xref: /freebsd/sys/kern/vfs_vnops.c (revision 5d06879adb95ac922703072a28fc11048d809a4b)
1 /*-
2  * Copyright (c) 1982, 1986, 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
11  * Copyright (c) 2013, 2014 The FreeBSD Foundation
12  *
13  * Portions of this software were developed by Konstantin Belousov
14  * under sponsorship from the FreeBSD Foundation.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	@(#)vfs_vnops.c	8.2 (Berkeley) 1/21/94
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/disk.h>
49 #include <sys/fail.h>
50 #include <sys/fcntl.h>
51 #include <sys/file.h>
52 #include <sys/kdb.h>
53 #include <sys/stat.h>
54 #include <sys/priv.h>
55 #include <sys/proc.h>
56 #include <sys/limits.h>
57 #include <sys/lock.h>
58 #include <sys/mman.h>
59 #include <sys/mount.h>
60 #include <sys/mutex.h>
61 #include <sys/namei.h>
62 #include <sys/vnode.h>
63 #include <sys/bio.h>
64 #include <sys/buf.h>
65 #include <sys/filio.h>
66 #include <sys/resourcevar.h>
67 #include <sys/rwlock.h>
68 #include <sys/sx.h>
69 #include <sys/sysctl.h>
70 #include <sys/ttycom.h>
71 #include <sys/conf.h>
72 #include <sys/syslog.h>
73 #include <sys/unistd.h>
74 #include <sys/user.h>
75 
76 #include <security/audit/audit.h>
77 #include <security/mac/mac_framework.h>
78 
79 #include <vm/vm.h>
80 #include <vm/vm_extern.h>
81 #include <vm/pmap.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vnode_pager.h>
86 
87 static fo_rdwr_t	vn_read;
88 static fo_rdwr_t	vn_write;
89 static fo_rdwr_t	vn_io_fault;
90 static fo_truncate_t	vn_truncate;
91 static fo_ioctl_t	vn_ioctl;
92 static fo_poll_t	vn_poll;
93 static fo_kqfilter_t	vn_kqfilter;
94 static fo_stat_t	vn_statfile;
95 static fo_close_t	vn_closefile;
96 static fo_mmap_t	vn_mmap;
97 
98 struct 	fileops vnops = {
99 	.fo_read = vn_io_fault,
100 	.fo_write = vn_io_fault,
101 	.fo_truncate = vn_truncate,
102 	.fo_ioctl = vn_ioctl,
103 	.fo_poll = vn_poll,
104 	.fo_kqfilter = vn_kqfilter,
105 	.fo_stat = vn_statfile,
106 	.fo_close = vn_closefile,
107 	.fo_chmod = vn_chmod,
108 	.fo_chown = vn_chown,
109 	.fo_sendfile = vn_sendfile,
110 	.fo_seek = vn_seek,
111 	.fo_fill_kinfo = vn_fill_kinfo,
112 	.fo_mmap = vn_mmap,
113 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
114 };
115 
116 static const int io_hold_cnt = 16;
117 static int vn_io_fault_enable = 1;
118 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
119     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
120 static int vn_io_fault_prefault = 0;
121 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
122     &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
123 static u_long vn_io_faults_cnt;
124 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
125     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
126 
127 /*
128  * Returns true if vn_io_fault mode of handling the i/o request should
129  * be used.
130  */
131 static bool
132 do_vn_io_fault(struct vnode *vp, struct uio *uio)
133 {
134 	struct mount *mp;
135 
136 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
137 	    (mp = vp->v_mount) != NULL &&
138 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
139 }
140 
141 /*
142  * Structure used to pass arguments to vn_io_fault1(), to do either
143  * file- or vnode-based I/O calls.
144  */
145 struct vn_io_fault_args {
146 	enum {
147 		VN_IO_FAULT_FOP,
148 		VN_IO_FAULT_VOP
149 	} kind;
150 	struct ucred *cred;
151 	int flags;
152 	union {
153 		struct fop_args_tag {
154 			struct file *fp;
155 			fo_rdwr_t *doio;
156 		} fop_args;
157 		struct vop_args_tag {
158 			struct vnode *vp;
159 		} vop_args;
160 	} args;
161 };
162 
163 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
164     struct vn_io_fault_args *args, struct thread *td);
165 
166 int
167 vn_open(ndp, flagp, cmode, fp)
168 	struct nameidata *ndp;
169 	int *flagp, cmode;
170 	struct file *fp;
171 {
172 	struct thread *td = ndp->ni_cnd.cn_thread;
173 
174 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
175 }
176 
177 /*
178  * Common code for vnode open operations via a name lookup.
179  * Lookup the vnode and invoke VOP_CREATE if needed.
180  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
181  *
182  * Note that this does NOT free nameidata for the successful case,
183  * due to the NDINIT being done elsewhere.
184  */
185 int
186 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
187     struct ucred *cred, struct file *fp)
188 {
189 	struct vnode *vp;
190 	struct mount *mp;
191 	struct thread *td = ndp->ni_cnd.cn_thread;
192 	struct vattr vat;
193 	struct vattr *vap = &vat;
194 	int fmode, error;
195 
196 restart:
197 	fmode = *flagp;
198 	if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
199 	    O_EXCL | O_DIRECTORY))
200 		return (EINVAL);
201 	else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
202 		ndp->ni_cnd.cn_nameiop = CREATE;
203 		/*
204 		 * Set NOCACHE to avoid flushing the cache when
205 		 * rolling in many files at once.
206 		*/
207 		ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
208 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
209 			ndp->ni_cnd.cn_flags |= FOLLOW;
210 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
211 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
212 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
213 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
214 		bwillwrite();
215 		if ((error = namei(ndp)) != 0)
216 			return (error);
217 		if (ndp->ni_vp == NULL) {
218 			VATTR_NULL(vap);
219 			vap->va_type = VREG;
220 			vap->va_mode = cmode;
221 			if (fmode & O_EXCL)
222 				vap->va_vaflags |= VA_EXCLUSIVE;
223 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
224 				NDFREE(ndp, NDF_ONLY_PNBUF);
225 				vput(ndp->ni_dvp);
226 				if ((error = vn_start_write(NULL, &mp,
227 				    V_XSLEEP | PCATCH)) != 0)
228 					return (error);
229 				goto restart;
230 			}
231 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
232 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
233 #ifdef MAC
234 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
235 			    &ndp->ni_cnd, vap);
236 			if (error == 0)
237 #endif
238 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
239 						   &ndp->ni_cnd, vap);
240 			vput(ndp->ni_dvp);
241 			vn_finished_write(mp);
242 			if (error) {
243 				NDFREE(ndp, NDF_ONLY_PNBUF);
244 				return (error);
245 			}
246 			fmode &= ~O_TRUNC;
247 			vp = ndp->ni_vp;
248 		} else {
249 			if (ndp->ni_dvp == ndp->ni_vp)
250 				vrele(ndp->ni_dvp);
251 			else
252 				vput(ndp->ni_dvp);
253 			ndp->ni_dvp = NULL;
254 			vp = ndp->ni_vp;
255 			if (fmode & O_EXCL) {
256 				error = EEXIST;
257 				goto bad;
258 			}
259 			fmode &= ~O_CREAT;
260 		}
261 	} else {
262 		ndp->ni_cnd.cn_nameiop = LOOKUP;
263 		ndp->ni_cnd.cn_flags = ISOPEN |
264 		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
265 		if (!(fmode & FWRITE))
266 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
267 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
268 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
269 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
270 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
271 		if ((error = namei(ndp)) != 0)
272 			return (error);
273 		vp = ndp->ni_vp;
274 	}
275 	error = vn_open_vnode(vp, fmode, cred, td, fp);
276 	if (error)
277 		goto bad;
278 	*flagp = fmode;
279 	return (0);
280 bad:
281 	NDFREE(ndp, NDF_ONLY_PNBUF);
282 	vput(vp);
283 	*flagp = fmode;
284 	ndp->ni_vp = NULL;
285 	return (error);
286 }
287 
288 /*
289  * Common code for vnode open operations once a vnode is located.
290  * Check permissions, and call the VOP_OPEN routine.
291  */
292 int
293 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
294     struct thread *td, struct file *fp)
295 {
296 	struct mount *mp;
297 	accmode_t accmode;
298 	struct flock lf;
299 	int error, have_flock, lock_flags, type;
300 
301 	if (vp->v_type == VLNK)
302 		return (EMLINK);
303 	if (vp->v_type == VSOCK)
304 		return (EOPNOTSUPP);
305 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
306 		return (ENOTDIR);
307 	accmode = 0;
308 	if (fmode & (FWRITE | O_TRUNC)) {
309 		if (vp->v_type == VDIR)
310 			return (EISDIR);
311 		accmode |= VWRITE;
312 	}
313 	if (fmode & FREAD)
314 		accmode |= VREAD;
315 	if (fmode & FEXEC)
316 		accmode |= VEXEC;
317 	if ((fmode & O_APPEND) && (fmode & FWRITE))
318 		accmode |= VAPPEND;
319 #ifdef MAC
320 	if (fmode & O_CREAT)
321 		accmode |= VCREAT;
322 	if (fmode & O_VERIFY)
323 		accmode |= VVERIFY;
324 	error = mac_vnode_check_open(cred, vp, accmode);
325 	if (error)
326 		return (error);
327 
328 	accmode &= ~(VCREAT | VVERIFY);
329 #endif
330 	if ((fmode & O_CREAT) == 0) {
331 		if (accmode & VWRITE) {
332 			error = vn_writechk(vp);
333 			if (error)
334 				return (error);
335 		}
336 		if (accmode) {
337 		        error = VOP_ACCESS(vp, accmode, cred, td);
338 			if (error)
339 				return (error);
340 		}
341 	}
342 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
343 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
344 	if ((error = VOP_OPEN(vp, fmode, cred, td, fp)) != 0)
345 		return (error);
346 
347 	if (fmode & (O_EXLOCK | O_SHLOCK)) {
348 		KASSERT(fp != NULL, ("open with flock requires fp"));
349 		lock_flags = VOP_ISLOCKED(vp);
350 		VOP_UNLOCK(vp, 0);
351 		lf.l_whence = SEEK_SET;
352 		lf.l_start = 0;
353 		lf.l_len = 0;
354 		if (fmode & O_EXLOCK)
355 			lf.l_type = F_WRLCK;
356 		else
357 			lf.l_type = F_RDLCK;
358 		type = F_FLOCK;
359 		if ((fmode & FNONBLOCK) == 0)
360 			type |= F_WAIT;
361 		error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
362 		have_flock = (error == 0);
363 		vn_lock(vp, lock_flags | LK_RETRY);
364 		if (error == 0 && vp->v_iflag & VI_DOOMED)
365 			error = ENOENT;
366 		/*
367 		 * Another thread might have used this vnode as an
368 		 * executable while the vnode lock was dropped.
