xref: /freebsd/sys/kern/vfs_vnops.c (revision 3e5645b78f476816ca3b5acc28b29bbafbb9c444)
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/fcntl.h>
50 #include <sys/file.h>
51 #include <sys/kdb.h>
52 #include <sys/stat.h>
53 #include <sys/priv.h>
54 #include <sys/proc.h>
55 #include <sys/limits.h>
56 #include <sys/lock.h>
57 #include <sys/mount.h>
58 #include <sys/mutex.h>
59 #include <sys/namei.h>
60 #include <sys/vnode.h>
61 #include <sys/bio.h>
62 #include <sys/buf.h>
63 #include <sys/filio.h>
64 #include <sys/resourcevar.h>
65 #include <sys/rwlock.h>
66 #include <sys/sx.h>
67 #include <sys/sysctl.h>
68 #include <sys/ttycom.h>
69 #include <sys/conf.h>
70 #include <sys/syslog.h>
71 #include <sys/unistd.h>
72 #include <sys/user.h>
73 
74 #include <security/audit/audit.h>
75 #include <security/mac/mac_framework.h>
76 
77 #include <vm/vm.h>
78 #include <vm/vm_extern.h>
79 #include <vm/pmap.h>
80 #include <vm/vm_map.h>
81 #include <vm/vm_object.h>
82 #include <vm/vm_page.h>
83 
84 static fo_rdwr_t	vn_read;
85 static fo_rdwr_t	vn_write;
86 static fo_rdwr_t	vn_io_fault;
87 static fo_truncate_t	vn_truncate;
88 static fo_ioctl_t	vn_ioctl;
89 static fo_poll_t	vn_poll;
90 static fo_kqfilter_t	vn_kqfilter;
91 static fo_stat_t	vn_statfile;
92 static fo_close_t	vn_closefile;
93 
94 struct 	fileops vnops = {
95 	.fo_read = vn_io_fault,
96 	.fo_write = vn_io_fault,
97 	.fo_truncate = vn_truncate,
98 	.fo_ioctl = vn_ioctl,
99 	.fo_poll = vn_poll,
100 	.fo_kqfilter = vn_kqfilter,
101 	.fo_stat = vn_statfile,
102 	.fo_close = vn_closefile,
103 	.fo_chmod = vn_chmod,
104 	.fo_chown = vn_chown,
105 	.fo_sendfile = vn_sendfile,
106 	.fo_seek = vn_seek,
107 	.fo_fill_kinfo = vn_fill_kinfo,
108 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
109 };
110 
111 static const int io_hold_cnt = 16;
112 static int vn_io_fault_enable = 1;
113 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
114     &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
115 static u_long vn_io_faults_cnt;
116 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
117     &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
118 
119 /*
120  * Returns true if vn_io_fault mode of handling the i/o request should
121  * be used.
122  */
123 static bool
124 do_vn_io_fault(struct vnode *vp, struct uio *uio)
125 {
126 	struct mount *mp;
127 
128 	return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
129 	    (mp = vp->v_mount) != NULL &&
130 	    (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
131 }
132 
133 /*
134  * Structure used to pass arguments to vn_io_fault1(), to do either
135  * file- or vnode-based I/O calls.
136  */
137 struct vn_io_fault_args {
138 	enum {
139 		VN_IO_FAULT_FOP,
140 		VN_IO_FAULT_VOP
141 	} kind;
142 	struct ucred *cred;
143 	int flags;
144 	union {
145 		struct fop_args_tag {
146 			struct file *fp;
147 			fo_rdwr_t *doio;
148 		} fop_args;
149 		struct vop_args_tag {
150 			struct vnode *vp;
151 		} vop_args;
152 	} args;
153 };
154 
155 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
156     struct vn_io_fault_args *args, struct thread *td);
157 
158 int
159 vn_open(ndp, flagp, cmode, fp)
160 	struct nameidata *ndp;
161 	int *flagp, cmode;
162 	struct file *fp;
163 {
164 	struct thread *td = ndp->ni_cnd.cn_thread;
165 
166 	return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
167 }
168 
169 /*
170  * Common code for vnode open operations via a name lookup.
171  * Lookup the vnode and invoke VOP_CREATE if needed.
172  * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
173  *
174  * Note that this does NOT free nameidata for the successful case,
175  * due to the NDINIT being done elsewhere.
176  */
177 int
178 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
179     struct ucred *cred, struct file *fp)
180 {
181 	struct vnode *vp;
182 	struct mount *mp;
183 	struct thread *td = ndp->ni_cnd.cn_thread;
184 	struct vattr vat;
185 	struct vattr *vap = &vat;
186 	int fmode, error;
187 
188 restart:
189 	fmode = *flagp;
190 	if (fmode & O_CREAT) {
191 		ndp->ni_cnd.cn_nameiop = CREATE;
192 		/*
193 		 * Set NOCACHE to avoid flushing the cache when
194 		 * rolling in many files at once.
195 		*/
196 		ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
197 		if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
198 			ndp->ni_cnd.cn_flags |= FOLLOW;
199 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
200 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
201 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
202 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
203 		bwillwrite();
204 		if ((error = namei(ndp)) != 0)
205 			return (error);
206 		if (ndp->ni_vp == NULL) {
207 			VATTR_NULL(vap);
208 			vap->va_type = VREG;
209 			vap->va_mode = cmode;
210 			if (fmode & O_EXCL)
211 				vap->va_vaflags |= VA_EXCLUSIVE;
212 			if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
213 				NDFREE(ndp, NDF_ONLY_PNBUF);
214 				vput(ndp->ni_dvp);
215 				if ((error = vn_start_write(NULL, &mp,
216 				    V_XSLEEP | PCATCH)) != 0)
217 					return (error);
218 				goto restart;
219 			}
220 			if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
221 				ndp->ni_cnd.cn_flags |= MAKEENTRY;
222 #ifdef MAC
223 			error = mac_vnode_check_create(cred, ndp->ni_dvp,
224 			    &ndp->ni_cnd, vap);
225 			if (error == 0)
226 #endif
227 				error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
228 						   &ndp->ni_cnd, vap);
229 			vput(ndp->ni_dvp);
230 			vn_finished_write(mp);
231 			if (error) {
232 				NDFREE(ndp, NDF_ONLY_PNBUF);
233 				return (error);
234 			}
235 			fmode &= ~O_TRUNC;
236 			vp = ndp->ni_vp;
237 		} else {
238 			if (ndp->ni_dvp == ndp->ni_vp)
239 				vrele(ndp->ni_dvp);
240 			else
241 				vput(ndp->ni_dvp);
242 			ndp->ni_dvp = NULL;
243 			vp = ndp->ni_vp;
244 			if (fmode & O_EXCL) {
245 				error = EEXIST;
246 				goto bad;
247 			}
248 			fmode &= ~O_CREAT;
249 		}
250 	} else {
251 		ndp->ni_cnd.cn_nameiop = LOOKUP;
252 		ndp->ni_cnd.cn_flags = ISOPEN |
253 		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
254 		if (!(fmode & FWRITE))
255 			ndp->ni_cnd.cn_flags |= LOCKSHARED;
256 		if (!(vn_open_flags & VN_OPEN_NOAUDIT))
257 			ndp->ni_cnd.cn_flags |= AUDITVNODE1;
258 		if (vn_open_flags & VN_OPEN_NOCAPCHECK)
259 			ndp->ni_cnd.cn_flags |= NOCAPCHECK;
260 		if ((error = namei(ndp)) != 0)
261 			return (error);
262 		vp = ndp->ni_vp;
263 	}
264 	error = vn_open_vnode(vp, fmode, cred, td, fp);
265 	if (error)
266 		goto bad;
267 	*flagp = fmode;
268 	return (0);
269 bad:
270 	NDFREE(ndp, NDF_ONLY_PNBUF);
271 	vput(vp);
272 	*flagp = fmode;
273 	ndp->ni_vp = NULL;
274 	return (error);
275 }
276 
277 /*
278  * Common code for vnode open operations once a vnode is located.
