
611 			}
612 		} else
613 			rl_cookie = NULL;
614 		mp = NULL;
615 		if (rw == UIO_WRITE) {
616 			if (vp->v_type != VCHR &&
617 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
618 			    != 0)
619 				goto out;
620 			if (MNT_SHARED_WRITES(mp) ||
621 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
622 				lock_flags = LK_SHARED;
623 			else
624 				lock_flags = LK_EXCLUSIVE;
625 		} else
626 			lock_flags = LK_SHARED;
627 		vn_lock(vp, lock_flags | LK_RETRY);
628 	} else
629 		rl_cookie = NULL;
630 
631 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
632 #ifdef MAC
633 	if ((ioflg & IO_NOMACCHECK) == 0) {
634 		if (rw == UIO_READ)
635 			error = mac_vnode_check_read(active_cred, file_cred,
636 			    vp);
637 		else
638 			error = mac_vnode_check_write(active_cred, file_cred,
639 			    vp);
640 	}
641 #endif
642 	if (error == 0) {
643 		if (file_cred != NULL)
644 			cred = file_cred;
645 		else
646 			cred = active_cred;
647 		if (do_vn_io_fault(vp, &auio)) {
648 			args.kind = VN_IO_FAULT_VOP;
649 			args.cred = cred;
650 			args.flags = ioflg;
651 			args.args.vop_args.vp = vp;
652 			error = vn_io_fault1(vp, &auio, &args, td);
653 		} else if (rw == UIO_READ) {
654 			error = VOP_READ(vp, &auio, ioflg, cred);
655 		} else /* if (rw == UIO_WRITE) */ {
656 			error = VOP_WRITE(vp, &auio, ioflg, cred);
657 		}
658 	}
659 	if (aresid)
660 		*aresid = auio.uio_resid;
661 	else
662 		if (auio.uio_resid && error == 0)
663 			error = EIO;
664 	if ((ioflg & IO_NODELOCKED) == 0) {
665 		VOP_UNLOCK(vp);
666 		if (mp != NULL)
667 			vn_finished_write(mp);
668 	}
669  out:
670 	if (rl_cookie != NULL)
671 		vn_rangelock_unlock(vp, rl_cookie);
672 	return (error);
673 }
674 
675 /*
676  * Package up an I/O request on a vnode into a uio and do it.  The I/O
677  * request is split up into smaller chunks and we try to avoid saturating
678  * the buffer cache while potentially holding a vnode locked, so we
679  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
680  * to give other processes a chance to lock the vnode (either other processes
681  * core'ing the same binary, or unrelated processes scanning the directory).
682  */
683 int
684 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len,
685     off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred,
686     struct ucred *file_cred, size_t *aresid, struct thread *td)
687 {
688 	int error = 0;
689 	ssize_t iaresid;
690 
691 	do {
692 		int chunk;
693 
694 		/*
695 		 * Force `offset' to a multiple of MAXBSIZE except possibly
696 		 * for the first chunk, so that filesystems only need to
697 		 * write full blocks except possibly for the first and last
698 		 * chunks.
699 		 */
700 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
701 
702 		if (chunk > len)
703 			chunk = len;
704 		if (rw != UIO_READ && vp->v_type == VREG)
705 			bwillwrite();
706 		iaresid = 0;
707 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
708 		    ioflg, active_cred, file_cred, &iaresid, td);
709 		len -= chunk;	/* aresid calc already includes length */
710 		if (error)
711 			break;
712 		offset += chunk;
713 		base = (char *)base + chunk;
714 		kern_yield(PRI_USER);
715 	} while (len);
716 	if (aresid)
717 		*aresid = len + iaresid;
718 	return (error);
719 }
720 
721 #if OFF_MAX <= LONG_MAX
722 off_t
723 foffset_lock(struct file *fp, int flags)
724 {
725 	volatile short *flagsp;
726 	off_t res;
727 	short state;
728 
729 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
730 
731 	if ((flags & FOF_NOLOCK) != 0)
732 		return (atomic_load_long(&fp->f_offset));
733 
734 	/*
735 	 * According to McKusick the vn lock was protecting f_offset here.
736 	 * It is now protected by the FOFFSET_LOCKED flag.
737 	 */
738 	flagsp = &fp->f_vnread_flags;
739 	if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED))
740 		return (atomic_load_long(&fp->f_offset));
741 
742 	sleepq_lock(&fp->f_vnread_flags);
743 	state = atomic_load_16(flagsp);
744 	for (;;) {
745 		if ((state & FOFFSET_LOCKED) == 0) {
746 			if (!atomic_fcmpset_acq_16(flagsp, &state,
747 			    FOFFSET_LOCKED))
748 				continue;
749 			break;
750 		}
751 		if ((state & FOFFSET_LOCK_WAITING) == 0) {
752 			if (!atomic_fcmpset_acq_16(flagsp, &state,
753 			    state | FOFFSET_LOCK_WAITING))
754 				continue;
755 		}
756 		DROP_GIANT();
757 		sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0);
758 		sleepq_wait(&fp->f_vnread_flags, PUSER -1);
759 		PICKUP_GIANT();
760 		sleepq_lock(&fp->f_vnread_flags);
761 		state = atomic_load_16(flagsp);
762 	}
763 	res = atomic_load_long(&fp->f_offset);
764 	sleepq_release(&fp->f_vnread_flags);
765 	return (res);
766 }
767 
768 void
769 foffset_unlock(struct file *fp, off_t val, int flags)
770 {
771 	volatile short *flagsp;
772 	short state;
773 
774 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
775 
776 	if ((flags & FOF_NOUPDATE) == 0)
777 		atomic_store_long(&fp->f_offset, val);
778 	if ((flags & FOF_NEXTOFF_R) != 0)
779 		fp->f_nextoff[UIO_READ] = val;
780 	if ((flags & FOF_NEXTOFF_W) != 0)
781 		fp->f_nextoff[UIO_WRITE] = val;
782 
783 	if ((flags & FOF_NOLOCK) != 0)
784 		return;
785 
786 	flagsp = &fp->f_vnread_flags;
787 	state = atomic_load_16(flagsp);
788 	if ((state & FOFFSET_LOCK_WAITING) == 0 &&
789 	    atomic_cmpset_rel_16(flagsp, state, 0))
790 		return;
791 
792 	sleepq_lock(&fp->f_vnread_flags);
793 	MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0);
794 	MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0);
795 	fp->f_vnread_flags = 0;
796 	sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0);
797 	sleepq_release(&fp->f_vnread_flags);
798 }
799 #else
800 off_t
801 foffset_lock(struct file *fp, int flags)
802 {
803 	struct mtx *mtxp;
804 	off_t res;
805 
806 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
807 
808 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
809 	mtx_lock(mtxp);
810 	if ((flags & FOF_NOLOCK) == 0) {
811 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
812 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
813 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
814 			    "vofflock", 0);
815 		}
816 		fp->f_vnread_flags |= FOFFSET_LOCKED;
817 	}
818 	res = fp->f_offset;
819 	mtx_unlock(mtxp);
820 	return (res);
821 }
822 
823 void
824 foffset_unlock(struct file *fp, off_t val, int flags)
825 {
826 	struct mtx *mtxp;
827 
828 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
829 
830 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
831 	mtx_lock(mtxp);
832 	if ((flags & FOF_NOUPDATE) == 0)
833 		fp->f_offset = val;
834 	if ((flags & FOF_NEXTOFF_R) != 0)
835 		fp->f_nextoff[UIO_READ] = val;
836 	if ((flags & FOF_NEXTOFF_W) != 0)
837 		fp->f_nextoff[UIO_WRITE] = val;
838 	if ((flags & FOF_NOLOCK) == 0) {
839 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
840 		    ("Lost FOFFSET_LOCKED"));
841 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
842 			wakeup(&fp->f_vnread_flags);
843 		fp->f_vnread_flags = 0;
844 	}
845 	mtx_unlock(mtxp);
846 }
847 #endif
848 
849 void
850 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
851 {
852 
853 	if ((flags & FOF_OFFSET) == 0)
854 		uio->uio_offset = foffset_lock(fp, flags);
855 }
856 
857 void
858 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
859 {
860 
861 	if ((flags & FOF_OFFSET) == 0)
862 		foffset_unlock(fp, uio->uio_offset, flags);
863 }
864 
865 static int
866 get_advice(struct file *fp, struct uio *uio)
867 {
868 	struct mtx *mtxp;
869 	int ret;
870 
871 	ret = POSIX_FADV_NORMAL;
872 	if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
873 		return (ret);
874 
875 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
876 	mtx_lock(mtxp);
877 	if (fp->f_advice != NULL &&
878 	    uio->uio_offset >= fp->f_advice->fa_start &&
879 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
880 		ret = fp->f_advice->fa_advice;
881 	mtx_unlock(mtxp);
882 	return (ret);
883 }
884 
885 int
886 vn_read_from_obj(struct vnode *vp, struct uio *uio)
887 {
888 	vm_object_t obj;
889 	vm_page_t ma[io_hold_cnt + 2];
890 	off_t off, vsz;
891 	ssize_t resid;
892 	int error, i, j;
893 
894 	MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2));
895 	obj = atomic_load_ptr(&vp->v_object);
896 	if (obj == NULL)
897 		return (EJUSTRETURN);
898 
899 	/*
900 	 * Depends on type stability of vm_objects.
901 	 */
902 	vm_object_pip_add(obj, 1);
903 	if ((obj->flags & OBJ_DEAD) != 0) {
904 		/*
905 		 * Note that object might be already reused from the
906 		 * vnode, and the OBJ_DEAD flag cleared.  This is fine,
907 		 * we recheck for DOOMED vnode state after all pages
908 		 * are busied, and retract then.
909 		 *
910 		 * But we check for OBJ_DEAD to ensure that we do not
911 		 * busy pages while vm_object_terminate_pages()
912 		 * processes the queue.
913 		 */
914 		error = EJUSTRETURN;
915 		goto out_pip;
916 	}
917 
918 	resid = uio->uio_resid;
919 	off = uio->uio_offset;
920 	for (i = 0; resid > 0; i++) {
921 		MPASS(i < io_hold_cnt + 2);
922 		ma[i] = vm_page_grab_unlocked(obj, atop(off),
923 		    VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY |
924 		    VM_ALLOC_NOWAIT);
925 		if (ma[i] == NULL)
926 			break;
927 
928 		/*
929 		 * Skip invalid pages.  Valid mask can be partial only
930 		 * at EOF, and we clip later.
931 		 */
932 		if (vm_page_none_valid(ma[i])) {
933 			vm_page_sunbusy(ma[i]);
934 			break;
935 		}
936 
937 		resid -= PAGE_SIZE;
938 		off += PAGE_SIZE;
939 	}
940 	if (i == 0) {
941 		error = EJUSTRETURN;
942 		goto out_pip;
943 	}
944 
945 	/*
946 	 * Check VIRF_DOOMED after we busied our pages.  Since
947 	 * vgonel() terminates the vnode' vm_object, it cannot
948 	 * process past pages busied by us.
949 	 */
950 	if (VN_IS_DOOMED(vp)) {
951 		error = EJUSTRETURN;
952 		goto out;
953 	}
954 
955 	resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1);
956 	if (resid > uio->uio_resid)
957 		resid = uio->uio_resid;
958 
959 	/*
960 	 * Unlocked read of vnp_size is safe because truncation cannot
961 	 * pass busied page.  But we load vnp_size into a local
962 	 * variable so that possible concurrent extension does not
963 	 * break calculation.