369 		 * Ensure the vnode is still able to be opened for
370 		 * writing after the lock has been obtained.
371 		 */
372 		if (error == 0 && accmode & VWRITE)
373 			error = vn_writechk(vp);
374 		if (error) {
375 			VOP_UNLOCK(vp, 0);
376 			if (have_flock) {
377 				lf.l_whence = SEEK_SET;
378 				lf.l_start = 0;
379 				lf.l_len = 0;
380 				lf.l_type = F_UNLCK;
381 				(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf,
382 				    F_FLOCK);
383 			}
384 			vn_start_write(vp, &mp, V_WAIT);
385 			vn_lock(vp, lock_flags | LK_RETRY);
386 			(void)VOP_CLOSE(vp, fmode, cred, td);
387 			vn_finished_write(mp);
388 			/* Prevent second close from fdrop()->vn_close(). */
389 			if (fp != NULL)
390 				fp->f_ops= &badfileops;
391 			return (error);
392 		}
393 		fp->f_flag |= FHASLOCK;
394 	}
395 	if (fmode & FWRITE) {
396 		VOP_ADD_WRITECOUNT(vp, 1);
397 		CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
398 		    __func__, vp, vp->v_writecount);
399 	}
400 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
401 	return (0);
402 }
403 
404 /*
405  * Check for write permissions on the specified vnode.
406  * Prototype text segments cannot be written.
407  */
408 int
409 vn_writechk(vp)
410 	register struct vnode *vp;
411 {
412 
413 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
414 	/*
415 	 * If there's shared text associated with
416 	 * the vnode, try to free it up once.  If
417 	 * we fail, we can't allow writing.
418 	 */
419 	if (VOP_IS_TEXT(vp))
420 		return (ETXTBSY);
421 
422 	return (0);
423 }
424 
425 /*
426  * Vnode close call
427  */
428 int
429 vn_close(vp, flags, file_cred, td)
430 	register struct vnode *vp;
431 	int flags;
432 	struct ucred *file_cred;
433 	struct thread *td;
434 {
435 	struct mount *mp;
436 	int error, lock_flags;
437 
438 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
439 	    MNT_EXTENDED_SHARED(vp->v_mount))
440 		lock_flags = LK_SHARED;
441 	else
442 		lock_flags = LK_EXCLUSIVE;
443 
444 	vn_start_write(vp, &mp, V_WAIT);
445 	vn_lock(vp, lock_flags | LK_RETRY);
446 	if (flags & FWRITE) {
447 		VNASSERT(vp->v_writecount > 0, vp,
448 		    ("vn_close: negative writecount"));
449 		VOP_ADD_WRITECOUNT(vp, -1);
450 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
451 		    __func__, vp, vp->v_writecount);
452 	}
453 	error = VOP_CLOSE(vp, flags, file_cred, td);
454 	vput(vp);
455 	vn_finished_write(mp);
456 	return (error);
457 }
458 
459 /*
460  * Heuristic to detect sequential operation.
461  */
462 static int
463 sequential_heuristic(struct uio *uio, struct file *fp)
464 {
465 
466 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
467 	if (fp->f_flag & FRDAHEAD)
468 		return (fp->f_seqcount << IO_SEQSHIFT);
469 
470 	/*
471 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
472 	 * that the first I/O is normally considered to be slightly
473 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
474 	 * unless previous seeks have reduced f_seqcount to 0, in which
475 	 * case offset 0 is not special.
476 	 */
477 	if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
478 	    uio->uio_offset == fp->f_nextoff) {
479 		/*
480 		 * f_seqcount is in units of fixed-size blocks so that it
481 		 * depends mainly on the amount of sequential I/O and not
482 		 * much on the number of sequential I/O's.  The fixed size
483 		 * of 16384 is hard-coded here since it is (not quite) just
484 		 * a magic size that works well here.  This size is more
485 		 * closely related to the best I/O size for real disks than
486 		 * to any block size used by software.
487 		 */
488 		fp->f_seqcount += howmany(uio->uio_resid, 16384);
489 		if (fp->f_seqcount > IO_SEQMAX)
490 			fp->f_seqcount = IO_SEQMAX;
491 		return (fp->f_seqcount << IO_SEQSHIFT);
492 	}
493 
494 	/* Not sequential.  Quickly draw-down sequentiality. */
495 	if (fp->f_seqcount > 1)
496 		fp->f_seqcount = 1;
497 	else
498 		fp->f_seqcount = 0;
499 	return (0);
500 }
501 
502 /*
503  * Package up an I/O request on a vnode into a uio and do it.
504  */
505 int
506 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
507     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
508     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
509 {
510 	struct uio auio;
511 	struct iovec aiov;
512 	struct mount *mp;
513 	struct ucred *cred;
514 	void *rl_cookie;
515 	struct vn_io_fault_args args;
516 	int error, lock_flags;
517 
518 	auio.uio_iov = &aiov;
519 	auio.uio_iovcnt = 1;
520 	aiov.iov_base = base;
521 	aiov.iov_len = len;
522 	auio.uio_resid = len;
523 	auio.uio_offset = offset;
524 	auio.uio_segflg = segflg;
525 	auio.uio_rw = rw;
526 	auio.uio_td = td;
527 	error = 0;
528 
529 	if ((ioflg & IO_NODELOCKED) == 0) {
530 		if ((ioflg & IO_RANGELOCKED) == 0) {
531 			if (rw == UIO_READ) {
532 				rl_cookie = vn_rangelock_rlock(vp, offset,
533 				    offset + len);
534 			} else {
535 				rl_cookie = vn_rangelock_wlock(vp, offset,
536 				    offset + len);
537 			}
538 		} else
539 			rl_cookie = NULL;
540 		mp = NULL;
541 		if (rw == UIO_WRITE) {
542 			if (vp->v_type != VCHR &&
543 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
544 			    != 0)
545 				goto out;
546 			if (MNT_SHARED_WRITES(mp) ||
547 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
548 				lock_flags = LK_SHARED;
549 			else
550 				lock_flags = LK_EXCLUSIVE;
551 		} else
552 			lock_flags = LK_SHARED;
553 		vn_lock(vp, lock_flags | LK_RETRY);
554 	} else
555 		rl_cookie = NULL;
556 
557 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
558 #ifdef MAC
559 	if ((ioflg & IO_NOMACCHECK) == 0) {
560 		if (rw == UIO_READ)
561 			error = mac_vnode_check_read(active_cred, file_cred,
562 			    vp);
563 		else
564 			error = mac_vnode_check_write(active_cred, file_cred,
565 			    vp);
566 	}
567 #endif
568 	if (error == 0) {
569 		if (file_cred != NULL)
570 			cred = file_cred;
571 		else
572 			cred = active_cred;
573 		if (do_vn_io_fault(vp, &auio)) {
574 			args.kind = VN_IO_FAULT_VOP;
575 			args.cred = cred;
576 			args.flags = ioflg;
577 			args.args.vop_args.vp = vp;
578 			error = vn_io_fault1(vp, &auio, &args, td);
579 		} else if (rw == UIO_READ) {
580 			error = VOP_READ(vp, &auio, ioflg, cred);
581 		} else /* if (rw == UIO_WRITE) */ {
582 			error = VOP_WRITE(vp, &auio, ioflg, cred);
583 		}
584 	}
585 	if (aresid)
586 		*aresid = auio.uio_resid;
587 	else
588 		if (auio.uio_resid && error == 0)
589 			error = EIO;
590 	if ((ioflg & IO_NODELOCKED) == 0) {
591 		VOP_UNLOCK(vp, 0);
592 		if (mp != NULL)
593 			vn_finished_write(mp);
594 	}
595  out:
596 	if (rl_cookie != NULL)
597 		vn_rangelock_unlock(vp, rl_cookie);
598 	return (error);
599 }
600 
601 /*
602  * Package up an I/O request on a vnode into a uio and do it.  The I/O
603  * request is split up into smaller chunks and we try to avoid saturating
604  * the buffer cache while potentially holding a vnode locked, so we
605  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
606  * to give other processes a chance to lock the vnode (either other processes
607  * core'ing the same binary, or unrelated processes scanning the directory).
608  */
609 int
610 vn_rdwr_inchunks(rw, vp, base, len, offset, segflg, ioflg, active_cred,
611     file_cred, aresid, td)
612 	enum uio_rw rw;
613 	struct vnode *vp;
614 	void *base;
615 	size_t len;
616 	off_t offset;
617 	enum uio_seg segflg;
618 	int ioflg;
619 	struct ucred *active_cred;
620 	struct ucred *file_cred;
621 	size_t *aresid;
622 	struct thread *td;
623 {
624 	int error = 0;
625 	ssize_t iaresid;
626 
627 	do {
628 		int chunk;
629 
630 		/*
631 		 * Force `offset' to a multiple of MAXBSIZE except possibly
632 		 * for the first chunk, so that filesystems only need to
633 		 * write full blocks except possibly for the first and last
634 		 * chunks.