279  * Check permissions, and call the VOP_OPEN routine.
280  */
281 int
282 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
283     struct thread *td, struct file *fp)
284 {
285 	struct mount *mp;
286 	accmode_t accmode;
287 	struct flock lf;
288 	int error, have_flock, lock_flags, type;
289 
290 	if (vp->v_type == VLNK)
291 		return (EMLINK);
292 	if (vp->v_type == VSOCK)
293 		return (EOPNOTSUPP);
294 	if (vp->v_type != VDIR && fmode & O_DIRECTORY)
295 		return (ENOTDIR);
296 	accmode = 0;
297 	if (fmode & (FWRITE | O_TRUNC)) {
298 		if (vp->v_type == VDIR)
299 			return (EISDIR);
300 		accmode |= VWRITE;
301 	}
302 	if (fmode & FREAD)
303 		accmode |= VREAD;
304 	if (fmode & FEXEC)
305 		accmode |= VEXEC;
306 	if ((fmode & O_APPEND) && (fmode & FWRITE))
307 		accmode |= VAPPEND;
308 #ifdef MAC
309 	if (fmode & O_CREAT)
310 		accmode |= VCREAT;
311 	if (fmode & O_VERIFY)
312 		accmode |= VVERIFY;
313 	error = mac_vnode_check_open(cred, vp, accmode);
314 	if (error)
315 		return (error);
316 
317 	accmode &= ~(VCREAT | VVERIFY);
318 #endif
319 	if ((fmode & O_CREAT) == 0) {
320 		if (accmode & VWRITE) {
321 			error = vn_writechk(vp);
322 			if (error)
323 				return (error);
324 		}
325 		if (accmode) {
326 		        error = VOP_ACCESS(vp, accmode, cred, td);
327 			if (error)
328 				return (error);
329 		}
330 	}
331 	if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
332 		vn_lock(vp, LK_UPGRADE | LK_RETRY);
333 	if ((error = VOP_OPEN(vp, fmode, cred, td, fp)) != 0)
334 		return (error);
335 
336 	if (fmode & (O_EXLOCK | O_SHLOCK)) {
337 		KASSERT(fp != NULL, ("open with flock requires fp"));
338 		lock_flags = VOP_ISLOCKED(vp);
339 		VOP_UNLOCK(vp, 0);
340 		lf.l_whence = SEEK_SET;
341 		lf.l_start = 0;
342 		lf.l_len = 0;
343 		if (fmode & O_EXLOCK)
344 			lf.l_type = F_WRLCK;
345 		else
346 			lf.l_type = F_RDLCK;
347 		type = F_FLOCK;
348 		if ((fmode & FNONBLOCK) == 0)
349 			type |= F_WAIT;
350 		error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
351 		have_flock = (error == 0);
352 		vn_lock(vp, lock_flags | LK_RETRY);
353 		if (error == 0 && vp->v_iflag & VI_DOOMED)
354 			error = ENOENT;
355 		/*
356 		 * Another thread might have used this vnode as an
357 		 * executable while the vnode lock was dropped.
358 		 * Ensure the vnode is still able to be opened for
359 		 * writing after the lock has been obtained.
360 		 */
361 		if (error == 0 && accmode & VWRITE)
362 			error = vn_writechk(vp);
363 		if (error) {
364 			VOP_UNLOCK(vp, 0);
365 			if (have_flock) {
366 				lf.l_whence = SEEK_SET;
367 				lf.l_start = 0;
368 				lf.l_len = 0;
369 				lf.l_type = F_UNLCK;
370 				(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf,
371 				    F_FLOCK);
372 			}
373 			vn_start_write(vp, &mp, V_WAIT);
374 			vn_lock(vp, lock_flags | LK_RETRY);
375 			(void)VOP_CLOSE(vp, fmode, cred, td);
376 			vn_finished_write(mp);
377 			/* Prevent second close from fdrop()->vn_close(). */
378 			if (fp != NULL)
379 				fp->f_ops= &badfileops;
380 			return (error);
381 		}
382 		fp->f_flag |= FHASLOCK;
383 	}
384 	if (fmode & FWRITE) {
385 		VOP_ADD_WRITECOUNT(vp, 1);
386 		CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
387 		    __func__, vp, vp->v_writecount);
388 	}
389 	ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
390 	return (0);
391 }
392 
393 /*
394  * Check for write permissions on the specified vnode.
395  * Prototype text segments cannot be written.
396  */
397 int
398 vn_writechk(vp)
399 	register struct vnode *vp;
400 {
401 
402 	ASSERT_VOP_LOCKED(vp, "vn_writechk");
403 	/*
404 	 * If there's shared text associated with
405 	 * the vnode, try to free it up once.  If
406 	 * we fail, we can't allow writing.
407 	 */
408 	if (VOP_IS_TEXT(vp))
409 		return (ETXTBSY);
410 
411 	return (0);
412 }
413 
414 /*
415  * Vnode close call
416  */
417 int
418 vn_close(vp, flags, file_cred, td)
419 	register struct vnode *vp;
420 	int flags;
421 	struct ucred *file_cred;
422 	struct thread *td;
423 {
424 	struct mount *mp;
425 	int error, lock_flags;
426 
427 	if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
428 	    MNT_EXTENDED_SHARED(vp->v_mount))
429 		lock_flags = LK_SHARED;
430 	else
431 		lock_flags = LK_EXCLUSIVE;
432 
433 	vn_start_write(vp, &mp, V_WAIT);
434 	vn_lock(vp, lock_flags | LK_RETRY);
435 	if (flags & FWRITE) {
436 		VNASSERT(vp->v_writecount > 0, vp,
437 		    ("vn_close: negative writecount"));
438 		VOP_ADD_WRITECOUNT(vp, -1);
439 		CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
440 		    __func__, vp, vp->v_writecount);
441 	}
442 	error = VOP_CLOSE(vp, flags, file_cred, td);
443 	vput(vp);
444 	vn_finished_write(mp);
445 	return (error);
446 }
447 
448 /*
449  * Heuristic to detect sequential operation.
450  */
451 static int
452 sequential_heuristic(struct uio *uio, struct file *fp)
453 {
454 
455 	ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
456 	if (fp->f_flag & FRDAHEAD)
457 		return (fp->f_seqcount << IO_SEQSHIFT);
458 
459 	/*
460 	 * Offset 0 is handled specially.  open() sets f_seqcount to 1 so
461 	 * that the first I/O is normally considered to be slightly
462 	 * sequential.  Seeking to offset 0 doesn't change sequentiality
463 	 * unless previous seeks have reduced f_seqcount to 0, in which
464 	 * case offset 0 is not special.
465 	 */
466 	if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
467 	    uio->uio_offset == fp->f_nextoff) {
468 		/*
469 		 * f_seqcount is in units of fixed-size blocks so that it
470 		 * depends mainly on the amount of sequential I/O and not
471 		 * much on the number of sequential I/O's.  The fixed size
472 		 * of 16384 is hard-coded here since it is (not quite) just
473 		 * a magic size that works well here.  This size is more
474 		 * closely related to the best I/O size for real disks than
475 		 * to any block size used by software.