964 	 */
965 #if defined(__powerpc__) && !defined(__powerpc64__)
966 	vsz = obj->un_pager.vnp.vnp_size;
967 #else
968 	vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size);
969 #endif
970 	if (uio->uio_offset >= vsz) {
971 		error = EJUSTRETURN;
972 		goto out;
973 	}
974 	if (uio->uio_offset + resid > vsz)
975 		resid = vsz - uio->uio_offset;
976 
977 	error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio);
978 
979 out:
980 	for (j = 0; j < i; j++) {
981 		if (error == 0)
982 			vm_page_reference(ma[j]);
983 		vm_page_sunbusy(ma[j]);
984 	}
985 out_pip:
986 	vm_object_pip_wakeup(obj);
987 	if (error != 0)
988 		return (error);
989 	return (uio->uio_resid == 0 ? 0 : EJUSTRETURN);
990 }
991 
992 /*
993  * File table vnode read routine.
994  */
995 static int
996 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
997     struct thread *td)
998 {
999 	struct vnode *vp;
1000 	off_t orig_offset;
1001 	int error, ioflag;
1002 	int advice;
1003 
1004 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1005 	    uio->uio_td, td));
1006 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1007 	vp = fp->f_vnode;
1008 	ioflag = 0;
1009 	if (fp->f_flag & FNONBLOCK)
1010 		ioflag |= IO_NDELAY;
1011 	if (fp->f_flag & O_DIRECT)
1012 		ioflag |= IO_DIRECT;
1013 
1014 	/*
1015 	 * Try to read from page cache.  VIRF_DOOMED check is racy but
1016 	 * allows us to avoid unneeded work outright.
1017 	 */
1018 	if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() &&
1019 	    (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) {
1020 		error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred);
1021 		if (error == 0) {
1022 			fp->f_nextoff[UIO_READ] = uio->uio_offset;
1023 			return (0);
1024 		}
1025 		if (error != EJUSTRETURN)
1026 			return (error);
1027 	}
1028 
1029 	advice = get_advice(fp, uio);
1030 	vn_lock(vp, LK_SHARED | LK_RETRY);
1031 
1032 	switch (advice) {
1033 	case POSIX_FADV_NORMAL:
1034 	case POSIX_FADV_SEQUENTIAL:
1035 	case POSIX_FADV_NOREUSE:
1036 		ioflag |= sequential_heuristic(uio, fp);
1037 		break;
1038 	case POSIX_FADV_RANDOM:
1039 		/* Disable read-ahead for random I/O. */
1040 		break;
1041 	}
1042 	orig_offset = uio->uio_offset;
1043 
1044 #ifdef MAC
1045 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
1046 	if (error == 0)
1047 #endif
1048 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
1049 	fp->f_nextoff[UIO_READ] = uio->uio_offset;
1050 	VOP_UNLOCK(vp);
1051 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1052 	    orig_offset != uio->uio_offset)
1053 		/*
1054 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1055 		 * for the backing file after a POSIX_FADV_NOREUSE
1056 		 * read(2).
1057 		 */
1058 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1059 		    POSIX_FADV_DONTNEED);
1060 	return (error);
1061 }
1062 
1063 /*
1064  * File table vnode write routine.
1065  */
1066 static int
1067 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
1068     struct thread *td)
1069 {
1070 	struct vnode *vp;
1071 	struct mount *mp;
1072 	off_t orig_offset;
1073 	int error, ioflag, lock_flags;
1074 	int advice;
1075 
1076 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
1077 	    uio->uio_td, td));
1078 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
1079 	vp = fp->f_vnode;
1080 	if (vp->v_type == VREG)
1081 		bwillwrite();
1082 	ioflag = IO_UNIT;
1083 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
1084 		ioflag |= IO_APPEND;
1085 	if (fp->f_flag & FNONBLOCK)
1086 		ioflag |= IO_NDELAY;
1087 	if (fp->f_flag & O_DIRECT)
1088 		ioflag |= IO_DIRECT;
1089 	if ((fp->f_flag & O_FSYNC) ||
1090 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
1091 		ioflag |= IO_SYNC;
1092 	/*
1093 	 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
1094 	 * implementations that don't understand IO_DATASYNC fall back to full
1095 	 * O_SYNC behavior.
1096 	 */
1097 	if (fp->f_flag & O_DSYNC)
1098 		ioflag |= IO_SYNC | IO_DATASYNC;
1099 	mp = NULL;
1100 	if (vp->v_type != VCHR &&
1101 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
1102 		goto unlock;
1103 
1104 	advice = get_advice(fp, uio);
1105 
1106 	if (MNT_SHARED_WRITES(mp) ||
1107 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
1108 		lock_flags = LK_SHARED;
1109 	} else {
1110 		lock_flags = LK_EXCLUSIVE;
1111 	}
1112 
1113 	vn_lock(vp, lock_flags | LK_RETRY);
1114 	switch (advice) {
1115 	case POSIX_FADV_NORMAL:
1116 	case POSIX_FADV_SEQUENTIAL:
1117 	case POSIX_FADV_NOREUSE:
1118 		ioflag |= sequential_heuristic(uio, fp);
1119 		break;
1120 	case POSIX_FADV_RANDOM:
1121 		/* XXX: Is this correct? */
1122 		break;
1123 	}
1124 	orig_offset = uio->uio_offset;
1125 
1126 #ifdef MAC
1127 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1128 	if (error == 0)
1129 #endif
1130 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
1131 	fp->f_nextoff[UIO_WRITE] = uio->uio_offset;
1132 	VOP_UNLOCK(vp);
1133 	if (vp->v_type != VCHR)
1134 		vn_finished_write(mp);
1135 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
1136 	    orig_offset != uio->uio_offset)
1137 		/*
1138 		 * Use POSIX_FADV_DONTNEED to flush pages and buffers
1139 		 * for the backing file after a POSIX_FADV_NOREUSE
1140 		 * write(2).
1141 		 */
1142 		error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1,
1143 		    POSIX_FADV_DONTNEED);
1144 unlock:
1145 	return (error);
1146 }
1147 
1148 /*
1149  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
1150  * prevent the following deadlock:
1151  *
1152  * Assume that the thread A reads from the vnode vp1 into userspace
1153  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
1154  * currently not resident, then system ends up with the call chain
1155  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
1156  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
1157  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
1158  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
1159  * backed by the pages of vnode vp1, and some page in buf2 is not
1160  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
1161  *
1162  * To prevent the lock order reversal and deadlock, vn_io_fault() does
1163  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
1164  * Instead, it first tries to do the whole range i/o with pagefaults
1165  * disabled. If all pages in the i/o buffer are resident and mapped,
1166  * VOP will succeed (ignoring the genuine filesystem errors).
1167  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
1168  * i/o in chunks, with all pages in the chunk prefaulted and held
1169  * using vm_fault_quick_hold_pages().
1170  *
1171  * Filesystems using this deadlock avoidance scheme should use the
1172  * array of the held pages from uio, saved in the curthread->td_ma,
1173  * instead of doing uiomove().  A helper function
1174  * vn_io_fault_uiomove() converts uiomove request into
1175  * uiomove_fromphys() over td_ma array.
1176  *
1177  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1178  * make the current i/o request atomic with respect to other i/os and
1179  * truncations.
1180  */
1181 
1182 /*
1183  * Decode vn_io_fault_args and perform the corresponding i/o.
1184  */
1185 static int
1186 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1187     struct thread *td)
1188 {
1189 	int error, save;
1190 
1191 	error = 0;
1192 	save = vm_fault_disable_pagefaults();
1193 	switch (args->kind) {
1194 	case VN_IO_FAULT_FOP:
1195 		error = (args->args.fop_args.doio)(args->args.fop_args.fp,
1196 		    uio, args->cred, args->flags, td);
1197 		break;
1198 	case VN_IO_FAULT_VOP:
1199 		if (uio->uio_rw == UIO_READ) {
1200 			error = VOP_READ(args->args.vop_args.vp, uio,
1201 			    args->flags, args->cred);
1202 		} else if (uio->uio_rw == UIO_WRITE) {
1203 			error = VOP_WRITE(args->args.vop_args.vp, uio,
1204 			    args->flags, args->cred);
1205 		}
1206 		break;
1207 	default:
1208 		panic("vn_io_fault_doio: unknown kind of io %d %d",
1209 		    args->kind, uio->uio_rw);
1210 	}
1211 	vm_fault_enable_pagefaults(save);
1212 	return (error);
1213 }
1214 
1215 static int
1216 vn_io_fault_touch(char *base, const struct uio *uio)
1217 {
1218 	int r;
1219 
1220 	r = fubyte(base);
1221 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1222 		return (EFAULT);
1223 	return (0);
1224 }
1225 
1226 static int
1227 vn_io_fault_prefault_user(const struct uio *uio)
1228 {
1229 	char *base;
1230 	const struct iovec *iov;
1231 	size_t len;
1232 	ssize_t resid;
1233 	int error, i;
1234 
1235 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1236 	    ("vn_io_fault_prefault userspace"));
1237 
1238 	error = i = 0;
1239 	iov = uio->uio_iov;
1240 	resid = uio->uio_resid;
1241 	base = iov->iov_base;
1242 	len = iov->iov_len;
1243 	while (resid > 0) {
1244 		error = vn_io_fault_touch(base, uio);
1245 		if (error != 0)
1246 			break;
1247 		if (len < PAGE_SIZE) {
1248 			if (len != 0) {
1249 				error = vn_io_fault_touch(base + len - 1, uio);
1250 				if (error != 0)
1251 					break;
1252 				resid -= len;
1253 			}
1254 			if (++i >= uio->uio_iovcnt)
1255 				break;
1256 			iov = uio->uio_iov + i;
1257 			base = iov->iov_base;
1258 			len = iov->iov_len;
1259 		} else {
1260 			len -= PAGE_SIZE;
1261 			base += PAGE_SIZE;
1262 			resid -= PAGE_SIZE;
1263 		}
1264 	}
1265 	return (error);
1266 }
1267 
1268 /*
1269  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1270  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1271  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1272  * into args and call vn_io_fault1() to handle faults during the user
1273  * mode buffer accesses.
1274  */
1275 static int
1276 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1277     struct thread *td)
1278 {
1279 	vm_page_t ma[io_hold_cnt + 2];
1280 	struct uio *uio_clone, short_uio;
1281 	struct iovec short_iovec[1];
1282 	vm_page_t *prev_td_ma;
1283 	vm_prot_t prot;
1284 	vm_offset_t addr, end;
1285 	size_t len, resid;
1286 	ssize_t adv;
1287 	int error, cnt, saveheld, prev_td_ma_cnt;
1288 
1289 	if (vn_io_fault_prefault) {
1290 		error = vn_io_fault_prefault_user(uio);
1291 		if (error != 0)
1292 			return (error); /* Or ignore ? */
1293 	}
1294 
1295 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1296 
1297 	/*
1298 	 * The UFS follows IO_UNIT directive and replays back both
1299 	 * uio_offset and uio_resid if an error is encountered during the
1300 	 * operation.  But, since the iovec may be already advanced,
1301 	 * uio is still in an inconsistent state.
1302 	 *
1303 	 * Cache a copy of the original uio, which is advanced to the redo
1304 	 * point using UIO_NOCOPY below.
1305 	 */
1306 	uio_clone = cloneuio(uio);
1307 	resid = uio->uio_resid;
1308 
1309 	short_uio.uio_segflg = UIO_USERSPACE;
1310 	short_uio.uio_rw = uio->uio_rw;
1311 	short_uio.uio_td = uio->uio_td;
1312 
1313 	error = vn_io_fault_doio(args, uio, td);
1314 	if (error != EFAULT)
1315 		goto out;
1316 
1317 	atomic_add_long(&vn_io_faults_cnt, 1);
1318 	uio_clone->uio_segflg = UIO_NOCOPY;
1319 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1320 	uio_clone->uio_segflg = uio->uio_segflg;
1321 
1322 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1323 	prev_td_ma = td->td_ma;
1324 	prev_td_ma_cnt = td->td_ma_cnt;
1325 
1326 	while (uio_clone->uio_resid != 0) {
1327 		len = uio_clone->uio_iov->iov_len;
1328 		if (len == 0) {
1329 			KASSERT(uio_clone->uio_iovcnt >= 1,
1330 			    ("iovcnt underflow"));
1331 			uio_clone->uio_iov++;
1332 			uio_clone->uio_iovcnt--;
1333 			continue;
1334 		}
1335 		if (len > ptoa(io_hold_cnt))
1336 			len = ptoa(io_hold_cnt);
1337 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1338 		end = round_page(addr + len);
1339 		if (end < addr) {
1340 			error = EFAULT;
1341 			break;
1342 		}
1343 		cnt = atop(end - trunc_page(addr));
1344 		/*
1345 		 * A perfectly misaligned address and length could cause
1346 		 * both the start and the end of the chunk to use partial
1347 		 * page.  +2 accounts for such a situation.