635 		 */
636 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
637 
638 		if (chunk > len)
639 			chunk = len;
640 		if (rw != UIO_READ && vp->v_type == VREG)
641 			bwillwrite();
642 		iaresid = 0;
643 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
644 		    ioflg, active_cred, file_cred, &iaresid, td);
645 		len -= chunk;	/* aresid calc already includes length */
646 		if (error)
647 			break;
648 		offset += chunk;
649 		base = (char *)base + chunk;
650 		kern_yield(PRI_USER);
651 	} while (len);
652 	if (aresid)
653 		*aresid = len + iaresid;
654 	return (error);
655 }
656 
657 off_t
658 foffset_lock(struct file *fp, int flags)
659 {
660 	struct mtx *mtxp;
661 	off_t res;
662 
663 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
664 
665 #if OFF_MAX <= LONG_MAX
666 	/*
667 	 * Caller only wants the current f_offset value.  Assume that
668 	 * the long and shorter integer types reads are atomic.
669 	 */
670 	if ((flags & FOF_NOLOCK) != 0)
671 		return (fp->f_offset);
672 #endif
673 
674 	/*
675 	 * According to McKusick the vn lock was protecting f_offset here.
676 	 * It is now protected by the FOFFSET_LOCKED flag.
677 	 */
678 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
679 	mtx_lock(mtxp);
680 	if ((flags & FOF_NOLOCK) == 0) {
681 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
682 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
683 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
684 			    "vofflock", 0);
685 		}
686 		fp->f_vnread_flags |= FOFFSET_LOCKED;
687 	}
688 	res = fp->f_offset;
689 	mtx_unlock(mtxp);
690 	return (res);
691 }
692 
693 void
694 foffset_unlock(struct file *fp, off_t val, int flags)
695 {
696 	struct mtx *mtxp;
697 
698 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
699 
700 #if OFF_MAX <= LONG_MAX
701 	if ((flags & FOF_NOLOCK) != 0) {
702 		if ((flags & FOF_NOUPDATE) == 0)
703 			fp->f_offset = val;
704 		if ((flags & FOF_NEXTOFF) != 0)
705 			fp->f_nextoff = val;
706 		return;
707 	}
708 #endif
709 
710 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
711 	mtx_lock(mtxp);
712 	if ((flags & FOF_NOUPDATE) == 0)
713 		fp->f_offset = val;
714 	if ((flags & FOF_NEXTOFF) != 0)
715 		fp->f_nextoff = val;
716 	if ((flags & FOF_NOLOCK) == 0) {
717 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
718 		    ("Lost FOFFSET_LOCKED"));
719 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
720 			wakeup(&fp->f_vnread_flags);
721 		fp->f_vnread_flags = 0;
722 	}
723 	mtx_unlock(mtxp);
724 }
725 
726 void
727 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
728 {
729 
730 	if ((flags & FOF_OFFSET) == 0)
731 		uio->uio_offset = foffset_lock(fp, flags);
732 }
733 
734 void
735 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
736 {
737 
738 	if ((flags & FOF_OFFSET) == 0)
739 		foffset_unlock(fp, uio->uio_offset, flags);
740 }
741 
742 static int
743 get_advice(struct file *fp, struct uio *uio)
744 {
745 	struct mtx *mtxp;
746 	int ret;
747 
748 	ret = POSIX_FADV_NORMAL;
749 	if (fp->f_advice == NULL)
750 		return (ret);
751 
752 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
753 	mtx_lock(mtxp);
754 	if (uio->uio_offset >= fp->f_advice->fa_start &&
755 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
756 		ret = fp->f_advice->fa_advice;
757 	mtx_unlock(mtxp);
758 	return (ret);
759 }
760 
761 /*
762  * File table vnode read routine.
763  */
764 static int
765 vn_read(fp, uio, active_cred, flags, td)
766 	struct file *fp;
767 	struct uio *uio;
768 	struct ucred *active_cred;
769 	int flags;
770 	struct thread *td;
771 {
772 	struct vnode *vp;
773 	struct mtx *mtxp;
774 	int error, ioflag;
775 	int advice;
776 	off_t offset, start, end;
777 
778 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
779 	    uio->uio_td, td));
780 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
781 	vp = fp->f_vnode;
782 	ioflag = 0;
783 	if (fp->f_flag & FNONBLOCK)
784 		ioflag |= IO_NDELAY;
785 	if (fp->f_flag & O_DIRECT)
786 		ioflag |= IO_DIRECT;
787 	advice = get_advice(fp, uio);
788 	vn_lock(vp, LK_SHARED | LK_RETRY);
789 
790 	switch (advice) {
791 	case POSIX_FADV_NORMAL:
792 	case POSIX_FADV_SEQUENTIAL:
793 	case POSIX_FADV_NOREUSE:
794 		ioflag |= sequential_heuristic(uio, fp);
795 		break;
796 	case POSIX_FADV_RANDOM:
797 		/* Disable read-ahead for random I/O. */
798 		break;
799 	}
800 	offset = uio->uio_offset;
801 
802 #ifdef MAC
803 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
804 	if (error == 0)
805 #endif
806 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
807 	fp->f_nextoff = uio->uio_offset;
808 	VOP_UNLOCK(vp, 0);
809 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
810 	    offset != uio->uio_offset) {
811 		/*
812 		 * Use POSIX_FADV_DONTNEED to flush clean pages and
813 		 * buffers for the backing file after a
814 		 * POSIX_FADV_NOREUSE read(2).  To optimize the common
815 		 * case of using POSIX_FADV_NOREUSE with sequential
816 		 * access, track the previous implicit DONTNEED
817 		 * request and grow this request to include the
818 		 * current read(2) in addition to the previous
819 		 * DONTNEED.  With purely sequential access this will
820 		 * cause the DONTNEED requests to continously grow to
821 		 * cover all of the previously read regions of the
822 		 * file.  This allows filesystem blocks that are
823 		 * accessed by multiple calls to read(2) to be flushed
824 		 * once the last read(2) finishes.
825 		 */
826 		start = offset;
827 		end = uio->uio_offset - 1;
828 		mtxp = mtx_pool_find(mtxpool_sleep, fp);
829 		mtx_lock(mtxp);
830 		if (fp->f_advice != NULL &&
831 		    fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
832 			if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
833 				start = fp->f_advice->fa_prevstart;
834 			else if (fp->f_advice->fa_prevstart != 0 &&
835 			    fp->f_advice->fa_prevstart == end + 1)
836 				end = fp->f_advice->fa_prevend;
837 			fp->f_advice->fa_prevstart = start;
838 			fp->f_advice->fa_prevend = end;
839 		}
840 		mtx_unlock(mtxp);
841 		error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
842 	}
843 	return (error);
844 }
845 
846 /*
847  * File table vnode write routine.
848  */
849 static int
850 vn_write(fp, uio, active_cred, flags, td)
851 	struct file *fp;
852 	struct uio *uio;
853 	struct ucred *active_cred;
854 	int flags;
855 	struct thread *td;
856 {
857 	struct vnode *vp;
858 	struct mount *mp;
859 	struct mtx *mtxp;
860 	int error, ioflag, lock_flags;
861 	int advice;
862 	off_t offset, start, end;
863 
864 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
865 	    uio->uio_td, td));
866 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
867 	vp = fp->f_vnode;
868 	if (vp->v_type == VREG)
869 		bwillwrite();
870 	ioflag = IO_UNIT;
871 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
872 		ioflag |= IO_APPEND;
873 	if (fp->f_flag & FNONBLOCK)
874 		ioflag |= IO_NDELAY;
875 	if (fp->f_flag & O_DIRECT)
876 		ioflag |= IO_DIRECT;
877 	if ((fp->f_flag & O_FSYNC) ||
878 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
879 		ioflag |= IO_SYNC;
880 	mp = NULL;
881 	if (vp->v_type != VCHR &&
882 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
883 		goto unlock;
884 
885 	advice = get_advice(fp, uio);
886 
887 	if (MNT_SHARED_WRITES(mp) ||
888 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
889 		lock_flags = LK_SHARED;
890 	} else {
891 		lock_flags = LK_EXCLUSIVE;
892 	}
893 
894 	vn_lock(vp, lock_flags | LK_RETRY);
895 	switch (advice) {
896 	case POSIX_FADV_NORMAL:
897 	case POSIX_FADV_SEQUENTIAL:
898 	case POSIX_FADV_NOREUSE:
899 		ioflag |= sequential_heuristic(uio, fp);
900 		break;
901 	case POSIX_FADV_RANDOM:
902 		/* XXX: Is this correct? */
903 		break;
904 	}
905 	offset = uio->uio_offset;
906 
907 #ifdef MAC
908 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
909 	if (error == 0)
910 #endif
911 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
912 	fp->f_nextoff = uio->uio_offset;
913 	VOP_UNLOCK(vp, 0);
914 	if (vp->v_type != VCHR)
915 		vn_finished_write(mp);
916 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
917 	    offset != uio->uio_offset) {
918 		/*
919 		 * Use POSIX_FADV_DONTNEED to flush clean pages and
920 		 * buffers for the backing file after a
921 		 * POSIX_FADV_NOREUSE write(2).  To optimize the
922 		 * common case of using POSIX_FADV_NOREUSE with
923 		 * sequential access, track the previous implicit
924 		 * DONTNEED request and grow this request to include
925 		 * the current write(2) in addition to the previous
926 		 * DONTNEED.  With purely sequential access this will
927 		 * cause the DONTNEED requests to continously grow to
928 		 * cover all of the previously written regions of the
929 		 * file.
930 		 *
931 		 * Note that the blocks just written are almost
932 		 * certainly still dirty, so this only works when
933 		 * VOP_ADVISE() calls from subsequent writes push out
934 		 * the data written by this write(2) once the backing
935 		 * buffers are clean.  However, as compared to forcing
936 		 * IO_DIRECT, this gives much saner behavior.  Write
937 		 * clustering is still allowed, and clean pages are
938 		 * merely moved to the cache page queue rather than
939 		 * outright thrown away.  This means a subsequent
940 		 * read(2) can still avoid hitting the disk if the
941 		 * pages have not been reclaimed.
942 		 *
943 		 * This does make POSIX_FADV_NOREUSE largely useless
944 		 * with non-sequential access.  However, sequential
945 		 * access is the more common use case and the flag is
946 		 * merely advisory.