476 		 */
477 		fp->f_seqcount += howmany(uio->uio_resid, 16384);
478 		if (fp->f_seqcount > IO_SEQMAX)
479 			fp->f_seqcount = IO_SEQMAX;
480 		return (fp->f_seqcount << IO_SEQSHIFT);
481 	}
482 
483 	/* Not sequential.  Quickly draw-down sequentiality. */
484 	if (fp->f_seqcount > 1)
485 		fp->f_seqcount = 1;
486 	else
487 		fp->f_seqcount = 0;
488 	return (0);
489 }
490 
491 /*
492  * Package up an I/O request on a vnode into a uio and do it.
493  */
494 int
495 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
496     enum uio_seg segflg, int ioflg, struct ucred *active_cred,
497     struct ucred *file_cred, ssize_t *aresid, struct thread *td)
498 {
499 	struct uio auio;
500 	struct iovec aiov;
501 	struct mount *mp;
502 	struct ucred *cred;
503 	void *rl_cookie;
504 	struct vn_io_fault_args args;
505 	int error, lock_flags;
506 
507 	auio.uio_iov = &aiov;
508 	auio.uio_iovcnt = 1;
509 	aiov.iov_base = base;
510 	aiov.iov_len = len;
511 	auio.uio_resid = len;
512 	auio.uio_offset = offset;
513 	auio.uio_segflg = segflg;
514 	auio.uio_rw = rw;
515 	auio.uio_td = td;
516 	error = 0;
517 
518 	if ((ioflg & IO_NODELOCKED) == 0) {
519 		if ((ioflg & IO_RANGELOCKED) == 0) {
520 			if (rw == UIO_READ) {
521 				rl_cookie = vn_rangelock_rlock(vp, offset,
522 				    offset + len);
523 			} else {
524 				rl_cookie = vn_rangelock_wlock(vp, offset,
525 				    offset + len);
526 			}
527 		} else
528 			rl_cookie = NULL;
529 		mp = NULL;
530 		if (rw == UIO_WRITE) {
531 			if (vp->v_type != VCHR &&
532 			    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
533 			    != 0)
534 				goto out;
535 			if (MNT_SHARED_WRITES(mp) ||
536 			    ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
537 				lock_flags = LK_SHARED;
538 			else
539 				lock_flags = LK_EXCLUSIVE;
540 		} else
541 			lock_flags = LK_SHARED;
542 		vn_lock(vp, lock_flags | LK_RETRY);
543 	} else
544 		rl_cookie = NULL;
545 
546 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
547 #ifdef MAC
548 	if ((ioflg & IO_NOMACCHECK) == 0) {
549 		if (rw == UIO_READ)
550 			error = mac_vnode_check_read(active_cred, file_cred,
551 			    vp);
552 		else
553 			error = mac_vnode_check_write(active_cred, file_cred,
554 			    vp);
555 	}
556 #endif
557 	if (error == 0) {
558 		if (file_cred != NULL)
559 			cred = file_cred;
560 		else
561 			cred = active_cred;
562 		if (do_vn_io_fault(vp, &auio)) {
563 			args.kind = VN_IO_FAULT_VOP;
564 			args.cred = cred;
565 			args.flags = ioflg;
566 			args.args.vop_args.vp = vp;
567 			error = vn_io_fault1(vp, &auio, &args, td);
568 		} else if (rw == UIO_READ) {
569 			error = VOP_READ(vp, &auio, ioflg, cred);
570 		} else /* if (rw == UIO_WRITE) */ {
571 			error = VOP_WRITE(vp, &auio, ioflg, cred);
572 		}
573 	}
574 	if (aresid)
575 		*aresid = auio.uio_resid;
576 	else
577 		if (auio.uio_resid && error == 0)
578 			error = EIO;
579 	if ((ioflg & IO_NODELOCKED) == 0) {
580 		VOP_UNLOCK(vp, 0);
581 		if (mp != NULL)
582 			vn_finished_write(mp);
583 	}
584  out:
585 	if (rl_cookie != NULL)
586 		vn_rangelock_unlock(vp, rl_cookie);
587 	return (error);
588 }
589 
590 /*
591  * Package up an I/O request on a vnode into a uio and do it.  The I/O
592  * request is split up into smaller chunks and we try to avoid saturating
593  * the buffer cache while potentially holding a vnode locked, so we
594  * check bwillwrite() before calling vn_rdwr().  We also call kern_yield()
595  * to give other processes a chance to lock the vnode (either other processes
596  * core'ing the same binary, or unrelated processes scanning the directory).
597  */
598 int
599 vn_rdwr_inchunks(rw, vp, base, len, offset, segflg, ioflg, active_cred,
600     file_cred, aresid, td)
601 	enum uio_rw rw;
602 	struct vnode *vp;
603 	void *base;
604 	size_t len;
605 	off_t offset;
606 	enum uio_seg segflg;
607 	int ioflg;
608 	struct ucred *active_cred;
609 	struct ucred *file_cred;
610 	size_t *aresid;
611 	struct thread *td;
612 {
613 	int error = 0;
614 	ssize_t iaresid;
615 
616 	do {
617 		int chunk;
618 
619 		/*
620 		 * Force `offset' to a multiple of MAXBSIZE except possibly
621 		 * for the first chunk, so that filesystems only need to
622 		 * write full blocks except possibly for the first and last
623 		 * chunks.
624 		 */
625 		chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
626 
627 		if (chunk > len)
628 			chunk = len;
629 		if (rw != UIO_READ && vp->v_type == VREG)
630 			bwillwrite();
631 		iaresid = 0;
632 		error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
633 		    ioflg, active_cred, file_cred, &iaresid, td);
634 		len -= chunk;	/* aresid calc already includes length */
635 		if (error)
636 			break;
637 		offset += chunk;
638 		base = (char *)base + chunk;
639 		kern_yield(PRI_USER);
640 	} while (len);
641 	if (aresid)
642 		*aresid = len + iaresid;
643 	return (error);
644 }
645 
646 off_t
647 foffset_lock(struct file *fp, int flags)
648 {
649 	struct mtx *mtxp;
650 	off_t res;
651 
652 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
653 
654 #if OFF_MAX <= LONG_MAX
655 	/*
656 	 * Caller only wants the current f_offset value.  Assume that
657 	 * the long and shorter integer types reads are atomic.
658 	 */
659 	if ((flags & FOF_NOLOCK) != 0)
660 		return (fp->f_offset);
661 #endif
662 
663 	/*
664 	 * According to McKusick the vn lock was protecting f_offset here.
665 	 * It is now protected by the FOFFSET_LOCKED flag.