1348 		 */
1349 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1350 		    addr, len, prot, ma, io_hold_cnt + 2);
1351 		if (cnt == -1) {
1352 			error = EFAULT;
1353 			break;
1354 		}
1355 		short_uio.uio_iov = &short_iovec[0];
1356 		short_iovec[0].iov_base = (void *)addr;
1357 		short_uio.uio_iovcnt = 1;
1358 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1359 		short_uio.uio_offset = uio_clone->uio_offset;
1360 		td->td_ma = ma;
1361 		td->td_ma_cnt = cnt;
1362 
1363 		error = vn_io_fault_doio(args, &short_uio, td);
1364 		vm_page_unhold_pages(ma, cnt);
1365 		adv = len - short_uio.uio_resid;
1366 
1367 		uio_clone->uio_iov->iov_base =
1368 		    (char *)uio_clone->uio_iov->iov_base + adv;
1369 		uio_clone->uio_iov->iov_len -= adv;
1370 		uio_clone->uio_resid -= adv;
1371 		uio_clone->uio_offset += adv;
1372 
1373 		uio->uio_resid -= adv;
1374 		uio->uio_offset += adv;
1375 
1376 		if (error != 0 || adv == 0)
1377 			break;
1378 	}
1379 	td->td_ma = prev_td_ma;
1380 	td->td_ma_cnt = prev_td_ma_cnt;
1381 	curthread_pflags_restore(saveheld);
1382 out:
1383 	free(uio_clone, M_IOV);
1384 	return (error);
1385 }
1386 
1387 static int
1388 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1389     int flags, struct thread *td)
1390 {
1391 	fo_rdwr_t *doio;
1392 	struct vnode *vp;
1393 	void *rl_cookie;
1394 	struct vn_io_fault_args args;
1395 	int error;
1396 
1397 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1398 	vp = fp->f_vnode;
1399 
1400 	/*
1401 	 * The ability to read(2) on a directory has historically been
1402 	 * allowed for all users, but this can and has been the source of
1403 	 * at least one security issue in the past.  As such, it is now hidden
1404 	 * away behind a sysctl for those that actually need it to use it, and
1405 	 * restricted to root when it's turned on to make it relatively safe to
1406 	 * leave on for longer sessions of need.
1407 	 */
1408 	if (vp->v_type == VDIR) {
1409 		KASSERT(uio->uio_rw == UIO_READ,
1410 		    ("illegal write attempted on a directory"));
1411 		if (!vfs_allow_read_dir)
1412 			return (EISDIR);
1413 		if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0)
1414 			return (EISDIR);
1415 	}
1416 
1417 	foffset_lock_uio(fp, uio, flags);
1418 	if (do_vn_io_fault(vp, uio)) {
1419 		args.kind = VN_IO_FAULT_FOP;
1420 		args.args.fop_args.fp = fp;
1421 		args.args.fop_args.doio = doio;
1422 		args.cred = active_cred;
1423 		args.flags = flags | FOF_OFFSET;
1424 		if (uio->uio_rw == UIO_READ) {
1425 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1426 			    uio->uio_offset + uio->uio_resid);
1427 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1428 		    (flags & FOF_OFFSET) == 0) {
1429 			/* For appenders, punt and lock the whole range. */
1430 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1431 		} else {
1432 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1433 			    uio->uio_offset + uio->uio_resid);
1434 		}
1435 		error = vn_io_fault1(vp, uio, &args, td);
1436 		vn_rangelock_unlock(vp, rl_cookie);
1437 	} else {
1438 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1439 	}
1440 	foffset_unlock_uio(fp, uio, flags);
1441 	return (error);
1442 }
1443 
1444 /*
1445  * Helper function to perform the requested uiomove operation using
1446  * the held pages for io->uio_iov[0].iov_base buffer instead of
1447  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1448  * instead of iov_base prevents page faults that could occur due to
1449  * pmap_collect() invalidating the mapping created by
1450  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1451  * object cleanup revoking the write access from page mappings.
1452  *
1453  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1454  * instead of plain uiomove().
1455  */
1456 int
1457 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1458 {
1459 	struct uio transp_uio;
1460 	struct iovec transp_iov[1];
1461 	struct thread *td;
1462 	size_t adv;
1463 	int error, pgadv;
1464 
1465 	td = curthread;
1466 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1467 	    uio->uio_segflg != UIO_USERSPACE)
1468 		return (uiomove(data, xfersize, uio));
1469 
1470 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1471 	transp_iov[0].iov_base = data;
1472 	transp_uio.uio_iov = &transp_iov[0];
1473 	transp_uio.uio_iovcnt = 1;
1474 	if (xfersize > uio->uio_resid)
1475 		xfersize = uio->uio_resid;
1476 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1477 	transp_uio.uio_offset = 0;
1478 	transp_uio.uio_segflg = UIO_SYSSPACE;
1479 	/*
1480 	 * Since transp_iov points to data, and td_ma page array
1481 	 * corresponds to original uio->uio_iov, we need to invert the
1482 	 * direction of the i/o operation as passed to
1483 	 * uiomove_fromphys().
1484 	 */
1485 	switch (uio->uio_rw) {
1486 	case UIO_WRITE:
1487 		transp_uio.uio_rw = UIO_READ;
1488 		break;
1489 	case UIO_READ:
1490 		transp_uio.uio_rw = UIO_WRITE;
1491 		break;
1492 	}
1493 	transp_uio.uio_td = uio->uio_td;
1494 	error = uiomove_fromphys(td->td_ma,
1495 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1496 	    xfersize, &transp_uio);
1497 	adv = xfersize - transp_uio.uio_resid;
1498 	pgadv =
1499 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1500 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1501 	td->td_ma += pgadv;
1502 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1503 	    pgadv));
1504 	td->td_ma_cnt -= pgadv;
1505 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1506 	uio->uio_iov->iov_len -= adv;
1507 	uio->uio_resid -= adv;
1508 	uio->uio_offset += adv;
1509 	return (error);
1510 }
1511 
1512 int
1513 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1514     struct uio *uio)
1515 {
1516 	struct thread *td;
1517 	vm_offset_t iov_base;
1518 	int cnt, pgadv;
1519 
1520 	td = curthread;
1521 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1522 	    uio->uio_segflg != UIO_USERSPACE)
1523 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1524 
1525 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1526 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1527 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1528 	switch (uio->uio_rw) {
1529 	case UIO_WRITE:
1530 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1531 		    offset, cnt);
1532 		break;
1533 	case UIO_READ:
1534 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1535 		    cnt);
1536 		break;
1537 	}
1538 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1539 	td->td_ma += pgadv;
1540 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1541 	    pgadv));
1542 	td->td_ma_cnt -= pgadv;
1543 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1544 	uio->uio_iov->iov_len -= cnt;
1545 	uio->uio_resid -= cnt;
1546 	uio->uio_offset += cnt;
1547 	return (0);
1548 }
1549 
1550 /*
1551  * File table truncate routine.
1552  */
1553 static int
1554 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1555     struct thread *td)
1556 {
1557 	struct mount *mp;
1558 	struct vnode *vp;
1559 	void *rl_cookie;
1560 	int error;
1561 
1562 	vp = fp->f_vnode;
1563 
1564 retry:
1565 	/*
1566 	 * Lock the whole range for truncation.  Otherwise split i/o
1567 	 * might happen partly before and partly after the truncation.
1568 	 */
1569 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1570 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1571 	if (error)
1572 		goto out1;
1573 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1574 	AUDIT_ARG_VNODE1(vp);
1575 	if (vp->v_type == VDIR) {
1576 		error = EISDIR;
1577 		goto out;
1578 	}
1579 #ifdef MAC
1580 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1581 	if (error)
1582 		goto out;
1583 #endif
1584 	error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0,
1585 	    fp->f_cred);
1586 out:
1587 	VOP_UNLOCK(vp);
1588 	vn_finished_write(mp);
1589 out1:
1590 	vn_rangelock_unlock(vp, rl_cookie);
1591 	if (error == ERELOOKUP)
1592 		goto retry;
1593 	return (error);
1594 }
1595 
1596 /*
1597  * Truncate a file that is already locked.
1598  */
1599 int
1600 vn_truncate_locked(struct vnode *vp, off_t length, bool sync,
1601     struct ucred *cred)
1602 {
1603 	struct vattr vattr;
1604 	int error;
1605 
1606 	error = VOP_ADD_WRITECOUNT(vp, 1);
1607 	if (error == 0) {
1608 		VATTR_NULL(&vattr);
1609 		vattr.va_size = length;
1610 		if (sync)
1611 			vattr.va_vaflags |= VA_SYNC;
1612 		error = VOP_SETATTR(vp, &vattr, cred);
1613 		VOP_ADD_WRITECOUNT_CHECKED(vp, -1);
1614 	}
1615 	return (error);
1616 }
1617 
1618 /*
1619  * File table vnode stat routine.
1620  */
1621 static int
1622 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred,
1623     struct thread *td)
1624 {
1625 	struct vnode *vp = fp->f_vnode;
1626 	int error;
1627 
1628 	vn_lock(vp, LK_SHARED | LK_RETRY);
1629 	error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td);
1630 	VOP_UNLOCK(vp);
1631 
1632 	return (error);
1633 }
1634 
1635 /*
1636  * File table vnode ioctl routine.
1637  */
1638 static int
1639 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
1640     struct thread *td)
1641 {
1642 	struct vattr vattr;
1643 	struct vnode *vp;
1644 	struct fiobmap2_arg *bmarg;
1645 	int error;
1646 
1647 	vp = fp->f_vnode;
1648 	switch (vp->v_type) {
1649 	case VDIR:
1650 	case VREG:
1651 		switch (com) {
1652 		case FIONREAD:
1653 			vn_lock(vp, LK_SHARED | LK_RETRY);
1654 			error = VOP_GETATTR(vp, &vattr, active_cred);
1655 			VOP_UNLOCK(vp);
1656 			if (error == 0)
1657 				*(int *)data = vattr.va_size - fp->f_offset;
1658 			return (error);
1659 		case FIOBMAP2:
1660 			bmarg = (struct fiobmap2_arg *)data;
1661 			vn_lock(vp, LK_SHARED | LK_RETRY);
1662 #ifdef MAC
1663 			error = mac_vnode_check_read(active_cred, fp->f_cred,
1664 			    vp);
1665 			if (error == 0)
1666 #endif
1667 				error = VOP_BMAP(vp, bmarg->bn, NULL,
1668 				    &bmarg->bn, &bmarg->runp, &bmarg->runb);
1669 			VOP_UNLOCK(vp);
1670 			return (error);
1671 		case FIONBIO:
1672 		case FIOASYNC:
1673 			return (0);
1674 		default:
1675 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1676 			    active_cred, td));
1677 		}
1678 		break;
1679 	case VCHR:
1680 		return (VOP_IOCTL(vp, com, data, fp->f_flag,
1681 		    active_cred, td));
1682 	default:
1683 		return (ENOTTY);
1684 	}
1685 }
1686 
1687 /*
1688  * File table vnode poll routine.