947 		 */
948 		start = offset;
949 		end = uio->uio_offset - 1;
950 		mtxp = mtx_pool_find(mtxpool_sleep, fp);
951 		mtx_lock(mtxp);
952 		if (fp->f_advice != NULL &&
953 		    fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
954 			if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
955 				start = fp->f_advice->fa_prevstart;
956 			else if (fp->f_advice->fa_prevstart != 0 &&
957 			    fp->f_advice->fa_prevstart == end + 1)
958 				end = fp->f_advice->fa_prevend;
959 			fp->f_advice->fa_prevstart = start;
960 			fp->f_advice->fa_prevend = end;
961 		}
962 		mtx_unlock(mtxp);
963 		error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
964 	}
965 
966 unlock:
967 	return (error);
968 }
969 
970 /*
971  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
972  * prevent the following deadlock:
973  *
974  * Assume that the thread A reads from the vnode vp1 into userspace
975  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
976  * currently not resident, then system ends up with the call chain
977  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
978  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
979  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
980  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
981  * backed by the pages of vnode vp1, and some page in buf2 is not
982  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
983  *
984  * To prevent the lock order reversal and deadlock, vn_io_fault() does
985  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
986  * Instead, it first tries to do the whole range i/o with pagefaults
987  * disabled. If all pages in the i/o buffer are resident and mapped,
988  * VOP will succeed (ignoring the genuine filesystem errors).
989  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
990  * i/o in chunks, with all pages in the chunk prefaulted and held
991  * using vm_fault_quick_hold_pages().
992  *
993  * Filesystems using this deadlock avoidance scheme should use the
994  * array of the held pages from uio, saved in the curthread->td_ma,
995  * instead of doing uiomove().  A helper function
996  * vn_io_fault_uiomove() converts uiomove request into
997  * uiomove_fromphys() over td_ma array.
998  *
999  * Since vnode locks do not cover the whole i/o anymore, rangelocks
1000  * make the current i/o request atomic with respect to other i/os and
1001  * truncations.
1002  */
1003 
1004 /*
1005  * Decode vn_io_fault_args and perform the corresponding i/o.
1006  */
1007 static int
1008 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1009     struct thread *td)
1010 {
1011 
1012 	switch (args->kind) {
1013 	case VN_IO_FAULT_FOP:
1014 		return ((args->args.fop_args.doio)(args->args.fop_args.fp,
1015 		    uio, args->cred, args->flags, td));
1016 	case VN_IO_FAULT_VOP:
1017 		if (uio->uio_rw == UIO_READ) {
1018 			return (VOP_READ(args->args.vop_args.vp, uio,
1019 			    args->flags, args->cred));
1020 		} else if (uio->uio_rw == UIO_WRITE) {
1021 			return (VOP_WRITE(args->args.vop_args.vp, uio,
1022 			    args->flags, args->cred));
1023 		}
1024 		break;
1025 	}
1026 	panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind,
1027 	    uio->uio_rw);
1028 }
1029 
1030 static int
1031 vn_io_fault_touch(char *base, const struct uio *uio)
1032 {
1033 	int r;
1034 
1035 	r = fubyte(base);
1036 	if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1037 		return (EFAULT);
1038 	return (0);
1039 }
1040 
1041 static int
1042 vn_io_fault_prefault_user(const struct uio *uio)
1043 {
1044 	char *base;
1045 	const struct iovec *iov;
1046 	size_t len;
1047 	ssize_t resid;
1048 	int error, i;
1049 
1050 	KASSERT(uio->uio_segflg == UIO_USERSPACE,
1051 	    ("vn_io_fault_prefault userspace"));
1052 
1053 	error = i = 0;
1054 	iov = uio->uio_iov;
1055 	resid = uio->uio_resid;
1056 	base = iov->iov_base;
1057 	len = iov->iov_len;
1058 	while (resid > 0) {
1059 		error = vn_io_fault_touch(base, uio);
1060 		if (error != 0)
1061 			break;
1062 		if (len < PAGE_SIZE) {
1063 			if (len != 0) {
1064 				error = vn_io_fault_touch(base + len - 1, uio);
1065 				if (error != 0)
1066 					break;
1067 				resid -= len;
1068 			}
1069 			if (++i >= uio->uio_iovcnt)
1070 				break;
1071 			iov = uio->uio_iov + i;
1072 			base = iov->iov_base;
1073 			len = iov->iov_len;
1074 		} else {
1075 			len -= PAGE_SIZE;
1076 			base += PAGE_SIZE;
1077 			resid -= PAGE_SIZE;
1078 		}
1079 	}
1080 	return (error);
1081 }
1082 
1083 /*
1084  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1085  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1086  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1087  * into args and call vn_io_fault1() to handle faults during the user
1088  * mode buffer accesses.
1089  */
1090 static int
1091 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1092     struct thread *td)
1093 {
1094 	vm_page_t ma[io_hold_cnt + 2];
1095 	struct uio *uio_clone, short_uio;
1096 	struct iovec short_iovec[1];
1097 	vm_page_t *prev_td_ma;
1098 	vm_prot_t prot;
1099 	vm_offset_t addr, end;
1100 	size_t len, resid;
1101 	ssize_t adv;
1102 	int error, cnt, save, saveheld, prev_td_ma_cnt;
1103 
1104 	if (vn_io_fault_prefault) {
1105 		error = vn_io_fault_prefault_user(uio);
1106 		if (error != 0)
1107 			return (error); /* Or ignore ? */
1108 	}
1109 
1110 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1111 
1112 	/*
1113 	 * The UFS follows IO_UNIT directive and replays back both
1114 	 * uio_offset and uio_resid if an error is encountered during the
1115 	 * operation.  But, since the iovec may be already advanced,
1116 	 * uio is still in an inconsistent state.
1117 	 *
1118 	 * Cache a copy of the original uio, which is advanced to the redo
1119 	 * point using UIO_NOCOPY below.
1120 	 */
1121 	uio_clone = cloneuio(uio);
1122 	resid = uio->uio_resid;
1123 
1124 	short_uio.uio_segflg = UIO_USERSPACE;
1125 	short_uio.uio_rw = uio->uio_rw;
1126 	short_uio.uio_td = uio->uio_td;
1127 
1128 	save = vm_fault_disable_pagefaults();
1129 	error = vn_io_fault_doio(args, uio, td);
1130 	if (error != EFAULT)
1131 		goto out;
1132 
1133 	atomic_add_long(&vn_io_faults_cnt, 1);
1134 	uio_clone->uio_segflg = UIO_NOCOPY;
1135 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1136 	uio_clone->uio_segflg = uio->uio_segflg;
1137 
1138 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1139 	prev_td_ma = td->td_ma;
1140 	prev_td_ma_cnt = td->td_ma_cnt;
1141 
1142 	while (uio_clone->uio_resid != 0) {
1143 		len = uio_clone->uio_iov->iov_len;
1144 		if (len == 0) {
1145 			KASSERT(uio_clone->uio_iovcnt >= 1,
1146 			    ("iovcnt underflow"));
1147 			uio_clone->uio_iov++;
1148 			uio_clone->uio_iovcnt--;
1149 			continue;
1150 		}
1151 		if (len > io_hold_cnt * PAGE_SIZE)
1152 			len = io_hold_cnt * PAGE_SIZE;
1153 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1154 		end = round_page(addr + len);
1155 		if (end < addr) {
1156 			error = EFAULT;
1157 			break;
1158 		}
1159 		cnt = atop(end - trunc_page(addr));
1160 		/*
1161 		 * A perfectly misaligned address and length could cause
1162 		 * both the start and the end of the chunk to use partial
1163 		 * page.  +2 accounts for such a situation.
1164 		 */
1165 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1166 		    addr, len, prot, ma, io_hold_cnt + 2);
1167 		if (cnt == -1) {
1168 			error = EFAULT;
1169 			break;
1170 		}
1171 		short_uio.uio_iov = &short_iovec[0];
1172 		short_iovec[0].iov_base = (void *)addr;
1173 		short_uio.uio_iovcnt = 1;
1174 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1175 		short_uio.uio_offset = uio_clone->uio_offset;
1176 		td->td_ma = ma;
1177 		td->td_ma_cnt = cnt;
1178 
1179 		error = vn_io_fault_doio(args, &short_uio, td);
1180 		vm_page_unhold_pages(ma, cnt);
1181 		adv = len - short_uio.uio_resid;
1182 
1183 		uio_clone->uio_iov->iov_base =
1184 		    (char *)uio_clone->uio_iov->iov_base + adv;
1185 		uio_clone->uio_iov->iov_len -= adv;
1186 		uio_clone->uio_resid -= adv;
1187 		uio_clone->uio_offset += adv;
1188 
1189 		uio->uio_resid -= adv;
1190 		uio->uio_offset += adv;
1191 
1192 		if (error != 0 || adv == 0)
1193 			break;
1194 	}
1195 	td->td_ma = prev_td_ma;
1196 	td->td_ma_cnt = prev_td_ma_cnt;
1197 	curthread_pflags_restore(saveheld);
1198 out:
1199 	vm_fault_enable_pagefaults(save);
1200 	free(uio_clone, M_IOV);
1201 	return (error);
1202 }
1203 
1204 static int
1205 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1206     int flags, struct thread *td)
1207 {
1208 	fo_rdwr_t *doio;
1209 	struct vnode *vp;
1210 	void *rl_cookie;
1211 	struct vn_io_fault_args args;
1212 	int error;
1213 
1214 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1215 	vp = fp->f_vnode;
1216 	foffset_lock_uio(fp, uio, flags);
1217 	if (do_vn_io_fault(vp, uio)) {
1218 		args.kind = VN_IO_FAULT_FOP;
1219 		args.args.fop_args.fp = fp;
1220 		args.args.fop_args.doio = doio;
1221 		args.cred = active_cred;
1222 		args.flags = flags | FOF_OFFSET;
1223 		if (uio->uio_rw == UIO_READ) {
1224 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1225 			    uio->uio_offset + uio->uio_resid);
1226 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1227 		    (flags & FOF_OFFSET) == 0) {
1228 			/* For appenders, punt and lock the whole range. */
1229 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1230 		} else {
1231 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1232 			    uio->uio_offset + uio->uio_resid);
1233 		}
1234 		error = vn_io_fault1(vp, uio, &args, td);
1235 		vn_rangelock_unlock(vp, rl_cookie);
1236 	} else {
1237 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1238 	}
1239 	foffset_unlock_uio(fp, uio, flags);
1240 	return (error);
1241 }
1242 
1243 /*
1244  * Helper function to perform the requested uiomove operation using
1245  * the held pages for io->uio_iov[0].iov_base buffer instead of
1246  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1247  * instead of iov_base prevents page faults that could occur due to
1248  * pmap_collect() invalidating the mapping created by
1249  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1250  * object cleanup revoking the write access from page mappings.