666 	 */
667 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
668 	mtx_lock(mtxp);
669 	if ((flags & FOF_NOLOCK) == 0) {
670 		while (fp->f_vnread_flags & FOFFSET_LOCKED) {
671 			fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
672 			msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
673 			    "vofflock", 0);
674 		}
675 		fp->f_vnread_flags |= FOFFSET_LOCKED;
676 	}
677 	res = fp->f_offset;
678 	mtx_unlock(mtxp);
679 	return (res);
680 }
681 
682 void
683 foffset_unlock(struct file *fp, off_t val, int flags)
684 {
685 	struct mtx *mtxp;
686 
687 	KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
688 
689 #if OFF_MAX <= LONG_MAX
690 	if ((flags & FOF_NOLOCK) != 0) {
691 		if ((flags & FOF_NOUPDATE) == 0)
692 			fp->f_offset = val;
693 		if ((flags & FOF_NEXTOFF) != 0)
694 			fp->f_nextoff = val;
695 		return;
696 	}
697 #endif
698 
699 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
700 	mtx_lock(mtxp);
701 	if ((flags & FOF_NOUPDATE) == 0)
702 		fp->f_offset = val;
703 	if ((flags & FOF_NEXTOFF) != 0)
704 		fp->f_nextoff = val;
705 	if ((flags & FOF_NOLOCK) == 0) {
706 		KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
707 		    ("Lost FOFFSET_LOCKED"));
708 		if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
709 			wakeup(&fp->f_vnread_flags);
710 		fp->f_vnread_flags = 0;
711 	}
712 	mtx_unlock(mtxp);
713 }
714 
715 void
716 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
717 {
718 
719 	if ((flags & FOF_OFFSET) == 0)
720 		uio->uio_offset = foffset_lock(fp, flags);
721 }
722 
723 void
724 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
725 {
726 
727 	if ((flags & FOF_OFFSET) == 0)
728 		foffset_unlock(fp, uio->uio_offset, flags);
729 }
730 
731 static int
732 get_advice(struct file *fp, struct uio *uio)
733 {
734 	struct mtx *mtxp;
735 	int ret;
736 
737 	ret = POSIX_FADV_NORMAL;
738 	if (fp->f_advice == NULL)
739 		return (ret);
740 
741 	mtxp = mtx_pool_find(mtxpool_sleep, fp);
742 	mtx_lock(mtxp);
743 	if (uio->uio_offset >= fp->f_advice->fa_start &&
744 	    uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
745 		ret = fp->f_advice->fa_advice;
746 	mtx_unlock(mtxp);
747 	return (ret);
748 }
749 
750 /*
751  * File table vnode read routine.
752  */
753 static int
754 vn_read(fp, uio, active_cred, flags, td)
755 	struct file *fp;
756 	struct uio *uio;
757 	struct ucred *active_cred;
758 	int flags;
759 	struct thread *td;
760 {
761 	struct vnode *vp;
762 	struct mtx *mtxp;
763 	int error, ioflag;
764 	int advice;
765 	off_t offset, start, end;
766 
767 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
768 	    uio->uio_td, td));
769 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
770 	vp = fp->f_vnode;
771 	ioflag = 0;
772 	if (fp->f_flag & FNONBLOCK)
773 		ioflag |= IO_NDELAY;
774 	if (fp->f_flag & O_DIRECT)
775 		ioflag |= IO_DIRECT;
776 	advice = get_advice(fp, uio);
777 	vn_lock(vp, LK_SHARED | LK_RETRY);
778 
779 	switch (advice) {
780 	case POSIX_FADV_NORMAL:
781 	case POSIX_FADV_SEQUENTIAL:
782 	case POSIX_FADV_NOREUSE:
783 		ioflag |= sequential_heuristic(uio, fp);
784 		break;
785 	case POSIX_FADV_RANDOM:
786 		/* Disable read-ahead for random I/O. */
787 		break;
788 	}
789 	offset = uio->uio_offset;
790 
791 #ifdef MAC
792 	error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
793 	if (error == 0)
794 #endif
795 		error = VOP_READ(vp, uio, ioflag, fp->f_cred);
796 	fp->f_nextoff = uio->uio_offset;
797 	VOP_UNLOCK(vp, 0);
798 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
799 	    offset != uio->uio_offset) {
800 		/*
801 		 * Use POSIX_FADV_DONTNEED to flush clean pages and
802 		 * buffers for the backing file after a
803 		 * POSIX_FADV_NOREUSE read(2).  To optimize the common
804 		 * case of using POSIX_FADV_NOREUSE with sequential
805 		 * access, track the previous implicit DONTNEED
806 		 * request and grow this request to include the
807 		 * current read(2) in addition to the previous
808 		 * DONTNEED.  With purely sequential access this will
809 		 * cause the DONTNEED requests to continously grow to
810 		 * cover all of the previously read regions of the
811 		 * file.  This allows filesystem blocks that are
812 		 * accessed by multiple calls to read(2) to be flushed
813 		 * once the last read(2) finishes.
814 		 */
815 		start = offset;
816 		end = uio->uio_offset - 1;
817 		mtxp = mtx_pool_find(mtxpool_sleep, fp);
818 		mtx_lock(mtxp);
819 		if (fp->f_advice != NULL &&
820 		    fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
821 			if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
822 				start = fp->f_advice->fa_prevstart;
823 			else if (fp->f_advice->fa_prevstart != 0 &&
824 			    fp->f_advice->fa_prevstart == end + 1)
825 				end = fp->f_advice->fa_prevend;
826 			fp->f_advice->fa_prevstart = start;
827 			fp->f_advice->fa_prevend = end;
828 		}
829 		mtx_unlock(mtxp);
830 		error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
831 	}
832 	return (error);
833 }
834 
835 /*
836  * File table vnode write routine.
837  */
838 static int
839 vn_write(fp, uio, active_cred, flags, td)
840 	struct file *fp;
841 	struct uio *uio;
842 	struct ucred *active_cred;
843 	int flags;
844 	struct thread *td;
845 {
846 	struct vnode *vp;
847 	struct mount *mp;
848 	struct mtx *mtxp;
849 	int error, ioflag, lock_flags;
850 	int advice;
851 	off_t offset, start, end;
852 
853 	KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
854 	    uio->uio_td, td));
855 	KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
856 	vp = fp->f_vnode;
857 	if (vp->v_type == VREG)
858 		bwillwrite();
859 	ioflag = IO_UNIT;
860 	if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
861 		ioflag |= IO_APPEND;
862 	if (fp->f_flag & FNONBLOCK)
863 		ioflag |= IO_NDELAY;
864 	if (fp->f_flag & O_DIRECT)
865 		ioflag |= IO_DIRECT;
866 	if ((fp->f_flag & O_FSYNC) ||
867 	    (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
868 		ioflag |= IO_SYNC;
869 	mp = NULL;
870 	if (vp->v_type != VCHR &&
871 	    (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
872 		goto unlock;
873 
874 	advice = get_advice(fp, uio);
875 
876 	if (MNT_SHARED_WRITES(mp) ||
877 	    (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
878 		lock_flags = LK_SHARED;
879 	} else {
880 		lock_flags = LK_EXCLUSIVE;
881 	}
882 
883 	vn_lock(vp, lock_flags | LK_RETRY);
884 	switch (advice) {
885 	case POSIX_FADV_NORMAL:
886 	case POSIX_FADV_SEQUENTIAL:
887 	case POSIX_FADV_NOREUSE:
888 		ioflag |= sequential_heuristic(uio, fp);
889 		break;
890 	case POSIX_FADV_RANDOM:
891 		/* XXX: Is this correct? */
892 		break;
893 	}
894 	offset = uio->uio_offset;
895 
896 #ifdef MAC
897 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
898 	if (error == 0)
899 #endif
900 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
901 	fp->f_nextoff = uio->uio_offset;
902 	VOP_UNLOCK(vp, 0);
903 	if (vp->v_type != VCHR)
904 		vn_finished_write(mp);
905 	if (error == 0 && advice == POSIX_FADV_NOREUSE &&
906 	    offset != uio->uio_offset) {
907 		/*
908 		 * Use POSIX_FADV_DONTNEED to flush clean pages and
909 		 * buffers for the backing file after a
910 		 * POSIX_FADV_NOREUSE write(2).  To optimize the
911 		 * common case of using POSIX_FADV_NOREUSE with
912 		 * sequential access, track the previous implicit
913 		 * DONTNEED request and grow this request to include
914 		 * the current write(2) in addition to the previous
915 		 * DONTNEED.  With purely sequential access this will
916 		 * cause the DONTNEED requests to continously grow to
917 		 * cover all of the previously written regions of the
918 		 * file.