1689  */
1690 static int
1691 vn_poll(struct file *fp, int events, struct ucred *active_cred,
1692     struct thread *td)
1693 {
1694 	struct vnode *vp;
1695 	int error;
1696 
1697 	vp = fp->f_vnode;
1698 #if defined(MAC) || defined(AUDIT)
1699 	if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) {
1700 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1701 		AUDIT_ARG_VNODE1(vp);
1702 		error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1703 		VOP_UNLOCK(vp);
1704 		if (error != 0)
1705 			return (error);
1706 	}
1707 #endif
1708 	error = VOP_POLL(vp, events, fp->f_cred, td);
1709 	return (error);
1710 }
1711 
1712 /*
1713  * Acquire the requested lock and then check for validity.  LK_RETRY
1714  * permits vn_lock to return doomed vnodes.
1715  */
1716 static int __noinline
1717 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line,
1718     int error)
1719 {
1720 
1721 	KASSERT((flags & LK_RETRY) == 0 || error == 0,
1722 	    ("vn_lock: error %d incompatible with flags %#x", error, flags));
1723 
1724 	if (error == 0)
1725 		VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed"));
1726 
1727 	if ((flags & LK_RETRY) == 0) {
1728 		if (error == 0) {
1729 			VOP_UNLOCK(vp);
1730 			error = ENOENT;
1731 		}
1732 		return (error);
1733 	}
1734 
1735 	/*
1736 	 * LK_RETRY case.
1737 	 *
1738 	 * Nothing to do if we got the lock.
1739 	 */
1740 	if (error == 0)
1741 		return (0);
1742 
1743 	/*
1744 	 * Interlock was dropped by the call in _vn_lock.
1745 	 */
1746 	flags &= ~LK_INTERLOCK;
1747 	do {
1748 		error = VOP_LOCK1(vp, flags, file, line);
1749 	} while (error != 0);
1750 	return (0);
1751 }
1752 
1753 int
1754 _vn_lock(struct vnode *vp, int flags, const char *file, int line)
1755 {
1756 	int error;
1757 
1758 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1759 	    ("vn_lock: no locktype (%d passed)", flags));
1760 	VNPASS(vp->v_holdcnt > 0, vp);
1761 	error = VOP_LOCK1(vp, flags, file, line);
1762 	if (__predict_false(error != 0 || VN_IS_DOOMED(vp)))
1763 		return (_vn_lock_fallback(vp, flags, file, line, error));
1764 	return (0);
1765 }
1766 
1767 /*
1768  * File table vnode close routine.
1769  */
1770 static int
1771 vn_closefile(struct file *fp, struct thread *td)
1772 {
1773 	struct vnode *vp;
1774 	struct flock lf;
1775 	int error;
1776 	bool ref;
1777 
1778 	vp = fp->f_vnode;
1779 	fp->f_ops = &badfileops;
1780 	ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1781 
1782 	error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1783 
1784 	if (__predict_false(ref)) {
1785 		lf.l_whence = SEEK_SET;
1786 		lf.l_start = 0;
1787 		lf.l_len = 0;
1788 		lf.l_type = F_UNLCK;
1789 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1790 		vrele(vp);
1791 	}
1792 	return (error);
1793 }
1794 
1795 /*
1796  * Preparing to start a filesystem write operation. If the operation is
1797  * permitted, then we bump the count of operations in progress and
1798  * proceed. If a suspend request is in progress, we wait until the
1799  * suspension is over, and then proceed.
1800  */
1801 static int
1802 vn_start_write_refed(struct mount *mp, int flags, bool mplocked)
1803 {
1804 	struct mount_pcpu *mpcpu;
1805 	int error, mflags;
1806 
1807 	if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 &&
1808 	    vfs_op_thread_enter(mp, mpcpu)) {
1809 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1810 		vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1);
1811 		vfs_op_thread_exit(mp, mpcpu);
1812 		return (0);
1813 	}
1814 
1815 	if (mplocked)
1816 		mtx_assert(MNT_MTX(mp), MA_OWNED);
1817 	else
1818 		MNT_ILOCK(mp);
1819 
1820 	error = 0;
1821 
1822 	/*
1823 	 * Check on status of suspension.
1824 	 */
1825 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1826 	    mp->mnt_susp_owner != curthread) {
1827 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1828 		    (flags & PCATCH) : 0) | (PUSER - 1);
1829 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1830 			if (flags & V_NOWAIT) {
1831 				error = EWOULDBLOCK;
1832 				goto unlock;
1833 			}
1834 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1835 			    "suspfs", 0);
1836 			if (error)
1837 				goto unlock;
1838 		}
1839 	}
1840 	if (flags & V_XSLEEP)
1841 		goto unlock;
1842 	mp->mnt_writeopcount++;
1843 unlock:
1844 	if (error != 0 || (flags & V_XSLEEP) != 0)
1845 		MNT_REL(mp);
1846 	MNT_IUNLOCK(mp);
1847 	return (error);
1848 }
1849 
1850 int
1851 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1852 {
1853 	struct mount *mp;
1854 	int error;
1855 
1856 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1857 	    ("V_MNTREF requires mp"));
1858 
1859 	error = 0;
1860 	/*
1861 	 * If a vnode is provided, get and return the mount point that
1862 	 * to which it will write.
1863 	 */
1864 	if (vp != NULL) {
1865 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1866 			*mpp = NULL;
1867 			if (error != EOPNOTSUPP)
1868 				return (error);
1869 			return (0);
1870 		}
1871 	}
1872 	if ((mp = *mpp) == NULL)
1873 		return (0);
1874 
1875 	/*
1876 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1877 	 * a vfs_ref().
1878 	 * As long as a vnode is not provided we need to acquire a
1879 	 * refcount for the provided mountpoint too, in order to
1880 	 * emulate a vfs_ref().
1881 	 */
1882 	if (vp == NULL && (flags & V_MNTREF) == 0)
1883 		vfs_ref(mp);
1884 
1885 	return (vn_start_write_refed(mp, flags, false));
1886 }
1887 
1888 /*
1889  * Secondary suspension. Used by operations such as vop_inactive
1890  * routines that are needed by the higher level functions. These
1891  * are allowed to proceed until all the higher level functions have
1892  * completed (indicated by mnt_writeopcount dropping to zero). At that
1893  * time, these operations are halted until the suspension is over.
1894  */
1895 int
1896 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1897 {
1898 	struct mount *mp;
1899 	int error;
1900 
1901 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1902 	    ("V_MNTREF requires mp"));
1903 
1904  retry:
1905 	if (vp != NULL) {
1906 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1907 			*mpp = NULL;
1908 			if (error != EOPNOTSUPP)
1909 				return (error);
1910 			return (0);
1911 		}
1912 	}
1913 	/*
1914 	 * If we are not suspended or have not yet reached suspended
1915 	 * mode, then let the operation proceed.
1916 	 */
1917 	if ((mp = *mpp) == NULL)
1918 		return (0);
1919 
1920 	/*
1921 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1922 	 * a vfs_ref().
1923 	 * As long as a vnode is not provided we need to acquire a
1924 	 * refcount for the provided mountpoint too, in order to
1925 	 * emulate a vfs_ref().
1926 	 */
1927 	MNT_ILOCK(mp);
1928 	if (vp == NULL && (flags & V_MNTREF) == 0)
1929 		MNT_REF(mp);
1930 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1931 		mp->mnt_secondary_writes++;
1932 		mp->mnt_secondary_accwrites++;
1933 		MNT_IUNLOCK(mp);
1934 		return (0);
1935 	}
1936 	if (flags & V_NOWAIT) {
1937 		MNT_REL(mp);
1938 		MNT_IUNLOCK(mp);
1939 		return (EWOULDBLOCK);
1940 	}
1941 	/*
1942 	 * Wait for the suspension to finish.
1943 	 */
1944 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1945 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1946 	    "suspfs", 0);
1947 	vfs_rel(mp);
1948 	if (error == 0)
1949 		goto retry;
1950 	return (error);
1951 }
1952 
1953 /*
1954  * Filesystem write operation has completed. If we are suspending and this
1955  * operation is the last one, notify the suspender that the suspension is
1956  * now in effect.
1957  */
1958 void
1959 vn_finished_write(struct mount *mp)
1960 {
1961 	struct mount_pcpu *mpcpu;
1962 	int c;
1963 
1964 	if (mp == NULL)
1965 		return;
1966 
1967 	if (vfs_op_thread_enter(mp, mpcpu)) {
1968 		vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1);
1969 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
1970 		vfs_op_thread_exit(mp, mpcpu);
1971 		return;
1972 	}
1973 
1974 	MNT_ILOCK(mp);
1975 	vfs_assert_mount_counters(mp);
1976 	MNT_REL(mp);
1977 	c = --mp->mnt_writeopcount;
1978 	if (mp->mnt_vfs_ops == 0) {
1979 		MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0);
1980 		MNT_IUNLOCK(mp);
1981 		return;
1982 	}
1983 	if (c < 0)
1984 		vfs_dump_mount_counters(mp);
1985 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0)
1986 		wakeup(&mp->mnt_writeopcount);
1987 	MNT_IUNLOCK(mp);
1988 }
1989 
1990 /*
1991  * Filesystem secondary write operation has completed. If we are
1992  * suspending and this operation is the last one, notify the suspender
1993  * that the suspension is now in effect.
1994  */
1995 void
1996 vn_finished_secondary_write(struct mount *mp)
1997 {
1998 	if (mp == NULL)
1999 		return;
2000 	MNT_ILOCK(mp);
2001 	MNT_REL(mp);
2002 	mp->mnt_secondary_writes--;
2003 	if (mp->mnt_secondary_writes < 0)
2004 		panic("vn_finished_secondary_write: neg cnt");
2005 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
2006 	    mp->mnt_secondary_writes <= 0)
2007 		wakeup(&mp->mnt_secondary_writes);
2008 	MNT_IUNLOCK(mp);
2009 }
2010 
2011 /*
2012  * Request a filesystem to suspend write operations.
2013  */
2014 int
2015 vfs_write_suspend(struct mount *mp, int flags)
2016 {
2017 	int error;
2018 
2019 	vfs_op_enter(mp);
2020 
2021 	MNT_ILOCK(mp);
2022 	vfs_assert_mount_counters(mp);
2023 	if (mp->mnt_susp_owner == curthread) {
2024 		vfs_op_exit_locked(mp);
2025 		MNT_IUNLOCK(mp);
2026 		return (EALREADY);
2027 	}
2028 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
2029 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
2030 
2031 	/*
2032 	 * Unmount holds a write reference on the mount point.  If we
2033 	 * own busy reference and drain for writers, we deadlock with
2034 	 * the reference draining in the unmount path.  Callers of
2035 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
2036 	 * vfs_busy() reference is owned and caller is not in the
2037 	 * unmount context.
2038 	 */
2039 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
2040 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
2041 		vfs_op_exit_locked(mp);
2042 		MNT_IUNLOCK(mp);
2043 		return (EBUSY);
2044 	}
2045 
2046 	mp->mnt_kern_flag |= MNTK_SUSPEND;
2047 	mp->mnt_susp_owner = curthread;
2048 	if (mp->mnt_writeopcount > 0)
2049 		(void) msleep(&mp->mnt_writeopcount,
2050 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
2051 	else
2052 		MNT_IUNLOCK(mp);
2053 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) {
2054 		vfs_write_resume(mp, 0);
2055 		/* vfs_write_resume does vfs_op_exit() for us */
2056 	}
2057 	return (error);
2058 }
2059 
2060 /*
2061  * Request a filesystem to resume write operations.