1251  *
1252  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1253  * instead of plain uiomove().
1254  */
1255 int
1256 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1257 {
1258 	struct uio transp_uio;
1259 	struct iovec transp_iov[1];
1260 	struct thread *td;
1261 	size_t adv;
1262 	int error, pgadv;
1263 
1264 	td = curthread;
1265 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1266 	    uio->uio_segflg != UIO_USERSPACE)
1267 		return (uiomove(data, xfersize, uio));
1268 
1269 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1270 	transp_iov[0].iov_base = data;
1271 	transp_uio.uio_iov = &transp_iov[0];
1272 	transp_uio.uio_iovcnt = 1;
1273 	if (xfersize > uio->uio_resid)
1274 		xfersize = uio->uio_resid;
1275 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1276 	transp_uio.uio_offset = 0;
1277 	transp_uio.uio_segflg = UIO_SYSSPACE;
1278 	/*
1279 	 * Since transp_iov points to data, and td_ma page array
1280 	 * corresponds to original uio->uio_iov, we need to invert the
1281 	 * direction of the i/o operation as passed to
1282 	 * uiomove_fromphys().
1283 	 */
1284 	switch (uio->uio_rw) {
1285 	case UIO_WRITE:
1286 		transp_uio.uio_rw = UIO_READ;
1287 		break;
1288 	case UIO_READ:
1289 		transp_uio.uio_rw = UIO_WRITE;
1290 		break;
1291 	}
1292 	transp_uio.uio_td = uio->uio_td;
1293 	error = uiomove_fromphys(td->td_ma,
1294 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1295 	    xfersize, &transp_uio);
1296 	adv = xfersize - transp_uio.uio_resid;
1297 	pgadv =
1298 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1299 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1300 	td->td_ma += pgadv;
1301 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1302 	    pgadv));
1303 	td->td_ma_cnt -= pgadv;
1304 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1305 	uio->uio_iov->iov_len -= adv;
1306 	uio->uio_resid -= adv;
1307 	uio->uio_offset += adv;
1308 	return (error);
1309 }
1310 
1311 int
1312 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1313     struct uio *uio)
1314 {
1315 	struct thread *td;
1316 	vm_offset_t iov_base;
1317 	int cnt, pgadv;
1318 
1319 	td = curthread;
1320 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1321 	    uio->uio_segflg != UIO_USERSPACE)
1322 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1323 
1324 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1325 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1326 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1327 	switch (uio->uio_rw) {
1328 	case UIO_WRITE:
1329 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1330 		    offset, cnt);
1331 		break;
1332 	case UIO_READ:
1333 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1334 		    cnt);
1335 		break;
1336 	}
1337 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1338 	td->td_ma += pgadv;
1339 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1340 	    pgadv));
1341 	td->td_ma_cnt -= pgadv;
1342 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1343 	uio->uio_iov->iov_len -= cnt;
1344 	uio->uio_resid -= cnt;
1345 	uio->uio_offset += cnt;
1346 	return (0);
1347 }
1348 
1349 
1350 /*
1351  * File table truncate routine.
1352  */
1353 static int
1354 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1355     struct thread *td)
1356 {
1357 	struct vattr vattr;
1358 	struct mount *mp;
1359 	struct vnode *vp;
1360 	void *rl_cookie;
1361 	int error;
1362 
1363 	vp = fp->f_vnode;
1364 
1365 	/*
1366 	 * Lock the whole range for truncation.  Otherwise split i/o
1367 	 * might happen partly before and partly after the truncation.
1368 	 */
1369 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1370 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1371 	if (error)
1372 		goto out1;
1373 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1374 	if (vp->v_type == VDIR) {
1375 		error = EISDIR;
1376 		goto out;
1377 	}
1378 #ifdef MAC
1379 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1380 	if (error)
1381 		goto out;
1382 #endif
1383 	error = vn_writechk(vp);
1384 	if (error == 0) {
1385 		VATTR_NULL(&vattr);
1386 		vattr.va_size = length;
1387 		error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1388 	}
1389 out:
1390 	VOP_UNLOCK(vp, 0);
1391 	vn_finished_write(mp);
1392 out1:
1393 	vn_rangelock_unlock(vp, rl_cookie);
1394 	return (error);
1395 }
1396 
1397 /*
1398  * File table vnode stat routine.
1399  */
1400 static int
1401 vn_statfile(fp, sb, active_cred, td)
1402 	struct file *fp;
1403 	struct stat *sb;
1404 	struct ucred *active_cred;
1405 	struct thread *td;
1406 {
1407 	struct vnode *vp = fp->f_vnode;
1408 	int error;
1409 
1410 	vn_lock(vp, LK_SHARED | LK_RETRY);
1411 	error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1412 	VOP_UNLOCK(vp, 0);
1413 
1414 	return (error);
1415 }
1416 
1417 /*
1418  * Stat a vnode; implementation for the stat syscall
1419  */
1420 int
1421 vn_stat(vp, sb, active_cred, file_cred, td)
1422 	struct vnode *vp;
1423 	register struct stat *sb;
1424 	struct ucred *active_cred;
1425 	struct ucred *file_cred;
1426 	struct thread *td;
1427 {
1428 	struct vattr vattr;
1429 	register struct vattr *vap;
1430 	int error;
1431 	u_short mode;
1432 
1433 #ifdef MAC
1434 	error = mac_vnode_check_stat(active_cred, file_cred, vp);
1435 	if (error)
1436 		return (error);
1437 #endif
1438 
1439 	vap = &vattr;
1440 
1441 	/*
1442 	 * Initialize defaults for new and unusual fields, so that file
1443 	 * systems which don't support these fields don't need to know
1444 	 * about them.
1445 	 */
1446 	vap->va_birthtime.tv_sec = -1;
1447 	vap->va_birthtime.tv_nsec = 0;
1448 	vap->va_fsid = VNOVAL;
1449 	vap->va_rdev = NODEV;
1450 
1451 	error = VOP_GETATTR(vp, vap, active_cred);
1452 	if (error)
1453 		return (error);
1454 
1455 	/*
1456 	 * Zero the spare stat fields
1457 	 */
1458 	bzero(sb, sizeof *sb);
1459 
1460 	/*
1461 	 * Copy from vattr table
1462 	 */
1463 	if (vap->va_fsid != VNOVAL)
1464 		sb->st_dev = vap->va_fsid;
1465 	else
1466 		sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1467 	sb->st_ino = vap->va_fileid;
1468 	mode = vap->va_mode;
1469 	switch (vap->va_type) {
1470 	case VREG:
1471 		mode |= S_IFREG;
1472 		break;
1473 	case VDIR:
1474 		mode |= S_IFDIR;
1475 		break;
1476 	case VBLK:
1477 		mode |= S_IFBLK;
1478 		break;
1479 	case VCHR:
1480 		mode |= S_IFCHR;
1481 		break;
1482 	case VLNK:
1483 		mode |= S_IFLNK;
1484 		break;
1485 	case VSOCK:
1486 		mode |= S_IFSOCK;
1487 		break;
1488 	case VFIFO:
1489 		mode |= S_IFIFO;
1490 		break;
1491 	default:
1492 		return (EBADF);
1493 	};
1494 	sb->st_mode = mode;
1495 	sb->st_nlink = vap->va_nlink;
1496 	sb->st_uid = vap->va_uid;
1497 	sb->st_gid = vap->va_gid;
1498 	sb->st_rdev = vap->va_rdev;
1499 	if (vap->va_size > OFF_MAX)
1500 		return (EOVERFLOW);
1501 	sb->st_size = vap->va_size;
1502 	sb->st_atim = vap->va_atime;
1503 	sb->st_mtim = vap->va_mtime;
1504 	sb->st_ctim = vap->va_ctime;
1505 	sb->st_birthtim = vap->va_birthtime;
1506 
1507         /*
1508 	 * According to www.opengroup.org, the meaning of st_blksize is
1509 	 *   "a filesystem-specific preferred I/O block size for this
1510 	 *    object.  In some filesystem types, this may vary from file
1511 	 *    to file"
1512 	 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1513 	 */
1514 
1515 	sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1516 
1517 	sb->st_flags = vap->va_flags;
1518 	if (priv_check(td, PRIV_VFS_GENERATION))
1519 		sb->st_gen = 0;
1520 	else
1521 		sb->st_gen = vap->va_gen;
1522 
1523 	sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1524 	return (0);
1525 }
1526 
1527 /*
1528  * File table vnode ioctl routine.
1529  */
1530 static int
1531 vn_ioctl(fp, com, data, active_cred, td)
1532 	struct file *fp;
1533 	u_long com;
1534 	void *data;
1535 	struct ucred *active_cred;
1536 	struct thread *td;
1537 {
1538 	struct vattr vattr;
1539 	struct vnode *vp;
1540 	int error;
1541 
1542 	vp = fp->f_vnode;
1543 	switch (vp->v_type) {
1544 	case VDIR:
1545 	case VREG:
1546 		switch (com) {
1547 		case FIONREAD:
1548 			vn_lock(vp, LK_SHARED | LK_RETRY);
1549 			error = VOP_GETATTR(vp, &vattr, active_cred);
1550 			VOP_UNLOCK(vp, 0);
1551 			if (error == 0)
1552 				*(int *)data = vattr.va_size - fp->f_offset;
1553 			return (error);
1554 		case FIONBIO:
1555 		case FIOASYNC:
1556 			return (0);
1557 		default:
1558 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1559 			    active_cred, td));
1560 		}
1561 	default:
1562 		return (ENOTTY);
1563 	}
1564 }
1565 
1566 /*
1567  * File table vnode poll routine.