919 		 *
920 		 * Note that the blocks just written are almost
921 		 * certainly still dirty, so this only works when
922 		 * VOP_ADVISE() calls from subsequent writes push out
923 		 * the data written by this write(2) once the backing
924 		 * buffers are clean.  However, as compared to forcing
925 		 * IO_DIRECT, this gives much saner behavior.  Write
926 		 * clustering is still allowed, and clean pages are
927 		 * merely moved to the cache page queue rather than
928 		 * outright thrown away.  This means a subsequent
929 		 * read(2) can still avoid hitting the disk if the
930 		 * pages have not been reclaimed.
931 		 *
932 		 * This does make POSIX_FADV_NOREUSE largely useless
933 		 * with non-sequential access.  However, sequential
934 		 * access is the more common use case and the flag is
935 		 * merely advisory.
936 		 */
937 		start = offset;
938 		end = uio->uio_offset - 1;
939 		mtxp = mtx_pool_find(mtxpool_sleep, fp);
940 		mtx_lock(mtxp);
941 		if (fp->f_advice != NULL &&
942 		    fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
943 			if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
944 				start = fp->f_advice->fa_prevstart;
945 			else if (fp->f_advice->fa_prevstart != 0 &&
946 			    fp->f_advice->fa_prevstart == end + 1)
947 				end = fp->f_advice->fa_prevend;
948 			fp->f_advice->fa_prevstart = start;
949 			fp->f_advice->fa_prevend = end;
950 		}
951 		mtx_unlock(mtxp);
952 		error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
953 	}
954 
955 unlock:
956 	return (error);
957 }
958 
959 /*
960  * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
961  * prevent the following deadlock:
962  *
963  * Assume that the thread A reads from the vnode vp1 into userspace
964  * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
965  * currently not resident, then system ends up with the call chain
966  *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
967  *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
968  * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
969  * If, at the same time, thread B reads from vnode vp2 into buffer buf2
970  * backed by the pages of vnode vp1, and some page in buf2 is not
971  * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
972  *
973  * To prevent the lock order reversal and deadlock, vn_io_fault() does
974  * not allow page faults to happen during VOP_READ() or VOP_WRITE().
975  * Instead, it first tries to do the whole range i/o with pagefaults
976  * disabled. If all pages in the i/o buffer are resident and mapped,
977  * VOP will succeed (ignoring the genuine filesystem errors).
978  * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
979  * i/o in chunks, with all pages in the chunk prefaulted and held
980  * using vm_fault_quick_hold_pages().
981  *
982  * Filesystems using this deadlock avoidance scheme should use the
983  * array of the held pages from uio, saved in the curthread->td_ma,
984  * instead of doing uiomove().  A helper function
985  * vn_io_fault_uiomove() converts uiomove request into
986  * uiomove_fromphys() over td_ma array.
987  *
988  * Since vnode locks do not cover the whole i/o anymore, rangelocks
989  * make the current i/o request atomic with respect to other i/os and
990  * truncations.
991  */
992 
993 /*
994  * Decode vn_io_fault_args and perform the corresponding i/o.
995  */
996 static int
997 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
998     struct thread *td)
999 {
1000 
1001 	switch (args->kind) {
1002 	case VN_IO_FAULT_FOP:
1003 		return ((args->args.fop_args.doio)(args->args.fop_args.fp,
1004 		    uio, args->cred, args->flags, td));
1005 	case VN_IO_FAULT_VOP:
1006 		if (uio->uio_rw == UIO_READ) {
1007 			return (VOP_READ(args->args.vop_args.vp, uio,
1008 			    args->flags, args->cred));
1009 		} else if (uio->uio_rw == UIO_WRITE) {
1010 			return (VOP_WRITE(args->args.vop_args.vp, uio,
1011 			    args->flags, args->cred));
1012 		}
1013 		break;
1014 	}
1015 	panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind,
1016 	    uio->uio_rw);
1017 }
1018 
1019 /*
1020  * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1021  * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1022  * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1023  * into args and call vn_io_fault1() to handle faults during the user
1024  * mode buffer accesses.
1025  */
1026 static int
1027 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1028     struct thread *td)
1029 {
1030 	vm_page_t ma[io_hold_cnt + 2];
1031 	struct uio *uio_clone, short_uio;
1032 	struct iovec short_iovec[1];
1033 	vm_page_t *prev_td_ma;
1034 	vm_prot_t prot;
1035 	vm_offset_t addr, end;
1036 	size_t len, resid;
1037 	ssize_t adv;
1038 	int error, cnt, save, saveheld, prev_td_ma_cnt;
1039 
1040 	prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1041 
1042 	/*
1043 	 * The UFS follows IO_UNIT directive and replays back both
1044 	 * uio_offset and uio_resid if an error is encountered during the
1045 	 * operation.  But, since the iovec may be already advanced,
1046 	 * uio is still in an inconsistent state.
1047 	 *
1048 	 * Cache a copy of the original uio, which is advanced to the redo
1049 	 * point using UIO_NOCOPY below.
1050 	 */
1051 	uio_clone = cloneuio(uio);
1052 	resid = uio->uio_resid;
1053 
1054 	short_uio.uio_segflg = UIO_USERSPACE;
1055 	short_uio.uio_rw = uio->uio_rw;
1056 	short_uio.uio_td = uio->uio_td;
1057 
1058 	save = vm_fault_disable_pagefaults();
1059 	error = vn_io_fault_doio(args, uio, td);
1060 	if (error != EFAULT)
1061 		goto out;
1062 
1063 	atomic_add_long(&vn_io_faults_cnt, 1);
1064 	uio_clone->uio_segflg = UIO_NOCOPY;
1065 	uiomove(NULL, resid - uio->uio_resid, uio_clone);
1066 	uio_clone->uio_segflg = uio->uio_segflg;
1067 
1068 	saveheld = curthread_pflags_set(TDP_UIOHELD);
1069 	prev_td_ma = td->td_ma;
1070 	prev_td_ma_cnt = td->td_ma_cnt;
1071 
1072 	while (uio_clone->uio_resid != 0) {
1073 		len = uio_clone->uio_iov->iov_len;
1074 		if (len == 0) {
1075 			KASSERT(uio_clone->uio_iovcnt >= 1,
1076 			    ("iovcnt underflow"));
1077 			uio_clone->uio_iov++;
1078 			uio_clone->uio_iovcnt--;
1079 			continue;
1080 		}
1081 		if (len > io_hold_cnt * PAGE_SIZE)
1082 			len = io_hold_cnt * PAGE_SIZE;
1083 		addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1084 		end = round_page(addr + len);
1085 		if (end < addr) {
1086 			error = EFAULT;
1087 			break;
1088 		}
1089 		cnt = atop(end - trunc_page(addr));
1090 		/*
1091 		 * A perfectly misaligned address and length could cause
1092 		 * both the start and the end of the chunk to use partial
1093 		 * page.  +2 accounts for such a situation.