2062  */
2063 void
2064 vfs_write_resume(struct mount *mp, int flags)
2065 {
2066 
2067 	MNT_ILOCK(mp);
2068 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
2069 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
2070 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
2071 				       MNTK_SUSPENDED);
2072 		mp->mnt_susp_owner = NULL;
2073 		wakeup(&mp->mnt_writeopcount);
2074 		wakeup(&mp->mnt_flag);
2075 		curthread->td_pflags &= ~TDP_IGNSUSP;
2076 		if ((flags & VR_START_WRITE) != 0) {
2077 			MNT_REF(mp);
2078 			mp->mnt_writeopcount++;
2079 		}
2080 		MNT_IUNLOCK(mp);
2081 		if ((flags & VR_NO_SUSPCLR) == 0)
2082 			VFS_SUSP_CLEAN(mp);
2083 		vfs_op_exit(mp);
2084 	} else if ((flags & VR_START_WRITE) != 0) {
2085 		MNT_REF(mp);
2086 		vn_start_write_refed(mp, 0, true);
2087 	} else {
2088 		MNT_IUNLOCK(mp);
2089 	}
2090 }
2091 
2092 /*
2093  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
2094  * methods.
2095  */
2096 int
2097 vfs_write_suspend_umnt(struct mount *mp)
2098 {
2099 	int error;
2100 
2101 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
2102 	    ("vfs_write_suspend_umnt: recursed"));
2103 
2104 	/* dounmount() already called vn_start_write(). */
2105 	for (;;) {
2106 		vn_finished_write(mp);
2107 		error = vfs_write_suspend(mp, 0);
2108 		if (error != 0) {
2109 			vn_start_write(NULL, &mp, V_WAIT);
2110 			return (error);
2111 		}
2112 		MNT_ILOCK(mp);
2113 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
2114 			break;
2115 		MNT_IUNLOCK(mp);
2116 		vn_start_write(NULL, &mp, V_WAIT);
2117 	}
2118 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
2119 	wakeup(&mp->mnt_flag);
2120 	MNT_IUNLOCK(mp);
2121 	curthread->td_pflags |= TDP_IGNSUSP;
2122 	return (0);
2123 }
2124 
2125 /*
2126  * Implement kqueues for files by translating it to vnode operation.
2127  */
2128 static int
2129 vn_kqfilter(struct file *fp, struct knote *kn)
2130 {
2131 
2132 	return (VOP_KQFILTER(fp->f_vnode, kn));
2133 }
2134 
2135 /*
2136  * Simplified in-kernel wrapper calls for extended attribute access.
2137  * Both calls pass in a NULL credential, authorizing as "kernel" access.
2138  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
2139  */
2140 int
2141 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
2142     const char *attrname, int *buflen, char *buf, struct thread *td)
2143 {
2144 	struct uio	auio;
2145 	struct iovec	iov;
2146 	int	error;
2147 
2148 	iov.iov_len = *buflen;
2149 	iov.iov_base = buf;
2150 
2151 	auio.uio_iov = &iov;
2152 	auio.uio_iovcnt = 1;
2153 	auio.uio_rw = UIO_READ;
2154 	auio.uio_segflg = UIO_SYSSPACE;
2155 	auio.uio_td = td;
2156 	auio.uio_offset = 0;
2157 	auio.uio_resid = *buflen;
2158 
2159 	if ((ioflg & IO_NODELOCKED) == 0)
2160 		vn_lock(vp, LK_SHARED | LK_RETRY);
2161 
2162 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2163 
2164 	/* authorize attribute retrieval as kernel */
2165 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2166 	    td);
2167 
2168 	if ((ioflg & IO_NODELOCKED) == 0)
2169 		VOP_UNLOCK(vp);
2170 
2171 	if (error == 0) {
2172 		*buflen = *buflen - auio.uio_resid;
2173 	}
2174 
2175 	return (error);
2176 }
2177 
2178 /*
2179  * XXX failure mode if partially written?
2180  */
2181 int
2182 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2183     const char *attrname, int buflen, char *buf, struct thread *td)
2184 {
2185 	struct uio	auio;
2186 	struct iovec	iov;
2187 	struct mount	*mp;
2188 	int	error;
2189 
2190 	iov.iov_len = buflen;
2191 	iov.iov_base = buf;
2192 
2193 	auio.uio_iov = &iov;
2194 	auio.uio_iovcnt = 1;
2195 	auio.uio_rw = UIO_WRITE;
2196 	auio.uio_segflg = UIO_SYSSPACE;
2197 	auio.uio_td = td;
2198 	auio.uio_offset = 0;
2199 	auio.uio_resid = buflen;
2200 
2201 	if ((ioflg & IO_NODELOCKED) == 0) {
2202 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2203 			return (error);
2204 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2205 	}
2206 
2207 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2208 
2209 	/* authorize attribute setting as kernel */
2210 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2211 
2212 	if ((ioflg & IO_NODELOCKED) == 0) {
2213 		vn_finished_write(mp);
2214 		VOP_UNLOCK(vp);
2215 	}
2216 
2217 	return (error);
2218 }
2219 
2220 int
2221 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2222     const char *attrname, struct thread *td)
2223 {
2224 	struct mount	*mp;
2225 	int	error;
2226 
2227 	if ((ioflg & IO_NODELOCKED) == 0) {
2228 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2229 			return (error);
2230 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2231 	}
2232 
2233 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2234 
2235 	/* authorize attribute removal as kernel */
2236 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2237 	if (error == EOPNOTSUPP)
2238 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2239 		    NULL, td);
2240 
2241 	if ((ioflg & IO_NODELOCKED) == 0) {
2242 		vn_finished_write(mp);
2243 		VOP_UNLOCK(vp);
2244 	}
2245 
2246 	return (error);
2247 }
2248 
2249 static int
2250 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2251     struct vnode **rvp)
2252 {
2253 
2254 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2255 }
2256 
2257 int
2258 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2259 {
2260 
2261 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2262 	    lkflags, rvp));
2263 }
2264 
2265 int
2266 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2267     int lkflags, struct vnode **rvp)
2268 {
2269 	struct mount *mp;
2270 	int ltype, error;
2271 
2272 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2273 	mp = vp->v_mount;
2274 	ltype = VOP_ISLOCKED(vp);
2275 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2276 	    ("vn_vget_ino: vp not locked"));
2277 	error = vfs_busy(mp, MBF_NOWAIT);
2278 	if (error != 0) {
2279 		vfs_ref(mp);
2280 		VOP_UNLOCK(vp);
2281 		error = vfs_busy(mp, 0);
2282 		vn_lock(vp, ltype | LK_RETRY);
2283 		vfs_rel(mp);
2284 		if (error != 0)
2285 			return (ENOENT);
2286 		if (VN_IS_DOOMED(vp)) {
2287 			vfs_unbusy(mp);
2288 			return (ENOENT);
2289 		}
2290 	}
2291 	VOP_UNLOCK(vp);
2292 	error = alloc(mp, alloc_arg, lkflags, rvp);
2293 	vfs_unbusy(mp);
2294 	if (error != 0 || *rvp != vp)
2295 		vn_lock(vp, ltype | LK_RETRY);
2296 	if (VN_IS_DOOMED(vp)) {
2297 		if (error == 0) {
2298 			if (*rvp == vp)
2299 				vunref(vp);
2300 			else
2301 				vput(*rvp);
2302 		}
2303 		error = ENOENT;
2304 	}
2305 	return (error);
2306 }
2307 
2308 int
2309 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2310     struct thread *td)
2311 {
2312 
2313 	if (vp->v_type != VREG || td == NULL)
2314 		return (0);
2315 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2316 	    lim_cur(td, RLIMIT_FSIZE)) {
2317 		PROC_LOCK(td->td_proc);
2318 		kern_psignal(td->td_proc, SIGXFSZ);
2319 		PROC_UNLOCK(td->td_proc);
2320 		return (EFBIG);
2321 	}
2322 	return (0);
2323 }
2324 
2325 int
2326 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2327     struct thread *td)
2328 {
2329 	struct vnode *vp;
2330 
2331 	vp = fp->f_vnode;
2332 #ifdef AUDIT
2333 	vn_lock(vp, LK_SHARED | LK_RETRY);
2334 	AUDIT_ARG_VNODE1(vp);
2335 	VOP_UNLOCK(vp);
2336 #endif
2337 	return (setfmode(td, active_cred, vp, mode));
2338 }
2339 
2340 int
2341 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2342     struct thread *td)
2343 {
2344 	struct vnode *vp;
2345 
2346 	vp = fp->f_vnode;
2347 #ifdef AUDIT
2348 	vn_lock(vp, LK_SHARED | LK_RETRY);
2349 	AUDIT_ARG_VNODE1(vp);
2350 	VOP_UNLOCK(vp);
2351 #endif
2352 	return (setfown(td, active_cred, vp, uid, gid));
2353 }
2354 
2355 void
2356 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2357 {
2358 	vm_object_t object;
2359 
2360 	if ((object = vp->v_object) == NULL)
2361 		return;
2362 	VM_OBJECT_WLOCK(object);
2363 	vm_object_page_remove(object, start, end, 0);
2364 	VM_OBJECT_WUNLOCK(object);
2365 }
2366 
2367 int
2368 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2369 {
2370 	struct vattr va;
2371 	daddr_t bn, bnp;
2372 	uint64_t bsize;
2373 	off_t noff;
2374 	int error;
2375 
2376 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2377 	    ("Wrong command %lu", cmd));
2378 
2379 	if (vn_lock(vp, LK_SHARED) != 0)
2380 		return (EBADF);
2381 	if (vp->v_type != VREG) {
2382 		error = ENOTTY;
2383 		goto unlock;
2384 	}
2385 	error = VOP_GETATTR(vp, &va, cred);
2386 	if (error != 0)
2387 		goto unlock;
2388 	noff = *off;
2389 	if (noff >= va.va_size) {
2390 		error = ENXIO;
2391 		goto unlock;
2392 	}
2393 	bsize = vp->v_mount->mnt_stat.f_iosize;
2394 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize -
2395 	    noff % bsize) {
2396 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2397 		if (error == EOPNOTSUPP) {
2398 			error = ENOTTY;
2399 			goto unlock;
2400 		}
2401 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2402 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2403 			noff = bn * bsize;
2404 			if (noff < *off)
2405 				noff = *off;
2406 			goto unlock;
2407 		}
2408 	}
2409 	if (noff > va.va_size)
2410 		noff = va.va_size;
2411 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2412 	if (cmd == FIOSEEKDATA)
2413 		error = ENXIO;
2414 unlock:
2415 	VOP_UNLOCK(vp);
2416 	if (error == 0)
2417 		*off = noff;
2418 	return (error);
2419 }
2420 
2421 int
2422 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2423 {
2424 	struct ucred *cred;
2425 	struct vnode *vp;
2426 	struct vattr vattr;
2427 	off_t foffset, size;
2428 	int error, noneg;
2429 
2430 	cred = td->td_ucred;
2431 	vp = fp->f_vnode;
2432 	foffset = foffset_lock(fp, 0);
2433 	noneg = (vp->v_type != VCHR);
2434 	error = 0;
2435 	switch (whence) {
2436 	case L_INCR:
2437 		if (noneg &&
2438 		    (foffset < 0 ||
2439 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2440 			error = EOVERFLOW;
2441 			break;
2442 		}
2443 		offset += foffset;
2444 		break;
2445 	case L_XTND:
2446 		vn_lock(vp, LK_SHARED | LK_RETRY);
2447 		error = VOP_GETATTR(vp, &vattr, cred);
2448 		VOP_UNLOCK(vp);
2449 		if (error)
2450 			break;
2451 
2452 		/*
2453 		 * If the file references a disk device, then fetch
2454 		 * the media size and use that to determine the ending
2455 		 * offset.