1568  */
1569 static int
1570 vn_poll(fp, events, active_cred, td)
1571 	struct file *fp;
1572 	int events;
1573 	struct ucred *active_cred;
1574 	struct thread *td;
1575 {
1576 	struct vnode *vp;
1577 	int error;
1578 
1579 	vp = fp->f_vnode;
1580 #ifdef MAC
1581 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1582 	error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1583 	VOP_UNLOCK(vp, 0);
1584 	if (!error)
1585 #endif
1586 
1587 	error = VOP_POLL(vp, events, fp->f_cred, td);
1588 	return (error);
1589 }
1590 
1591 /*
1592  * Acquire the requested lock and then check for validity.  LK_RETRY
1593  * permits vn_lock to return doomed vnodes.
1594  */
1595 int
1596 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1597 {
1598 	int error;
1599 
1600 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1601 	    ("vn_lock called with no locktype."));
1602 	do {
1603 #ifdef DEBUG_VFS_LOCKS
1604 		KASSERT(vp->v_holdcnt != 0,
1605 		    ("vn_lock %p: zero hold count", vp));
1606 #endif
1607 		error = VOP_LOCK1(vp, flags, file, line);
1608 		flags &= ~LK_INTERLOCK;	/* Interlock is always dropped. */
1609 		KASSERT((flags & LK_RETRY) == 0 || error == 0,
1610 		    ("LK_RETRY set with incompatible flags (0x%x) or an error occured (%d)",
1611 		    flags, error));
1612 		/*
1613 		 * Callers specify LK_RETRY if they wish to get dead vnodes.
1614 		 * If RETRY is not set, we return ENOENT instead.
1615 		 */
1616 		if (error == 0 && vp->v_iflag & VI_DOOMED &&
1617 		    (flags & LK_RETRY) == 0) {
1618 			VOP_UNLOCK(vp, 0);
1619 			error = ENOENT;
1620 			break;
1621 		}
1622 	} while (flags & LK_RETRY && error != 0);
1623 	return (error);
1624 }
1625 
1626 /*
1627  * File table vnode close routine.
1628  */
1629 static int
1630 vn_closefile(fp, td)
1631 	struct file *fp;
1632 	struct thread *td;
1633 {
1634 	struct vnode *vp;
1635 	struct flock lf;
1636 	int error;
1637 
1638 	vp = fp->f_vnode;
1639 	fp->f_ops = &badfileops;
1640 
1641 	if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK)
1642 		vref(vp);
1643 
1644 	error = vn_close(vp, fp->f_flag, fp->f_cred, td);
1645 
1646 	if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK) {
1647 		lf.l_whence = SEEK_SET;
1648 		lf.l_start = 0;
1649 		lf.l_len = 0;
1650 		lf.l_type = F_UNLCK;
1651 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1652 		vrele(vp);
1653 	}
1654 	return (error);
1655 }
1656 
1657 static bool
1658 vn_suspendable(struct mount *mp)
1659 {
1660 
1661 	return (mp->mnt_op->vfs_susp_clean != NULL);
1662 }
1663 
1664 /*
1665  * Preparing to start a filesystem write operation. If the operation is
1666  * permitted, then we bump the count of operations in progress and
1667  * proceed. If a suspend request is in progress, we wait until the
1668  * suspension is over, and then proceed.
1669  */
1670 static int
1671 vn_start_write_locked(struct mount *mp, int flags)
1672 {
1673 	int error, mflags;
1674 
1675 	mtx_assert(MNT_MTX(mp), MA_OWNED);
1676 	error = 0;
1677 
1678 	/*
1679 	 * Check on status of suspension.
1680 	 */
1681 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1682 	    mp->mnt_susp_owner != curthread) {
1683 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1684 		    (flags & PCATCH) : 0) | (PUSER - 1);
1685 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1686 			if (flags & V_NOWAIT) {
1687 				error = EWOULDBLOCK;
1688 				goto unlock;
1689 			}
1690 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1691 			    "suspfs", 0);
1692 			if (error)
1693 				goto unlock;
1694 		}
1695 	}
1696 	if (flags & V_XSLEEP)
1697 		goto unlock;
1698 	mp->mnt_writeopcount++;
1699 unlock:
1700 	if (error != 0 || (flags & V_XSLEEP) != 0)
1701 		MNT_REL(mp);
1702 	MNT_IUNLOCK(mp);
1703 	return (error);
1704 }
1705 
1706 int
1707 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1708 {
1709 	struct mount *mp;
1710 	int error;
1711 
1712 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1713 	    ("V_MNTREF requires mp"));
1714 
1715 	error = 0;
1716 	/*
1717 	 * If a vnode is provided, get and return the mount point that
1718 	 * to which it will write.
1719 	 */
1720 	if (vp != NULL) {
1721 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1722 			*mpp = NULL;
1723 			if (error != EOPNOTSUPP)
1724 				return (error);
1725 			return (0);
1726 		}
1727 	}
1728 	if ((mp = *mpp) == NULL)
1729 		return (0);
1730 
1731 	if (!vn_suspendable(mp)) {
1732 		if (vp != NULL || (flags & V_MNTREF) != 0)
1733 			vfs_rel(mp);
1734 		return (0);
1735 	}
1736 
1737 	/*
1738 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1739 	 * a vfs_ref().
1740 	 * As long as a vnode is not provided we need to acquire a
1741 	 * refcount for the provided mountpoint too, in order to
1742 	 * emulate a vfs_ref().
1743 	 */
1744 	MNT_ILOCK(mp);
1745 	if (vp == NULL && (flags & V_MNTREF) == 0)
1746 		MNT_REF(mp);
1747 
1748 	return (vn_start_write_locked(mp, flags));
1749 }
1750 
1751 /*
1752  * Secondary suspension. Used by operations such as vop_inactive
1753  * routines that are needed by the higher level functions. These
1754  * are allowed to proceed until all the higher level functions have
1755  * completed (indicated by mnt_writeopcount dropping to zero). At that
1756  * time, these operations are halted until the suspension is over.
1757  */
1758 int
1759 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1760 {
1761 	struct mount *mp;
1762 	int error;
1763 
1764 	KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1765 	    ("V_MNTREF requires mp"));
1766 
1767  retry:
1768 	if (vp != NULL) {
1769 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1770 			*mpp = NULL;
1771 			if (error != EOPNOTSUPP)
1772 				return (error);
1773 			return (0);
1774 		}
1775 	}
1776 	/*
1777 	 * If we are not suspended or have not yet reached suspended
1778 	 * mode, then let the operation proceed.
1779 	 */
1780 	if ((mp = *mpp) == NULL)
1781 		return (0);
1782 
1783 	if (!vn_suspendable(mp)) {
1784 		if (vp != NULL || (flags & V_MNTREF) != 0)
1785 			vfs_rel(mp);
1786 		return (0);
1787 	}
1788 
1789 	/*
1790 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1791 	 * a vfs_ref().
1792 	 * As long as a vnode is not provided we need to acquire a
1793 	 * refcount for the provided mountpoint too, in order to
1794 	 * emulate a vfs_ref().
1795 	 */
1796 	MNT_ILOCK(mp);
1797 	if (vp == NULL && (flags & V_MNTREF) == 0)
1798 		MNT_REF(mp);
1799 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1800 		mp->mnt_secondary_writes++;
1801 		mp->mnt_secondary_accwrites++;
1802 		MNT_IUNLOCK(mp);
1803 		return (0);
1804 	}
1805 	if (flags & V_NOWAIT) {
1806 		MNT_REL(mp);
1807 		MNT_IUNLOCK(mp);
1808 		return (EWOULDBLOCK);
1809 	}
1810 	/*
1811 	 * Wait for the suspension to finish.
1812 	 */
1813 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1814 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1815 	    "suspfs", 0);
1816 	vfs_rel(mp);
1817 	if (error == 0)
1818 		goto retry;
1819 	return (error);
1820 }
1821 
1822 /*
1823  * Filesystem write operation has completed. If we are suspending and this
1824  * operation is the last one, notify the suspender that the suspension is
1825  * now in effect.
1826  */
1827 void
1828 vn_finished_write(mp)
1829 	struct mount *mp;
1830 {
1831 	if (mp == NULL || !vn_suspendable(mp))
1832 		return;
1833 	MNT_ILOCK(mp);
1834 	MNT_REL(mp);
1835 	mp->mnt_writeopcount--;
1836 	if (mp->mnt_writeopcount < 0)
1837 		panic("vn_finished_write: neg cnt");
1838 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1839 	    mp->mnt_writeopcount <= 0)
1840 		wakeup(&mp->mnt_writeopcount);
1841 	MNT_IUNLOCK(mp);
1842 }
1843 
1844 
1845 /*
1846  * Filesystem secondary write operation has completed. If we are
1847  * suspending and this operation is the last one, notify the suspender
1848  * that the suspension is now in effect.
1849  */
1850 void
1851 vn_finished_secondary_write(mp)
1852 	struct mount *mp;
1853 {
1854 	if (mp == NULL || !vn_suspendable(mp))
1855 		return;
1856 	MNT_ILOCK(mp);
1857 	MNT_REL(mp);
1858 	mp->mnt_secondary_writes--;
1859 	if (mp->mnt_secondary_writes < 0)
1860 		panic("vn_finished_secondary_write: neg cnt");
1861 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1862 	    mp->mnt_secondary_writes <= 0)
1863 		wakeup(&mp->mnt_secondary_writes);
1864 	MNT_IUNLOCK(mp);
1865 }
1866 
1867 
1868 
1869 /*
1870  * Request a filesystem to suspend write operations.
1871  */
1872 int
1873 vfs_write_suspend(struct mount *mp, int flags)
1874 {
1875 	int error;
1876 
1877 	MPASS(vn_suspendable(mp));
1878 
1879 	MNT_ILOCK(mp);
1880 	if (mp->mnt_susp_owner == curthread) {
1881 		MNT_IUNLOCK(mp);
1882 		return (EALREADY);
1883 	}
1884 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
1885 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1886 
1887 	/*
1888 	 * Unmount holds a write reference on the mount point.  If we
1889 	 * own busy reference and drain for writers, we deadlock with
1890 	 * the reference draining in the unmount path.  Callers of
1891 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1892 	 * vfs_busy() reference is owned and caller is not in the
1893 	 * unmount context.