1094 		 */
1095 		cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1096 		    addr, len, prot, ma, io_hold_cnt + 2);
1097 		if (cnt == -1) {
1098 			error = EFAULT;
1099 			break;
1100 		}
1101 		short_uio.uio_iov = &short_iovec[0];
1102 		short_iovec[0].iov_base = (void *)addr;
1103 		short_uio.uio_iovcnt = 1;
1104 		short_uio.uio_resid = short_iovec[0].iov_len = len;
1105 		short_uio.uio_offset = uio_clone->uio_offset;
1106 		td->td_ma = ma;
1107 		td->td_ma_cnt = cnt;
1108 
1109 		error = vn_io_fault_doio(args, &short_uio, td);
1110 		vm_page_unhold_pages(ma, cnt);
1111 		adv = len - short_uio.uio_resid;
1112 
1113 		uio_clone->uio_iov->iov_base =
1114 		    (char *)uio_clone->uio_iov->iov_base + adv;
1115 		uio_clone->uio_iov->iov_len -= adv;
1116 		uio_clone->uio_resid -= adv;
1117 		uio_clone->uio_offset += adv;
1118 
1119 		uio->uio_resid -= adv;
1120 		uio->uio_offset += adv;
1121 
1122 		if (error != 0 || adv == 0)
1123 			break;
1124 	}
1125 	td->td_ma = prev_td_ma;
1126 	td->td_ma_cnt = prev_td_ma_cnt;
1127 	curthread_pflags_restore(saveheld);
1128 out:
1129 	vm_fault_enable_pagefaults(save);
1130 	free(uio_clone, M_IOV);
1131 	return (error);
1132 }
1133 
1134 static int
1135 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1136     int flags, struct thread *td)
1137 {
1138 	fo_rdwr_t *doio;
1139 	struct vnode *vp;
1140 	void *rl_cookie;
1141 	struct vn_io_fault_args args;
1142 	int error;
1143 
1144 	doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1145 	vp = fp->f_vnode;
1146 	foffset_lock_uio(fp, uio, flags);
1147 	if (do_vn_io_fault(vp, uio)) {
1148 		args.kind = VN_IO_FAULT_FOP;
1149 		args.args.fop_args.fp = fp;
1150 		args.args.fop_args.doio = doio;
1151 		args.cred = active_cred;
1152 		args.flags = flags | FOF_OFFSET;
1153 		if (uio->uio_rw == UIO_READ) {
1154 			rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1155 			    uio->uio_offset + uio->uio_resid);
1156 		} else if ((fp->f_flag & O_APPEND) != 0 ||
1157 		    (flags & FOF_OFFSET) == 0) {
1158 			/* For appenders, punt and lock the whole range. */
1159 			rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1160 		} else {
1161 			rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1162 			    uio->uio_offset + uio->uio_resid);
1163 		}
1164 		error = vn_io_fault1(vp, uio, &args, td);
1165 		vn_rangelock_unlock(vp, rl_cookie);
1166 	} else {
1167 		error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1168 	}
1169 	foffset_unlock_uio(fp, uio, flags);
1170 	return (error);
1171 }
1172 
1173 /*
1174  * Helper function to perform the requested uiomove operation using
1175  * the held pages for io->uio_iov[0].iov_base buffer instead of
1176  * copyin/copyout.  Access to the pages with uiomove_fromphys()
1177  * instead of iov_base prevents page faults that could occur due to
1178  * pmap_collect() invalidating the mapping created by
1179  * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1180  * object cleanup revoking the write access from page mappings.
1181  *
1182  * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1183  * instead of plain uiomove().
1184  */
1185 int
1186 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1187 {
1188 	struct uio transp_uio;
1189 	struct iovec transp_iov[1];
1190 	struct thread *td;
1191 	size_t adv;
1192 	int error, pgadv;
1193 
1194 	td = curthread;
1195 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1196 	    uio->uio_segflg != UIO_USERSPACE)
1197 		return (uiomove(data, xfersize, uio));
1198 
1199 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1200 	transp_iov[0].iov_base = data;
1201 	transp_uio.uio_iov = &transp_iov[0];
1202 	transp_uio.uio_iovcnt = 1;
1203 	if (xfersize > uio->uio_resid)
1204 		xfersize = uio->uio_resid;
1205 	transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1206 	transp_uio.uio_offset = 0;
1207 	transp_uio.uio_segflg = UIO_SYSSPACE;
1208 	/*
1209 	 * Since transp_iov points to data, and td_ma page array
1210 	 * corresponds to original uio->uio_iov, we need to invert the
1211 	 * direction of the i/o operation as passed to
1212 	 * uiomove_fromphys().
1213 	 */
1214 	switch (uio->uio_rw) {
1215 	case UIO_WRITE:
1216 		transp_uio.uio_rw = UIO_READ;
1217 		break;
1218 	case UIO_READ:
1219 		transp_uio.uio_rw = UIO_WRITE;
1220 		break;
1221 	}
1222 	transp_uio.uio_td = uio->uio_td;
1223 	error = uiomove_fromphys(td->td_ma,
1224 	    ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1225 	    xfersize, &transp_uio);
1226 	adv = xfersize - transp_uio.uio_resid;
1227 	pgadv =
1228 	    (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1229 	    (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1230 	td->td_ma += pgadv;
1231 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1232 	    pgadv));
1233 	td->td_ma_cnt -= pgadv;
1234 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1235 	uio->uio_iov->iov_len -= adv;
1236 	uio->uio_resid -= adv;
1237 	uio->uio_offset += adv;
1238 	return (error);
1239 }
1240 
1241 int
1242 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1243     struct uio *uio)
1244 {
1245 	struct thread *td;
1246 	vm_offset_t iov_base;
1247 	int cnt, pgadv;
1248 
1249 	td = curthread;
1250 	if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1251 	    uio->uio_segflg != UIO_USERSPACE)
1252 		return (uiomove_fromphys(ma, offset, xfersize, uio));
1253 
1254 	KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1255 	cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1256 	iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1257 	switch (uio->uio_rw) {
1258 	case UIO_WRITE:
1259 		pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1260 		    offset, cnt);
1261 		break;
1262 	case UIO_READ:
1263 		pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1264 		    cnt);
1265 		break;
1266 	}
1267 	pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1268 	td->td_ma += pgadv;
1269 	KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1270 	    pgadv));
1271 	td->td_ma_cnt -= pgadv;
1272 	uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1273 	uio->uio_iov->iov_len -= cnt;
1274 	uio->uio_resid -= cnt;
1275 	uio->uio_offset += cnt;
1276 	return (0);
1277 }
1278 
1279 
1280 /*
1281  * File table truncate routine.
1282  */
1283 static int
1284 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1285     struct thread *td)
1286 {
1287 	struct vattr vattr;
1288 	struct mount *mp;
1289 	struct vnode *vp;
1290 	void *rl_cookie;
1291 	int error;
1292 
1293 	vp = fp->f_vnode;
1294 
1295 	/*
1296 	 * Lock the whole range for truncation.  Otherwise split i/o
1297 	 * might happen partly before and partly after the truncation.
1298 	 */
1299 	rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1300 	error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1301 	if (error)
1302 		goto out1;
1303 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1304 	if (vp->v_type == VDIR) {
1305 		error = EISDIR;
1306 		goto out;
1307 	}
1308 #ifdef MAC
1309 	error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1310 	if (error)
1311 		goto out;
1312 #endif
1313 	error = vn_writechk(vp);
1314 	if (error == 0) {
1315 		VATTR_NULL(&vattr);
1316 		vattr.va_size = length;
1317 		error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1318 	}
1319 out:
1320 	VOP_UNLOCK(vp, 0);
1321 	vn_finished_write(mp);
1322 out1:
1323 	vn_rangelock_unlock(vp, rl_cookie);
1324 	return (error);
1325 }
1326 
1327 /*
1328  * File table vnode stat routine.