2456 		 */
2457 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2458 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2459 			vattr.va_size = size;
2460 		if (noneg &&
2461 		    (vattr.va_size > OFF_MAX ||
2462 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2463 			error = EOVERFLOW;
2464 			break;
2465 		}
2466 		offset += vattr.va_size;
2467 		break;
2468 	case L_SET:
2469 		break;
2470 	case SEEK_DATA:
2471 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2472 		if (error == ENOTTY)
2473 			error = EINVAL;
2474 		break;
2475 	case SEEK_HOLE:
2476 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2477 		if (error == ENOTTY)
2478 			error = EINVAL;
2479 		break;
2480 	default:
2481 		error = EINVAL;
2482 	}
2483 	if (error == 0 && noneg && offset < 0)
2484 		error = EINVAL;
2485 	if (error != 0)
2486 		goto drop;
2487 	VFS_KNOTE_UNLOCKED(vp, 0);
2488 	td->td_uretoff.tdu_off = offset;
2489 drop:
2490 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2491 	return (error);
2492 }
2493 
2494 int
2495 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2496     struct thread *td)
2497 {
2498 	int error;
2499 
2500 	/*
2501 	 * Grant permission if the caller is the owner of the file, or
2502 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2503 	 * on the file.  If the time pointer is null, then write
2504 	 * permission on the file is also sufficient.
2505 	 *
2506 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2507 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2508 	 * will be allowed to set the times [..] to the current
2509 	 * server time.
2510 	 */
2511 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2512 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2513 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2514 	return (error);
2515 }
2516 
2517 int
2518 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2519 {
2520 	struct vnode *vp;
2521 	int error;
2522 
2523 	if (fp->f_type == DTYPE_FIFO)
2524 		kif->kf_type = KF_TYPE_FIFO;
2525 	else
2526 		kif->kf_type = KF_TYPE_VNODE;
2527 	vp = fp->f_vnode;
2528 	vref(vp);
2529 	FILEDESC_SUNLOCK(fdp);
2530 	error = vn_fill_kinfo_vnode(vp, kif);
2531 	vrele(vp);
2532 	FILEDESC_SLOCK(fdp);
2533 	return (error);
2534 }
2535 
2536 static inline void
2537 vn_fill_junk(struct kinfo_file *kif)
2538 {
2539 	size_t len, olen;
2540 
2541 	/*
2542 	 * Simulate vn_fullpath returning changing values for a given
2543 	 * vp during e.g. coredump.
2544 	 */
2545 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2546 	olen = strlen(kif->kf_path);
2547 	if (len < olen)
2548 		strcpy(&kif->kf_path[len - 1], "$");
2549 	else
2550 		for (; olen < len; olen++)
2551 			strcpy(&kif->kf_path[olen], "A");
2552 }
2553 
2554 int
2555 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2556 {
2557 	struct vattr va;
2558 	char *fullpath, *freepath;
2559 	int error;
2560 
2561 	kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type);
2562 	freepath = NULL;
2563 	fullpath = "-";
2564 	error = vn_fullpath(vp, &fullpath, &freepath);
2565 	if (error == 0) {
2566 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2567 	}
2568 	if (freepath != NULL)
2569 		free(freepath, M_TEMP);
2570 
2571 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2572 		vn_fill_junk(kif);
2573 	);
2574 
2575 	/*
2576 	 * Retrieve vnode attributes.
2577 	 */
2578 	va.va_fsid = VNOVAL;
2579 	va.va_rdev = NODEV;
2580 	vn_lock(vp, LK_SHARED | LK_RETRY);
2581 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2582 	VOP_UNLOCK(vp);
2583 	if (error != 0)
2584 		return (error);
2585 	if (va.va_fsid != VNOVAL)
2586 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2587 	else
2588 		kif->kf_un.kf_file.kf_file_fsid =
2589 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2590 	kif->kf_un.kf_file.kf_file_fsid_freebsd11 =
2591 	    kif->kf_un.kf_file.kf_file_fsid; /* truncate */
2592 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2593 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2594 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2595 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2596 	kif->kf_un.kf_file.kf_file_rdev_freebsd11 =
2597 	    kif->kf_un.kf_file.kf_file_rdev; /* truncate */
2598 	return (0);
2599 }
2600 
2601 int
2602 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2603     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2604     struct thread *td)
2605 {
2606 #ifdef HWPMC_HOOKS
2607 	struct pmckern_map_in pkm;
2608 #endif
2609 	struct mount *mp;
2610 	struct vnode *vp;
2611 	vm_object_t object;
2612 	vm_prot_t maxprot;
2613 	boolean_t writecounted;
2614 	int error;
2615 
2616 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2617     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2618 	/*
2619 	 * POSIX shared-memory objects are defined to have
2620 	 * kernel persistence, and are not defined to support
2621 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2622 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2623 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2624 	 * flag to request this behavior.
2625 	 */
2626 	if ((fp->f_flag & FPOSIXSHM) != 0)
2627 		flags |= MAP_NOSYNC;
2628 #endif
2629 	vp = fp->f_vnode;
2630 
2631 	/*
2632 	 * Ensure that file and memory protections are
2633 	 * compatible.  Note that we only worry about
2634 	 * writability if mapping is shared; in this case,
2635 	 * current and max prot are dictated by the open file.
2636 	 * XXX use the vnode instead?  Problem is: what
2637 	 * credentials do we use for determination? What if
2638 	 * proc does a setuid?
2639 	 */
2640 	mp = vp->v_mount;
2641 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) {
2642 		maxprot = VM_PROT_NONE;
2643 		if ((prot & VM_PROT_EXECUTE) != 0)
2644 			return (EACCES);
2645 	} else
2646 		maxprot = VM_PROT_EXECUTE;
2647 	if ((fp->f_flag & FREAD) != 0)
2648 		maxprot |= VM_PROT_READ;
2649 	else if ((prot & VM_PROT_READ) != 0)
2650 		return (EACCES);
2651 
2652 	/*
2653 	 * If we are sharing potential changes via MAP_SHARED and we
2654 	 * are trying to get write permission although we opened it
2655 	 * without asking for it, bail out.
2656 	 */
2657 	if ((flags & MAP_SHARED) != 0) {
2658 		if ((fp->f_flag & FWRITE) != 0)
2659 			maxprot |= VM_PROT_WRITE;
2660 		else if ((prot & VM_PROT_WRITE) != 0)
2661 			return (EACCES);
2662 	} else {
2663 		maxprot |= VM_PROT_WRITE;
2664 		cap_maxprot |= VM_PROT_WRITE;
2665 	}
2666 	maxprot &= cap_maxprot;
2667 
2668 	/*
2669 	 * For regular files and shared memory, POSIX requires that
2670 	 * the value of foff be a legitimate offset within the data
2671 	 * object.  In particular, negative offsets are invalid.
2672 	 * Blocking negative offsets and overflows here avoids
2673 	 * possible wraparound or user-level access into reserved
2674 	 * ranges of the data object later.  In contrast, POSIX does
2675 	 * not dictate how offsets are used by device drivers, so in
2676 	 * the case of a device mapping a negative offset is passed
2677 	 * on.
2678 	 */
2679 	if (
2680 #ifdef _LP64
2681 	    size > OFF_MAX ||
2682 #endif
2683 	    foff > OFF_MAX - size)
2684 		return (EINVAL);
2685 
2686 	writecounted = FALSE;
2687 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2688 	    &foff, &object, &writecounted);
2689 	if (error != 0)
2690 		return (error);
2691 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2692 	    foff, writecounted, td);
2693 	if (error != 0) {
2694 		/*
2695 		 * If this mapping was accounted for in the vnode's
2696 		 * writecount, then undo that now.
2697 		 */
2698 		if (writecounted)
2699 			vm_pager_release_writecount(object, 0, size);
2700 		vm_object_deallocate(object);
2701 	}
2702 #ifdef HWPMC_HOOKS
2703 	/* Inform hwpmc(4) if an executable is being mapped. */
2704 	if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) {
2705 		if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) {
2706 			pkm.pm_file = vp;
2707 			pkm.pm_address = (uintptr_t) *addr;
2708 			PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm);
2709 		}
2710 	}
2711 #endif
2712 	return (error);
2713 }
2714 
2715 void
2716 vn_fsid(struct vnode *vp, struct vattr *va)
2717 {
2718 	fsid_t *f;
2719 
2720 	f = &vp->v_mount->mnt_stat.f_fsid;
2721 	va->va_fsid = (uint32_t)f->val[1];
2722 	va->va_fsid <<= sizeof(f->val[1]) * NBBY;
2723 	va->va_fsid += (uint32_t)f->val[0];
2724 }
2725 
2726 int
2727 vn_fsync_buf(struct vnode *vp, int waitfor)
2728 {
2729 	struct buf *bp, *nbp;
2730 	struct bufobj *bo;
2731 	struct mount *mp;
2732 	int error, maxretry;
2733 
2734 	error = 0;
2735 	maxretry = 10000;     /* large, arbitrarily chosen */
2736 	mp = NULL;
2737 	if (vp->v_type == VCHR) {
2738 		VI_LOCK(vp);
2739 		mp = vp->v_rdev->si_mountpt;
2740 		VI_UNLOCK(vp);
2741 	}
2742 	bo = &vp->v_bufobj;
2743 	BO_LOCK(bo);
2744 loop1:
2745 	/*
2746 	 * MARK/SCAN initialization to avoid infinite loops.
2747 	 */
2748         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2749 		bp->b_vflags &= ~BV_SCANNED;
2750 		bp->b_error = 0;
2751 	}
2752 
2753 	/*
2754 	 * Flush all dirty buffers associated with a vnode.
2755 	 */
2756 loop2:
2757 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2758 		if ((bp->b_vflags & BV_SCANNED) != 0)
2759 			continue;
2760 		bp->b_vflags |= BV_SCANNED;
2761 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) {
2762 			if (waitfor != MNT_WAIT)
2763 				continue;
2764 			if (BUF_LOCK(bp,
2765 			    LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL,
2766 			    BO_LOCKPTR(bo)) != 0) {
2767 				BO_LOCK(bo);
2768 				goto loop1;
2769 			}
2770 			BO_LOCK(bo);
2771 		}
2772 		BO_UNLOCK(bo);
2773 		KASSERT(bp->b_bufobj == bo,
2774 		    ("bp %p wrong b_bufobj %p should be %p",
2775 		    bp, bp->b_bufobj, bo));
2776 		if ((bp->b_flags & B_DELWRI) == 0)
2777 			panic("fsync: not dirty");
2778 		if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) {
2779 			vfs_bio_awrite(bp);
2780 		} else {
2781 			bremfree(bp);
2782 			bawrite(bp);
2783 		}
2784 		if (maxretry < 1000)
2785 			pause("dirty", hz < 1000 ? 1 : hz / 1000);
2786 		BO_LOCK(bo);
2787 		goto loop2;
2788 	}
2789 
2790 	/*
2791 	 * If synchronous the caller expects us to completely resolve all
2792 	 * dirty buffers in the system.  Wait for in-progress I/O to
2793 	 * complete (which could include background bitmap writes), then
2794 	 * retry if dirty blocks still exist.
2795 	 */
2796 	if (waitfor == MNT_WAIT) {
2797 		bufobj_wwait(bo, 0, 0);
2798 		if (bo->bo_dirty.bv_cnt > 0) {
2799 			/*
2800 			 * If we are unable to write any of these buffers
2801 			 * then we fail now rather than trying endlessly
2802 			 * to write them out.
2803 			 */
2804 			TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
2805 				if ((error = bp->b_error) != 0)
2806 					break;
2807 			if ((mp != NULL && mp->mnt_secondary_writes > 0) ||
2808 			    (error == 0 && --maxretry >= 0))
2809 				goto loop1;
2810 			if (error == 0)
2811 				error = EAGAIN;
2812 		}
2813 	}
2814 	BO_UNLOCK(bo);
2815 	if (error != 0)
2816 		vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error);
2817 
2818 	return (error);
2819 }
2820 
2821 /*
2822  * Copies a byte range from invp to outvp.  Calls VOP_COPY_FILE_RANGE()
2823  * or vn_generic_copy_file_range() after rangelocking the byte ranges,
2824  * to do the actual copy.