1894 	 */
1895 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1896 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1897 		MNT_IUNLOCK(mp);
1898 		return (EBUSY);
1899 	}
1900 
1901 	mp->mnt_kern_flag |= MNTK_SUSPEND;
1902 	mp->mnt_susp_owner = curthread;
1903 	if (mp->mnt_writeopcount > 0)
1904 		(void) msleep(&mp->mnt_writeopcount,
1905 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1906 	else
1907 		MNT_IUNLOCK(mp);
1908 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1909 		vfs_write_resume(mp, 0);
1910 	return (error);
1911 }
1912 
1913 /*
1914  * Request a filesystem to resume write operations.
1915  */
1916 void
1917 vfs_write_resume(struct mount *mp, int flags)
1918 {
1919 
1920 	MPASS(vn_suspendable(mp));
1921 
1922 	MNT_ILOCK(mp);
1923 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1924 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1925 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1926 				       MNTK_SUSPENDED);
1927 		mp->mnt_susp_owner = NULL;
1928 		wakeup(&mp->mnt_writeopcount);
1929 		wakeup(&mp->mnt_flag);
1930 		curthread->td_pflags &= ~TDP_IGNSUSP;
1931 		if ((flags & VR_START_WRITE) != 0) {
1932 			MNT_REF(mp);
1933 			mp->mnt_writeopcount++;
1934 		}
1935 		MNT_IUNLOCK(mp);
1936 		if ((flags & VR_NO_SUSPCLR) == 0)
1937 			VFS_SUSP_CLEAN(mp);
1938 	} else if ((flags & VR_START_WRITE) != 0) {
1939 		MNT_REF(mp);
1940 		vn_start_write_locked(mp, 0);
1941 	} else {
1942 		MNT_IUNLOCK(mp);
1943 	}
1944 }
1945 
1946 /*
1947  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1948  * methods.
1949  */
1950 int
1951 vfs_write_suspend_umnt(struct mount *mp)
1952 {
1953 	int error;
1954 
1955 	MPASS(vn_suspendable(mp));
1956 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1957 	    ("vfs_write_suspend_umnt: recursed"));
1958 
1959 	/* dounmount() already called vn_start_write(). */
1960 	for (;;) {
1961 		vn_finished_write(mp);
1962 		error = vfs_write_suspend(mp, 0);
1963 		if (error != 0) {
1964 			vn_start_write(NULL, &mp, V_WAIT);
1965 			return (error);
1966 		}
1967 		MNT_ILOCK(mp);
1968 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1969 			break;
1970 		MNT_IUNLOCK(mp);
1971 		vn_start_write(NULL, &mp, V_WAIT);
1972 	}
1973 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1974 	wakeup(&mp->mnt_flag);
1975 	MNT_IUNLOCK(mp);
1976 	curthread->td_pflags |= TDP_IGNSUSP;
1977 	return (0);
1978 }
1979 
1980 /*
1981  * Implement kqueues for files by translating it to vnode operation.
1982  */
1983 static int
1984 vn_kqfilter(struct file *fp, struct knote *kn)
1985 {
1986 
1987 	return (VOP_KQFILTER(fp->f_vnode, kn));
1988 }
1989 
1990 /*
1991  * Simplified in-kernel wrapper calls for extended attribute access.
1992  * Both calls pass in a NULL credential, authorizing as "kernel" access.
1993  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1994  */
1995 int
1996 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1997     const char *attrname, int *buflen, char *buf, struct thread *td)
1998 {
1999 	struct uio	auio;
2000 	struct iovec	iov;
2001 	int	error;
2002 
2003 	iov.iov_len = *buflen;
2004 	iov.iov_base = buf;
2005 
2006 	auio.uio_iov = &iov;
2007 	auio.uio_iovcnt = 1;
2008 	auio.uio_rw = UIO_READ;
2009 	auio.uio_segflg = UIO_SYSSPACE;
2010 	auio.uio_td = td;
2011 	auio.uio_offset = 0;
2012 	auio.uio_resid = *buflen;
2013 
2014 	if ((ioflg & IO_NODELOCKED) == 0)
2015 		vn_lock(vp, LK_SHARED | LK_RETRY);
2016 
2017 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2018 
2019 	/* authorize attribute retrieval as kernel */
2020 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
2021 	    td);
2022 
2023 	if ((ioflg & IO_NODELOCKED) == 0)
2024 		VOP_UNLOCK(vp, 0);
2025 
2026 	if (error == 0) {
2027 		*buflen = *buflen - auio.uio_resid;
2028 	}
2029 
2030 	return (error);
2031 }
2032 
2033 /*
2034  * XXX failure mode if partially written?
2035  */
2036 int
2037 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2038     const char *attrname, int buflen, char *buf, struct thread *td)
2039 {
2040 	struct uio	auio;
2041 	struct iovec	iov;
2042 	struct mount	*mp;
2043 	int	error;
2044 
2045 	iov.iov_len = buflen;
2046 	iov.iov_base = buf;
2047 
2048 	auio.uio_iov = &iov;
2049 	auio.uio_iovcnt = 1;
2050 	auio.uio_rw = UIO_WRITE;
2051 	auio.uio_segflg = UIO_SYSSPACE;
2052 	auio.uio_td = td;
2053 	auio.uio_offset = 0;
2054 	auio.uio_resid = buflen;
2055 
2056 	if ((ioflg & IO_NODELOCKED) == 0) {
2057 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2058 			return (error);
2059 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2060 	}
2061 
2062 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2063 
2064 	/* authorize attribute setting as kernel */
2065 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2066 
2067 	if ((ioflg & IO_NODELOCKED) == 0) {
2068 		vn_finished_write(mp);
2069 		VOP_UNLOCK(vp, 0);
2070 	}
2071 
2072 	return (error);
2073 }
2074 
2075 int
2076 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2077     const char *attrname, struct thread *td)
2078 {
2079 	struct mount	*mp;
2080 	int	error;
2081 
2082 	if ((ioflg & IO_NODELOCKED) == 0) {
2083 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2084 			return (error);
2085 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2086 	}
2087 
2088 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2089 
2090 	/* authorize attribute removal as kernel */
2091 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2092 	if (error == EOPNOTSUPP)
2093 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2094 		    NULL, td);
2095 
2096 	if ((ioflg & IO_NODELOCKED) == 0) {
2097 		vn_finished_write(mp);
2098 		VOP_UNLOCK(vp, 0);
2099 	}
2100 
2101 	return (error);
2102 }
2103 
2104 static int
2105 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2106     struct vnode **rvp)
2107 {
2108 
2109 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2110 }
2111 
2112 int
2113 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2114 {
2115 
2116 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2117 	    lkflags, rvp));
2118 }
2119 
2120 int
2121 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2122     int lkflags, struct vnode **rvp)
2123 {
2124 	struct mount *mp;
2125 	int ltype, error;
2126 
2127 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2128 	mp = vp->v_mount;
2129 	ltype = VOP_ISLOCKED(vp);
2130 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2131 	    ("vn_vget_ino: vp not locked"));
2132 	error = vfs_busy(mp, MBF_NOWAIT);
2133 	if (error != 0) {
2134 		vfs_ref(mp);
2135 		VOP_UNLOCK(vp, 0);
2136 		error = vfs_busy(mp, 0);
2137 		vn_lock(vp, ltype | LK_RETRY);
2138 		vfs_rel(mp);
2139 		if (error != 0)
2140 			return (ENOENT);
2141 		if (vp->v_iflag & VI_DOOMED) {
2142 			vfs_unbusy(mp);
2143 			return (ENOENT);
2144 		}
2145 	}
2146 	VOP_UNLOCK(vp, 0);
2147 	error = alloc(mp, alloc_arg, lkflags, rvp);
2148 	vfs_unbusy(mp);
2149 	if (*rvp != vp)
2150 		vn_lock(vp, ltype | LK_RETRY);
2151 	if (vp->v_iflag & VI_DOOMED) {
2152 		if (error == 0) {
2153 			if (*rvp == vp)
2154 				vunref(vp);
2155 			else
2156 				vput(*rvp);
2157 		}
2158 		error = ENOENT;
2159 	}
2160 	return (error);
2161 }
2162 
2163 int
2164 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2165     struct thread *td)
2166 {
2167 
2168 	if (vp->v_type != VREG || td == NULL)
2169 		return (0);
2170 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2171 	    lim_cur(td, RLIMIT_FSIZE)) {
2172 		PROC_LOCK(td->td_proc);
2173 		kern_psignal(td->td_proc, SIGXFSZ);
2174 		PROC_UNLOCK(td->td_proc);
2175 		return (EFBIG);
2176 	}
2177 	return (0);
2178 }
2179 
2180 int
2181 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2182     struct thread *td)
2183 {
2184 	struct vnode *vp;
2185 
2186 	vp = fp->f_vnode;
2187 #ifdef AUDIT
2188 	vn_lock(vp, LK_SHARED | LK_RETRY);
2189 	AUDIT_ARG_VNODE1(vp);
2190 	VOP_UNLOCK(vp, 0);
2191 #endif
2192 	return (setfmode(td, active_cred, vp, mode));
2193 }
2194 
2195 int
2196 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2197     struct thread *td)
2198 {
2199 	struct vnode *vp;
2200 
2201 	vp = fp->f_vnode;
2202 #ifdef AUDIT
2203 	vn_lock(vp, LK_SHARED | LK_RETRY);
2204 	AUDIT_ARG_VNODE1(vp);
2205 	VOP_UNLOCK(vp, 0);
2206 #endif
2207 	return (setfown(td, active_cred, vp, uid, gid));
2208 }
2209 
2210 void
2211 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2212 {
2213 	vm_object_t object;
2214 
2215 	if ((object = vp->v_object) == NULL)
2216 		return;
2217 	VM_OBJECT_WLOCK(object);
2218 	vm_object_page_remove(object, start, end, 0);
2219 	VM_OBJECT_WUNLOCK(object);
2220 }
2221 
2222 int
2223 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2224 {
2225 	struct vattr va;
2226 	daddr_t bn, bnp;
2227 	uint64_t bsize;
2228 	off_t noff;
2229 	int error;
2230 
2231 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2232 	    ("Wrong command %lu", cmd));
2233 
2234 	if (vn_lock(vp, LK_SHARED) != 0)
2235 		return (EBADF);
2236 	if (vp->v_type != VREG) {
2237 		error = ENOTTY;
2238 		goto unlock;
2239 	}
2240 	error = VOP_GETATTR(vp, &va, cred);
2241 	if (error != 0)
2242 		goto unlock;
2243 	noff = *off;
2244 	if (noff >= va.va_size) {
2245 		error = ENXIO;
2246 		goto unlock;
2247 	}
2248 	bsize = vp->v_mount->mnt_stat.f_iosize;
2249 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize) {
2250 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2251 		if (error == EOPNOTSUPP) {
2252 			error = ENOTTY;
2253 			goto unlock;
2254 		}
2255 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2256 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2257 			noff = bn * bsize;
2258 			if (noff < *off)
2259 				noff = *off;
2260 			goto unlock;
2261 		}
2262 	}
2263 	if (noff > va.va_size)
2264 		noff = va.va_size;
2265 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2266 	if (cmd == FIOSEEKDATA)
2267 		error = ENXIO;
2268 unlock:
2269 	VOP_UNLOCK(vp, 0);
2270 	if (error == 0)
2271 		*off = noff;
2272 	return (error);
2273 }
2274 
2275 int
2276 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2277 {
2278 	struct ucred *cred;
2279 	struct vnode *vp;
2280 	struct vattr vattr;
2281 	off_t foffset, size;
2282 	int error, noneg;
2283 
2284 	cred = td->td_ucred;
2285 	vp = fp->f_vnode;
2286 	foffset = foffset_lock(fp, 0);
2287 	noneg = (vp->v_type != VCHR);
2288 	error = 0;
2289 	switch (whence) {
2290 	case L_INCR:
2291 		if (noneg &&
2292 		    (foffset < 0 ||
2293 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2294 			error = EOVERFLOW;
2295 			break;
2296 		}
2297 		offset += foffset;
2298 		break;
2299 	case L_XTND:
2300 		vn_lock(vp, LK_SHARED | LK_RETRY);
2301 		error = VOP_GETATTR(vp, &vattr, cred);
2302 		VOP_UNLOCK(vp, 0);
2303 		if (error)
2304 			break;
2305 
2306 		/*
2307 		 * If the file references a disk device, then fetch
2308 		 * the media size and use that to determine the ending
2309 		 * offset.