1329  */
1330 static int
1331 vn_statfile(fp, sb, active_cred, td)
1332 	struct file *fp;
1333 	struct stat *sb;
1334 	struct ucred *active_cred;
1335 	struct thread *td;
1336 {
1337 	struct vnode *vp = fp->f_vnode;
1338 	int error;
1339 
1340 	vn_lock(vp, LK_SHARED | LK_RETRY);
1341 	error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1342 	VOP_UNLOCK(vp, 0);
1343 
1344 	return (error);
1345 }
1346 
1347 /*
1348  * Stat a vnode; implementation for the stat syscall
1349  */
1350 int
1351 vn_stat(vp, sb, active_cred, file_cred, td)
1352 	struct vnode *vp;
1353 	register struct stat *sb;
1354 	struct ucred *active_cred;
1355 	struct ucred *file_cred;
1356 	struct thread *td;
1357 {
1358 	struct vattr vattr;
1359 	register struct vattr *vap;
1360 	int error;
1361 	u_short mode;
1362 
1363 #ifdef MAC
1364 	error = mac_vnode_check_stat(active_cred, file_cred, vp);
1365 	if (error)
1366 		return (error);
1367 #endif
1368 
1369 	vap = &vattr;
1370 
1371 	/*
1372 	 * Initialize defaults for new and unusual fields, so that file
1373 	 * systems which don't support these fields don't need to know
1374 	 * about them.
1375 	 */
1376 	vap->va_birthtime.tv_sec = -1;
1377 	vap->va_birthtime.tv_nsec = 0;
1378 	vap->va_fsid = VNOVAL;
1379 	vap->va_rdev = NODEV;
1380 
1381 	error = VOP_GETATTR(vp, vap, active_cred);
1382 	if (error)
1383 		return (error);
1384 
1385 	/*
1386 	 * Zero the spare stat fields
1387 	 */
1388 	bzero(sb, sizeof *sb);
1389 
1390 	/*
1391 	 * Copy from vattr table
1392 	 */
1393 	if (vap->va_fsid != VNOVAL)
1394 		sb->st_dev = vap->va_fsid;
1395 	else
1396 		sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1397 	sb->st_ino = vap->va_fileid;
1398 	mode = vap->va_mode;
1399 	switch (vap->va_type) {
1400 	case VREG:
1401 		mode |= S_IFREG;
1402 		break;
1403 	case VDIR:
1404 		mode |= S_IFDIR;
1405 		break;
1406 	case VBLK:
1407 		mode |= S_IFBLK;
1408 		break;
1409 	case VCHR:
1410 		mode |= S_IFCHR;
1411 		break;
1412 	case VLNK:
1413 		mode |= S_IFLNK;
1414 		break;
1415 	case VSOCK:
1416 		mode |= S_IFSOCK;
1417 		break;
1418 	case VFIFO:
1419 		mode |= S_IFIFO;
1420 		break;
1421 	default:
1422 		return (EBADF);
1423 	};
1424 	sb->st_mode = mode;
1425 	sb->st_nlink = vap->va_nlink;
1426 	sb->st_uid = vap->va_uid;
1427 	sb->st_gid = vap->va_gid;
1428 	sb->st_rdev = vap->va_rdev;
1429 	if (vap->va_size > OFF_MAX)
1430 		return (EOVERFLOW);
1431 	sb->st_size = vap->va_size;
1432 	sb->st_atim = vap->va_atime;
1433 	sb->st_mtim = vap->va_mtime;
1434 	sb->st_ctim = vap->va_ctime;
1435 	sb->st_birthtim = vap->va_birthtime;
1436 
1437         /*
1438 	 * According to www.opengroup.org, the meaning of st_blksize is
1439 	 *   "a filesystem-specific preferred I/O block size for this
1440 	 *    object.  In some filesystem types, this may vary from file
1441 	 *    to file"
1442 	 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1443 	 */
1444 
1445 	sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1446 
1447 	sb->st_flags = vap->va_flags;
1448 	if (priv_check(td, PRIV_VFS_GENERATION))
1449 		sb->st_gen = 0;
1450 	else
1451 		sb->st_gen = vap->va_gen;
1452 
1453 	sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1454 	return (0);
1455 }
1456 
1457 /*
1458  * File table vnode ioctl routine.
1459  */
1460 static int
1461 vn_ioctl(fp, com, data, active_cred, td)
1462 	struct file *fp;
1463 	u_long com;
1464 	void *data;
1465 	struct ucred *active_cred;
1466 	struct thread *td;
1467 {
1468 	struct vattr vattr;
1469 	struct vnode *vp;
1470 	int error;
1471 
1472 	vp = fp->f_vnode;
1473 	switch (vp->v_type) {
1474 	case VDIR:
1475 	case VREG:
1476 		switch (com) {
1477 		case FIONREAD:
1478 			vn_lock(vp, LK_SHARED | LK_RETRY);
1479 			error = VOP_GETATTR(vp, &vattr, active_cred);
1480 			VOP_UNLOCK(vp, 0);
1481 			if (error == 0)
1482 				*(int *)data = vattr.va_size - fp->f_offset;
1483 			return (error);
1484 		case FIONBIO:
1485 		case FIOASYNC:
1486 			return (0);
1487 		default:
1488 			return (VOP_IOCTL(vp, com, data, fp->f_flag,
1489 			    active_cred, td));
1490 		}
1491 	default:
1492 		return (ENOTTY);
1493 	}
1494 }
1495 
1496 /*
1497  * File table vnode poll routine.
1498  */
1499 static int
1500 vn_poll(fp, events, active_cred, td)
1501 	struct file *fp;
1502 	int events;
1503 	struct ucred *active_cred;
1504 	struct thread *td;
1505 {
1506 	struct vnode *vp;
1507 	int error;
1508 
1509 	vp = fp->f_vnode;
1510 #ifdef MAC
1511 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1512 	error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1513 	VOP_UNLOCK(vp, 0);
1514 	if (!error)
1515 #endif
1516 
1517 	error = VOP_POLL(vp, events, fp->f_cred, td);
1518 	return (error);
1519 }
1520 
1521 /*
1522  * Acquire the requested lock and then check for validity.  LK_RETRY
1523  * permits vn_lock to return doomed vnodes.
1524  */
1525 int
1526 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1527 {
1528 	int error;
1529 
1530 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1531 	    ("vn_lock called with no locktype."));
1532 	do {
1533 #ifdef DEBUG_VFS_LOCKS
1534 		KASSERT(vp->v_holdcnt != 0,
1535 		    ("vn_lock %p: zero hold count", vp));
1536 #endif
1537 		error = VOP_LOCK1(vp, flags, file, line);
1538 		flags &= ~LK_INTERLOCK;	/* Interlock is always dropped. */
1539 		KASSERT((flags & LK_RETRY) == 0 || error == 0,
1540 		    ("LK_RETRY set with incompatible flags (0x%x) or an error occured (%d)",
1541 		    flags, error));
1542 		/*
1543 		 * Callers specify LK_RETRY if they wish to get dead vnodes.
1544 		 * If RETRY is not set, we return ENOENT instead.
1545 		 */
1546 		if (error == 0 && vp->v_iflag & VI_DOOMED &&
1547 		    (flags & LK_RETRY) == 0) {
1548 			VOP_UNLOCK(vp, 0);
1549 			error = ENOENT;
1550 			break;
1551 		}
1552 	} while (flags & LK_RETRY && error != 0);
1553 	return (error);
1554 }
1555 
1556 /*
1557  * File table vnode close routine.