2825  * vn_generic_copy_file_range() is factored out, so it can be called
2826  * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from
2827  * different file systems.
2828  */
2829 int
2830 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp,
2831     off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred,
2832     struct ucred *outcred, struct thread *fsize_td)
2833 {
2834 	int error;
2835 	size_t len;
2836 	uint64_t uval;
2837 
2838 	len = *lenp;
2839 	*lenp = 0;		/* For error returns. */
2840 	error = 0;
2841 
2842 	/* Do some sanity checks on the arguments. */
2843 	if (invp->v_type == VDIR || outvp->v_type == VDIR)
2844 		error = EISDIR;
2845 	else if (*inoffp < 0 || *outoffp < 0 ||
2846 	    invp->v_type != VREG || outvp->v_type != VREG)
2847 		error = EINVAL;
2848 	if (error != 0)
2849 		goto out;
2850 
2851 	/* Ensure offset + len does not wrap around. */
2852 	uval = *inoffp;
2853 	uval += len;
2854 	if (uval > INT64_MAX)
2855 		len = INT64_MAX - *inoffp;
2856 	uval = *outoffp;
2857 	uval += len;
2858 	if (uval > INT64_MAX)
2859 		len = INT64_MAX - *outoffp;
2860 	if (len == 0)
2861 		goto out;
2862 
2863 	/*
2864 	 * If the two vnode are for the same file system, call
2865 	 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range()
2866 	 * which can handle copies across multiple file systems.
2867 	 */
2868 	*lenp = len;
2869 	if (invp->v_mount == outvp->v_mount)
2870 		error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp,
2871 		    lenp, flags, incred, outcred, fsize_td);
2872 	else
2873 		error = vn_generic_copy_file_range(invp, inoffp, outvp,
2874 		    outoffp, lenp, flags, incred, outcred, fsize_td);
2875 out:
2876 	return (error);
2877 }
2878 
2879 /*
2880  * Test len bytes of data starting at dat for all bytes == 0.
2881  * Return true if all bytes are zero, false otherwise.
2882  * Expects dat to be well aligned.
2883  */
2884 static bool
2885 mem_iszero(void *dat, int len)
2886 {
2887 	int i;
2888 	const u_int *p;
2889 	const char *cp;
2890 
2891 	for (p = dat; len > 0; len -= sizeof(*p), p++) {
2892 		if (len >= sizeof(*p)) {
2893 			if (*p != 0)
2894 				return (false);
2895 		} else {
2896 			cp = (const char *)p;
2897 			for (i = 0; i < len; i++, cp++)
2898 				if (*cp != '\0')
2899 					return (false);
2900 		}
2901 	}
2902 	return (true);
2903 }
2904 
2905 /*
2906  * Look for a hole in the output file and, if found, adjust *outoffp
2907  * and *xferp to skip past the hole.
2908  * *xferp is the entire hole length to be written and xfer2 is how many bytes
2909  * to be written as 0's upon return.
2910  */
2911 static off_t
2912 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp,
2913     off_t *dataoffp, off_t *holeoffp, struct ucred *cred)
2914 {
2915 	int error;
2916 	off_t delta;
2917 
2918 	if (*holeoffp == 0 || *holeoffp <= *outoffp) {
2919 		*dataoffp = *outoffp;
2920 		error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred,
2921 		    curthread);
2922 		if (error == 0) {
2923 			*holeoffp = *dataoffp;
2924 			error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred,
2925 			    curthread);
2926 		}
2927 		if (error != 0 || *holeoffp == *dataoffp) {
2928 			/*
2929 			 * Since outvp is unlocked, it may be possible for
2930 			 * another thread to do a truncate(), lseek(), write()
2931 			 * creating a hole at startoff between the above
2932 			 * VOP_IOCTL() calls, if the other thread does not do
2933 			 * rangelocking.
2934 			 * If that happens, *holeoffp == *dataoffp and finding
2935 			 * the hole has failed, so disable vn_skip_hole().
2936 			 */
2937 			*holeoffp = -1;	/* Disable use of vn_skip_hole(). */
2938 			return (xfer2);
2939 		}
2940 		KASSERT(*dataoffp >= *outoffp,
2941 		    ("vn_skip_hole: dataoff=%jd < outoff=%jd",
2942 		    (intmax_t)*dataoffp, (intmax_t)*outoffp));
2943 		KASSERT(*holeoffp > *dataoffp,
2944 		    ("vn_skip_hole: holeoff=%jd <= dataoff=%jd",
2945 		    (intmax_t)*holeoffp, (intmax_t)*dataoffp));
2946 	}
2947 
2948 	/*
2949 	 * If there is a hole before the data starts, advance *outoffp and
2950 	 * *xferp past the hole.
2951 	 */
2952 	if (*dataoffp > *outoffp) {
2953 		delta = *dataoffp - *outoffp;
2954 		if (delta >= *xferp) {
2955 			/* Entire *xferp is a hole. */
2956 			*outoffp += *xferp;
2957 			*xferp = 0;
2958 			return (0);
2959 		}
2960 		*xferp -= delta;
2961 		*outoffp += delta;
2962 		xfer2 = MIN(xfer2, *xferp);
2963 	}
2964 
2965 	/*
2966 	 * If a hole starts before the end of this xfer2, reduce this xfer2 so
2967 	 * that the write ends at the start of the hole.
2968 	 * *holeoffp should always be greater than *outoffp, but for the
2969 	 * non-INVARIANTS case, check this to make sure xfer2 remains a sane
2970 	 * value.
2971 	 */
2972 	if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2)
2973 		xfer2 = *holeoffp - *outoffp;
2974 	return (xfer2);
2975 }
2976 
2977 /*
2978  * Write an xfer sized chunk to outvp in blksize blocks from dat.
2979  * dat is a maximum of blksize in length and can be written repeatedly in
2980  * the chunk.
2981  * If growfile == true, just grow the file via vn_truncate_locked() instead
2982  * of doing actual writes.
2983  * If checkhole == true, a hole is being punched, so skip over any hole
2984  * already in the output file.
2985  */
2986 static int
2987 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer,
2988     u_long blksize, bool growfile, bool checkhole, struct ucred *cred)
2989 {
2990 	struct mount *mp;
2991 	off_t dataoff, holeoff, xfer2;
2992 	int error, lckf;
2993 
2994 	/*
2995 	 * Loop around doing writes of blksize until write has been completed.
2996 	 * Lock/unlock on each loop iteration so that a bwillwrite() can be
2997 	 * done for each iteration, since the xfer argument can be very
2998 	 * large if there is a large hole to punch in the output file.
2999 	 */
3000 	error = 0;
3001 	holeoff = 0;
3002 	do {
3003 		xfer2 = MIN(xfer, blksize);
3004 		if (checkhole) {
3005 			/*
3006 			 * Punching a hole.  Skip writing if there is
3007 			 * already a hole in the output file.
3008 			 */
3009 			xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer,
3010 			    &dataoff, &holeoff, cred);
3011 			if (xfer == 0)
3012 				break;
3013 			if (holeoff < 0)
3014 				checkhole = false;
3015 			KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd",
3016 			    (intmax_t)xfer2));
3017 		}
3018 		bwillwrite();
3019 		mp = NULL;
3020 		error = vn_start_write(outvp, &mp, V_WAIT);
3021 		if (error != 0)
3022 			break;
3023 		if (growfile) {
3024 			error = vn_lock(outvp, LK_EXCLUSIVE);
3025 			if (error == 0) {
3026 				error = vn_truncate_locked(outvp, outoff + xfer,
3027 				    false, cred);
3028 				VOP_UNLOCK(outvp);
3029 			}
3030 		} else {
3031 			if (MNT_SHARED_WRITES(mp))
3032 				lckf = LK_SHARED;
3033 			else
3034 				lckf = LK_EXCLUSIVE;
3035 			error = vn_lock(outvp, lckf);
3036 			if (error == 0) {
3037 				error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2,
3038 				    outoff, UIO_SYSSPACE, IO_NODELOCKED,
3039 				    curthread->td_ucred, cred, NULL, curthread);
3040 				outoff += xfer2;
3041 				xfer -= xfer2;
3042 				VOP_UNLOCK(outvp);
3043 			}
3044 		}
3045 		if (mp != NULL)
3046 			vn_finished_write(mp);
3047 	} while (!growfile && xfer > 0 && error == 0);
3048 	return (error);
3049 }
3050 
3051 /*
3052  * Copy a byte range of one file to another.  This function can handle the
3053  * case where invp and outvp are on different file systems.
3054  * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there
3055  * is no better file system specific way to do it.
3056  */
3057 int
3058 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp,
3059     struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags,
3060     struct ucred *incred, struct ucred *outcred, struct thread *fsize_td)
3061 {
3062 	struct vattr va;
3063 	struct mount *mp;
3064 	struct uio io;
3065 	off_t startoff, endoff, xfer, xfer2;
3066 	u_long blksize;
3067 	int error, interrupted;
3068 	bool cantseek, readzeros, eof, lastblock;
3069 	ssize_t aresid;
3070 	size_t copylen, len, rem, savlen;
3071 	char *dat;
3072 	long holein, holeout;
3073 
3074 	holein = holeout = 0;
3075 	savlen = len = *lenp;
3076 	error = 0;
3077 	interrupted = 0;
3078 	dat = NULL;
3079 
3080 	error = vn_lock(invp, LK_SHARED);
3081 	if (error != 0)
3082 		goto out;
3083 	if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0)
3084 		holein = 0;
3085 	VOP_UNLOCK(invp);
3086 
3087 	mp = NULL;
3088 	error = vn_start_write(outvp, &mp, V_WAIT);
3089 	if (error == 0)
3090 		error = vn_lock(outvp, LK_EXCLUSIVE);
3091 	if (error == 0) {
3092 		/*
3093 		 * If fsize_td != NULL, do a vn_rlimit_fsize() call,
3094 		 * now that outvp is locked.
3095 		 */
3096 		if (fsize_td != NULL) {
3097 			io.uio_offset = *outoffp;
3098 			io.uio_resid = len;
3099 			error = vn_rlimit_fsize(outvp, &io, fsize_td);
3100 			if (error != 0)
3101 				error = EFBIG;
3102 		}
3103 		if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0)
3104 			holeout = 0;
3105 		/*
3106 		 * Holes that are past EOF do not need to be written as a block
3107 		 * of zero bytes.  So, truncate the output file as far as
3108 		 * possible and then use va.va_size to decide if writing 0
3109 		 * bytes is necessary in the loop below.
3110 		 */
3111 		if (error == 0)
3112 			error = VOP_GETATTR(outvp, &va, outcred);
3113 		if (error == 0 && va.va_size > *outoffp && va.va_size <=
3114 		    *outoffp + len) {
3115 #ifdef MAC
3116 			error = mac_vnode_check_write(curthread->td_ucred,
3117 			    outcred, outvp);
3118 			if (error == 0)
3119 #endif
3120 				error = vn_truncate_locked(outvp, *outoffp,
3121 				    false, outcred);
3122 			if (error == 0)
3123 				va.va_size = *outoffp;
3124 		}
3125 		VOP_UNLOCK(outvp);
3126 	}
3127 	if (mp != NULL)
3128 		vn_finished_write(mp);
3129 	if (error != 0)
3130 		goto out;
3131 
3132 	/*
3133 	 * Set the blksize to the larger of the hole sizes for invp and outvp.
3134 	 * If hole sizes aren't available, set the blksize to the larger
3135 	 * f_iosize of invp and outvp.
3136 	 * This code expects the hole sizes and f_iosizes to be powers of 2.
3137 	 * This value is clipped at 4Kbytes and 1Mbyte.