2310 		 */
2311 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2312 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2313 			vattr.va_size = size;
2314 		if (noneg &&
2315 		    (vattr.va_size > OFF_MAX ||
2316 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2317 			error = EOVERFLOW;
2318 			break;
2319 		}
2320 		offset += vattr.va_size;
2321 		break;
2322 	case L_SET:
2323 		break;
2324 	case SEEK_DATA:
2325 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2326 		break;
2327 	case SEEK_HOLE:
2328 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2329 		break;
2330 	default:
2331 		error = EINVAL;
2332 	}
2333 	if (error == 0 && noneg && offset < 0)
2334 		error = EINVAL;
2335 	if (error != 0)
2336 		goto drop;
2337 	VFS_KNOTE_UNLOCKED(vp, 0);
2338 	td->td_uretoff.tdu_off = offset;
2339 drop:
2340 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2341 	return (error);
2342 }
2343 
2344 int
2345 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2346     struct thread *td)
2347 {
2348 	int error;
2349 
2350 	/*
2351 	 * Grant permission if the caller is the owner of the file, or
2352 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2353 	 * on the file.  If the time pointer is null, then write
2354 	 * permission on the file is also sufficient.
2355 	 *
2356 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2357 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2358 	 * will be allowed to set the times [..] to the current
2359 	 * server time.
2360 	 */
2361 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2362 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2363 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2364 	return (error);
2365 }
2366 
2367 int
2368 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2369 {
2370 	struct vnode *vp;
2371 	int error;
2372 
2373 	if (fp->f_type == DTYPE_FIFO)
2374 		kif->kf_type = KF_TYPE_FIFO;
2375 	else
2376 		kif->kf_type = KF_TYPE_VNODE;
2377 	vp = fp->f_vnode;
2378 	vref(vp);
2379 	FILEDESC_SUNLOCK(fdp);
2380 	error = vn_fill_kinfo_vnode(vp, kif);
2381 	vrele(vp);
2382 	FILEDESC_SLOCK(fdp);
2383 	return (error);
2384 }
2385 
2386 static inline void
2387 vn_fill_junk(struct kinfo_file *kif)
2388 {
2389 	size_t len, olen;
2390 
2391 	/*
2392 	 * Simulate vn_fullpath returning changing values for a given
2393 	 * vp during e.g. coredump.
2394 	 */
2395 	len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1;
2396 	olen = strlen(kif->kf_path);
2397 	if (len < olen)
2398 		strcpy(&kif->kf_path[len - 1], "$");
2399 	else
2400 		for (; olen < len; olen++)
2401 			strcpy(&kif->kf_path[olen], "A");
2402 }
2403 
2404 int
2405 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2406 {
2407 	struct vattr va;
2408 	char *fullpath, *freepath;
2409 	int error;
2410 
2411 	kif->kf_vnode_type = vntype_to_kinfo(vp->v_type);
2412 	freepath = NULL;
2413 	fullpath = "-";
2414 	error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2415 	if (error == 0) {
2416 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2417 	}
2418 	if (freepath != NULL)
2419 		free(freepath, M_TEMP);
2420 
2421 	KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path,
2422 		vn_fill_junk(kif);
2423 	);
2424 
2425 	/*
2426 	 * Retrieve vnode attributes.
2427 	 */
2428 	va.va_fsid = VNOVAL;
2429 	va.va_rdev = NODEV;
2430 	vn_lock(vp, LK_SHARED | LK_RETRY);
2431 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2432 	VOP_UNLOCK(vp, 0);
2433 	if (error != 0)
2434 		return (error);
2435 	if (va.va_fsid != VNOVAL)
2436 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2437 	else
2438 		kif->kf_un.kf_file.kf_file_fsid =
2439 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2440 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2441 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2442 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2443 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2444 	return (0);
2445 }
2446 
2447 int
2448 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size,
2449     vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff,
2450     struct thread *td)
2451 {
2452 #ifdef HWPMC_HOOKS
2453 	struct pmckern_map_in pkm;
2454 #endif
2455 	struct mount *mp;
2456 	struct vnode *vp;
2457 	vm_object_t object;
2458 	vm_prot_t maxprot;
2459 	boolean_t writecounted;
2460 	int error;
2461 
2462 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
2463     defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
2464 	/*
2465 	 * POSIX shared-memory objects are defined to have
2466 	 * kernel persistence, and are not defined to support
2467 	 * read(2)/write(2) -- or even open(2).  Thus, we can
2468 	 * use MAP_ASYNC to trade on-disk coherence for speed.
2469 	 * The shm_open(3) library routine turns on the FPOSIXSHM
2470 	 * flag to request this behavior.
2471 	 */
2472 	if ((fp->f_flag & FPOSIXSHM) != 0)
2473 		flags |= MAP_NOSYNC;
2474 #endif
2475 	vp = fp->f_vnode;
2476 
2477 	/*
2478 	 * Ensure that file and memory protections are
2479 	 * compatible.  Note that we only worry about
2480 	 * writability if mapping is shared; in this case,
2481 	 * current and max prot are dictated by the open file.
2482 	 * XXX use the vnode instead?  Problem is: what
2483 	 * credentials do we use for determination? What if
2484 	 * proc does a setuid?
2485 	 */
2486 	mp = vp->v_mount;
2487 	if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0)
2488 		maxprot = VM_PROT_NONE;
2489 	else
2490 		maxprot = VM_PROT_EXECUTE;
2491 	if ((fp->f_flag & FREAD) != 0)
2492 		maxprot |= VM_PROT_READ;
2493 	else if ((prot & VM_PROT_READ) != 0)
2494 		return (EACCES);
2495 
2496 	/*
2497 	 * If we are sharing potential changes via MAP_SHARED and we
2498 	 * are trying to get write permission although we opened it
2499 	 * without asking for it, bail out.
2500 	 */
2501 	if ((flags & MAP_SHARED) != 0) {
2502 		if ((fp->f_flag & FWRITE) != 0)
2503 			maxprot |= VM_PROT_WRITE;
2504 		else if ((prot & VM_PROT_WRITE) != 0)
2505 			return (EACCES);
2506 	} else {
2507 		maxprot |= VM_PROT_WRITE;
2508 		cap_maxprot |= VM_PROT_WRITE;
2509 	}
2510 	maxprot &= cap_maxprot;
2511 
2512 	writecounted = FALSE;
2513 	error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp,
2514 	    &foff, &object, &writecounted);
2515 	if (error != 0)
2516 		return (error);
2517 	error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object,
2518 	    foff, writecounted, td);
2519 	if (error != 0) {
2520 		/*
2521 		 * If this mapping was accounted for in the vnode's
2522 		 * writecount, then undo that now.
2523 		 */
2524 		if (writecounted)
2525 			vnode_pager_release_writecount(object, 0, size);
2526 		vm_object_deallocate(object);
2527 	}
2528 #ifdef HWPMC_HOOKS
2529 	/* Inform hwpmc(4) if an executable is being mapped. */
2530 	if (error == 0 && (prot & VM_PROT_EXECUTE) != 0) {
2531 		pkm.pm_file = vp;
2532 		pkm.pm_address = (uintptr_t) addr;
2533 		PMC_CALL_HOOK(td, PMC_FN_MMAP, (void *) &pkm);
2534 	}
2535 #endif
2536 	return (error);
2537 }
2538