3138 	 */
3139 	blksize = MAX(holein, holeout);
3140 
3141 	/* Clip len to end at an exact multiple of hole size. */
3142 	if (blksize > 1) {
3143 		rem = *inoffp % blksize;
3144 		if (rem > 0)
3145 			rem = blksize - rem;
3146 		if (len - rem > blksize)
3147 			len = savlen = rounddown(len - rem, blksize) + rem;
3148 	}
3149 
3150 	if (blksize <= 1)
3151 		blksize = MAX(invp->v_mount->mnt_stat.f_iosize,
3152 		    outvp->v_mount->mnt_stat.f_iosize);
3153 	if (blksize < 4096)
3154 		blksize = 4096;
3155 	else if (blksize > 1024 * 1024)
3156 		blksize = 1024 * 1024;
3157 	dat = malloc(blksize, M_TEMP, M_WAITOK);
3158 
3159 	/*
3160 	 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA
3161 	 * to find holes.  Otherwise, just scan the read block for all 0s
3162 	 * in the inner loop where the data copying is done.
3163 	 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may
3164 	 * support holes on the server, but do not support FIOSEEKHOLE.
3165 	 */
3166 	eof = false;
3167 	while (len > 0 && error == 0 && !eof && interrupted == 0) {
3168 		endoff = 0;			/* To shut up compilers. */
3169 		cantseek = true;
3170 		startoff = *inoffp;
3171 		copylen = len;
3172 
3173 		/*
3174 		 * Find the next data area.  If there is just a hole to EOF,
3175 		 * FIOSEEKDATA should fail and then we drop down into the
3176 		 * inner loop and create the hole on the outvp file.
3177 		 * (I do not know if any file system will report a hole to
3178 		 *  EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA
3179 		 *  will fail for those file systems.)
3180 		 *
3181 		 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE,
3182 		 * the code just falls through to the inner copy loop.
3183 		 */
3184 		error = EINVAL;
3185 		if (holein > 0)
3186 			error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0,
3187 			    incred, curthread);
3188 		if (error == 0) {
3189 			endoff = startoff;
3190 			error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0,
3191 			    incred, curthread);
3192 			/*
3193 			 * Since invp is unlocked, it may be possible for
3194 			 * another thread to do a truncate(), lseek(), write()
3195 			 * creating a hole at startoff between the above
3196 			 * VOP_IOCTL() calls, if the other thread does not do
3197 			 * rangelocking.
3198 			 * If that happens, startoff == endoff and finding
3199 			 * the hole has failed, so set an error.
3200 			 */
3201 			if (error == 0 && startoff == endoff)
3202 				error = EINVAL; /* Any error. Reset to 0. */
3203 		}
3204 		if (error == 0) {
3205 			if (startoff > *inoffp) {
3206 				/* Found hole before data block. */
3207 				xfer = MIN(startoff - *inoffp, len);
3208 				if (*outoffp < va.va_size) {
3209 					/* Must write 0s to punch hole. */
3210 					xfer2 = MIN(va.va_size - *outoffp,
3211 					    xfer);
3212 					memset(dat, 0, MIN(xfer2, blksize));
3213 					error = vn_write_outvp(outvp, dat,
3214 					    *outoffp, xfer2, blksize, false,
3215 					    holeout > 0, outcred);
3216 				}
3217 
3218 				if (error == 0 && *outoffp + xfer >
3219 				    va.va_size && xfer == len)
3220 					/* Grow last block. */
3221 					error = vn_write_outvp(outvp, dat,
3222 					    *outoffp, xfer, blksize, true,
3223 					    false, outcred);
3224 				if (error == 0) {
3225 					*inoffp += xfer;
3226 					*outoffp += xfer;
3227 					len -= xfer;
3228 					if (len < savlen)
3229 						interrupted = sig_intr();
3230 				}
3231 			}
3232 			copylen = MIN(len, endoff - startoff);
3233 			cantseek = false;
3234 		} else {
3235 			cantseek = true;
3236 			startoff = *inoffp;
3237 			copylen = len;
3238 			error = 0;
3239 		}
3240 
3241 		xfer = blksize;
3242 		if (cantseek) {
3243 			/*
3244 			 * Set first xfer to end at a block boundary, so that
3245 			 * holes are more likely detected in the loop below via
3246 			 * the for all bytes 0 method.
3247 			 */
3248 			xfer -= (*inoffp % blksize);
3249 		}
3250 		/* Loop copying the data block. */
3251 		while (copylen > 0 && error == 0 && !eof && interrupted == 0) {
3252 			if (copylen < xfer)
3253 				xfer = copylen;
3254 			error = vn_lock(invp, LK_SHARED);
3255 			if (error != 0)
3256 				goto out;
3257 			error = vn_rdwr(UIO_READ, invp, dat, xfer,
3258 			    startoff, UIO_SYSSPACE, IO_NODELOCKED,
3259 			    curthread->td_ucred, incred, &aresid,
3260 			    curthread);
3261 			VOP_UNLOCK(invp);
3262 			lastblock = false;
3263 			if (error == 0 && aresid > 0) {
3264 				/* Stop the copy at EOF on the input file. */
3265 				xfer -= aresid;
3266 				eof = true;
3267 				lastblock = true;
3268 			}
3269 			if (error == 0) {
3270 				/*
3271 				 * Skip the write for holes past the initial EOF
3272 				 * of the output file, unless this is the last
3273 				 * write of the output file at EOF.
3274 				 */
3275 				readzeros = cantseek ? mem_iszero(dat, xfer) :
3276 				    false;
3277 				if (xfer == len)
3278 					lastblock = true;
3279 				if (!cantseek || *outoffp < va.va_size ||
3280 				    lastblock || !readzeros)
3281 					error = vn_write_outvp(outvp, dat,
3282 					    *outoffp, xfer, blksize,
3283 					    readzeros && lastblock &&
3284 					    *outoffp >= va.va_size, false,
3285 					    outcred);
3286 				if (error == 0) {
3287 					*inoffp += xfer;
3288 					startoff += xfer;
3289 					*outoffp += xfer;
3290 					copylen -= xfer;
3291 					len -= xfer;
3292 					if (len < savlen)
3293 						interrupted = sig_intr();
3294 				}
3295 			}
3296 			xfer = blksize;
3297 		}
3298 	}
3299 out:
3300 	*lenp = savlen - len;
3301 	free(dat, M_TEMP);
3302 	return (error);
3303 }
3304 
3305 static int
3306 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
3307 {
3308 	struct mount *mp;
3309 	struct vnode *vp;
3310 	off_t olen, ooffset;
3311 	int error;
3312 #ifdef AUDIT
3313 	int audited_vnode1 = 0;
3314 #endif
3315 
3316 	vp = fp->f_vnode;
3317 	if (vp->v_type != VREG)
3318 		return (ENODEV);
3319 
3320 	/* Allocating blocks may take a long time, so iterate. */
3321 	for (;;) {
3322 		olen = len;
3323 		ooffset = offset;
3324 
3325 		bwillwrite();
3326 		mp = NULL;
3327 		error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
3328 		if (error != 0)
3329 			break;
3330 		error = vn_lock(vp, LK_EXCLUSIVE);
3331 		if (error != 0) {
3332 			vn_finished_write(mp);
3333 			break;
3334 		}
3335 #ifdef AUDIT
3336 		if (!audited_vnode1) {
3337 			AUDIT_ARG_VNODE1(vp);
3338 			audited_vnode1 = 1;
3339 		}
3340 #endif
3341 #ifdef MAC
3342 		error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp);
3343 		if (error == 0)
3344 #endif
3345 			error = VOP_ALLOCATE(vp, &offset, &len);
3346 		VOP_UNLOCK(vp);
3347 		vn_finished_write(mp);
3348 
3349 		if (olen + ooffset != offset + len) {
3350 			panic("offset + len changed from %jx/%jx to %jx/%jx",
3351 			    ooffset, olen, offset, len);
3352 		}
3353 		if (error != 0 || len == 0)
3354 			break;
3355 		KASSERT(olen > len, ("Iteration did not make progress?"));
3356 		maybe_yield();
3357 	}
3358 
3359 	return (error);
3360 }
3361 
3362 static u_long vn_lock_pair_pause_cnt;
3363 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD,
3364     &vn_lock_pair_pause_cnt, 0,
3365     "Count of vn_lock_pair deadlocks");
3366 
3367 u_int vn_lock_pair_pause_max;
3368 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW,
3369     &vn_lock_pair_pause_max, 0,
3370     "Max ticks for vn_lock_pair deadlock avoidance sleep");
3371 
3372 static void
3373 vn_lock_pair_pause(const char *wmesg)
3374 {
3375 	atomic_add_long(&vn_lock_pair_pause_cnt, 1);
3376 	pause(wmesg, prng32_bounded(vn_lock_pair_pause_max));
3377 }
3378 
3379 /*
3380  * Lock pair of vnodes vp1, vp2, avoiding lock order reversal.
3381  * vp1_locked indicates whether vp1 is exclusively locked; if not, vp1
3382  * must be unlocked.  Same for vp2 and vp2_locked.  One of the vnodes
3383  * can be NULL.
3384  *
3385  * The function returns with both vnodes exclusively locked, and
3386  * guarantees that it does not create lock order reversal with other
3387  * threads during its execution.  Both vnodes could be unlocked
3388  * temporary (and reclaimed).
3389  */
3390 void
3391 vn_lock_pair(struct vnode *vp1, bool vp1_locked, struct vnode *vp2,
3392     bool vp2_locked)
3393 {
3394 	int error;
3395 
3396 	if (vp1 == NULL && vp2 == NULL)
3397 		return;
3398 	if (vp1 != NULL) {
3399 		if (vp1_locked)
3400 			ASSERT_VOP_ELOCKED(vp1, "vp1");
3401 		else
3402 			ASSERT_VOP_UNLOCKED(vp1, "vp1");
3403 	} else {
3404 		vp1_locked = true;
3405 	}
3406 	if (vp2 != NULL) {
3407 		if (vp2_locked)
3408 			ASSERT_VOP_ELOCKED(vp2, "vp2");
3409 		else
3410 			ASSERT_VOP_UNLOCKED(vp2, "vp2");
3411 	} else {
3412 		vp2_locked = true;
3413 	}
3414 	if (!vp1_locked && !vp2_locked) {
3415 		vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3416 		vp1_locked = true;
3417 	}
3418 
3419 	for (;;) {
3420 		if (vp1_locked && vp2_locked)
3421 			break;
3422 		if (vp1_locked && vp2 != NULL) {
3423 			if (vp1 != NULL) {
3424 				error = VOP_LOCK1(vp2, LK_EXCLUSIVE | LK_NOWAIT,
3425 				    __FILE__, __LINE__);
3426 				if (error == 0)
3427 					break;
3428 				VOP_UNLOCK(vp1);
3429 				vp1_locked = false;
3430 				vn_lock_pair_pause("vlp1");
3431 			}
3432 			vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3433 			vp2_locked = true;
3434 		}
3435 		if (vp2_locked && vp1 != NULL) {
3436 			if (vp2 != NULL) {
3437 				error = VOP_LOCK1(vp1, LK_EXCLUSIVE | LK_NOWAIT,
3438 				    __FILE__, __LINE__);
3439 				if (error == 0)
3440 					break;
3441 				VOP_UNLOCK(vp2);
3442 				vp2_locked = false;
3443 				vn_lock_pair_pause("vlp2");
3444 			}
3445 			vn_lock(vp1, LK_EXCLUSIVE | LK_RETRY);
3446 			vp1_locked = true;
3447 		}
3448 	}
3449 	if (vp1 != NULL)
3450 		ASSERT_VOP_ELOCKED(vp1, "vp1 ret");
3451 	if (vp2 != NULL)
3452 		ASSERT_VOP_ELOCKED(vp2, "vp2 ret");
3453 }
3454