1558  */
1559 static int
1560 vn_closefile(fp, td)
1561 	struct file *fp;
1562 	struct thread *td;
1563 {
1564 	struct vnode *vp;
1565 	struct flock lf;
1566 	int error;
1567 
1568 	vp = fp->f_vnode;
1569 	fp->f_ops = &badfileops;
1570 
1571 	if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK)
1572 		vref(vp);
1573 
1574 	error = vn_close(vp, fp->f_flag, fp->f_cred, td);
1575 
1576 	if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK) {
1577 		lf.l_whence = SEEK_SET;
1578 		lf.l_start = 0;
1579 		lf.l_len = 0;
1580 		lf.l_type = F_UNLCK;
1581 		(void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1582 		vrele(vp);
1583 	}
1584 	return (error);
1585 }
1586 
1587 static bool
1588 vn_suspendable_mp(struct mount *mp)
1589 {
1590 
1591 	return ((mp->mnt_kern_flag & MNTK_SUSPENDABLE) != 0);
1592 }
1593 
1594 static bool
1595 vn_suspendable(struct vnode *vp, struct mount **mpp)
1596 {
1597 
1598 	if (vp != NULL)
1599 		*mpp = vp->v_mount;
1600 	if (*mpp == NULL)
1601 		return (false);
1602 
1603 	return (vn_suspendable_mp(*mpp));
1604 }
1605 
1606 /*
1607  * Preparing to start a filesystem write operation. If the operation is
1608  * permitted, then we bump the count of operations in progress and
1609  * proceed. If a suspend request is in progress, we wait until the
1610  * suspension is over, and then proceed.
1611  */
1612 static int
1613 vn_start_write_locked(struct mount *mp, int flags)
1614 {
1615 	int error, mflags;
1616 
1617 	mtx_assert(MNT_MTX(mp), MA_OWNED);
1618 	error = 0;
1619 
1620 	/*
1621 	 * Check on status of suspension.
1622 	 */
1623 	if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1624 	    mp->mnt_susp_owner != curthread) {
1625 		mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1626 		    (flags & PCATCH) : 0) | (PUSER - 1);
1627 		while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1628 			if (flags & V_NOWAIT) {
1629 				error = EWOULDBLOCK;
1630 				goto unlock;
1631 			}
1632 			error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1633 			    "suspfs", 0);
1634 			if (error)
1635 				goto unlock;
1636 		}
1637 	}
1638 	if (flags & V_XSLEEP)
1639 		goto unlock;
1640 	mp->mnt_writeopcount++;
1641 unlock:
1642 	if (error != 0 || (flags & V_XSLEEP) != 0)
1643 		MNT_REL(mp);
1644 	MNT_IUNLOCK(mp);
1645 	return (error);
1646 }
1647 
1648 int
1649 vn_start_write(vp, mpp, flags)
1650 	struct vnode *vp;
1651 	struct mount **mpp;
1652 	int flags;
1653 {
1654 	struct mount *mp;
1655 	int error;
1656 
1657 	if (!vn_suspendable(vp, mpp))
1658 		return (0);
1659 
1660 	error = 0;
1661 	/*
1662 	 * If a vnode is provided, get and return the mount point that
1663 	 * to which it will write.
1664 	 */
1665 	if (vp != NULL) {
1666 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1667 			*mpp = NULL;
1668 			if (error != EOPNOTSUPP)
1669 				return (error);
1670 			return (0);
1671 		}
1672 	}
1673 	if ((mp = *mpp) == NULL)
1674 		return (0);
1675 
1676 	/*
1677 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1678 	 * a vfs_ref().
1679 	 * As long as a vnode is not provided we need to acquire a
1680 	 * refcount for the provided mountpoint too, in order to
1681 	 * emulate a vfs_ref().
1682 	 */
1683 	MNT_ILOCK(mp);
1684 	if (vp == NULL)
1685 		MNT_REF(mp);
1686 
1687 	return (vn_start_write_locked(mp, flags));
1688 }
1689 
1690 /*
1691  * Secondary suspension. Used by operations such as vop_inactive
1692  * routines that are needed by the higher level functions. These
1693  * are allowed to proceed until all the higher level functions have
1694  * completed (indicated by mnt_writeopcount dropping to zero). At that
1695  * time, these operations are halted until the suspension is over.
1696  */
1697 int
1698 vn_start_secondary_write(vp, mpp, flags)
1699 	struct vnode *vp;
1700 	struct mount **mpp;
1701 	int flags;
1702 {
1703 	struct mount *mp;
1704 	int error;
1705 
1706 	if (!vn_suspendable(vp, mpp))
1707 		return (0);
1708 
1709  retry:
1710 	if (vp != NULL) {
1711 		if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1712 			*mpp = NULL;
1713 			if (error != EOPNOTSUPP)
1714 				return (error);
1715 			return (0);
1716 		}
1717 	}
1718 	/*
1719 	 * If we are not suspended or have not yet reached suspended
1720 	 * mode, then let the operation proceed.
1721 	 */
1722 	if ((mp = *mpp) == NULL)
1723 		return (0);
1724 
1725 	/*
1726 	 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1727 	 * a vfs_ref().
1728 	 * As long as a vnode is not provided we need to acquire a
1729 	 * refcount for the provided mountpoint too, in order to
1730 	 * emulate a vfs_ref().
1731 	 */
1732 	MNT_ILOCK(mp);
1733 	if (vp == NULL)
1734 		MNT_REF(mp);
1735 	if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1736 		mp->mnt_secondary_writes++;
1737 		mp->mnt_secondary_accwrites++;
1738 		MNT_IUNLOCK(mp);
1739 		return (0);
1740 	}
1741 	if (flags & V_NOWAIT) {
1742 		MNT_REL(mp);
1743 		MNT_IUNLOCK(mp);
1744 		return (EWOULDBLOCK);
1745 	}
1746 	/*
1747 	 * Wait for the suspension to finish.
1748 	 */
1749 	error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1750 	    ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1751 	    "suspfs", 0);
1752 	vfs_rel(mp);
1753 	if (error == 0)
1754 		goto retry;
1755 	return (error);
1756 }
1757 
1758 /*
1759  * Filesystem write operation has completed. If we are suspending and this
1760  * operation is the last one, notify the suspender that the suspension is
1761  * now in effect.
1762  */
1763 void
1764 vn_finished_write(mp)
1765 	struct mount *mp;
1766 {
1767 	if (mp == NULL || !vn_suspendable_mp(mp))
1768 		return;
1769 	MNT_ILOCK(mp);
1770 	MNT_REL(mp);
1771 	mp->mnt_writeopcount--;
1772 	if (mp->mnt_writeopcount < 0)
1773 		panic("vn_finished_write: neg cnt");
1774 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1775 	    mp->mnt_writeopcount <= 0)
1776 		wakeup(&mp->mnt_writeopcount);
1777 	MNT_IUNLOCK(mp);
1778 }
1779 
1780 
1781 /*
1782  * Filesystem secondary write operation has completed. If we are
1783  * suspending and this operation is the last one, notify the suspender
1784  * that the suspension is now in effect.
1785  */
1786 void
1787 vn_finished_secondary_write(mp)
1788 	struct mount *mp;
1789 {
1790 	if (mp == NULL || !vn_suspendable_mp(mp))
1791 		return;
1792 	MNT_ILOCK(mp);
1793 	MNT_REL(mp);
1794 	mp->mnt_secondary_writes--;
1795 	if (mp->mnt_secondary_writes < 0)
1796 		panic("vn_finished_secondary_write: neg cnt");
1797 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1798 	    mp->mnt_secondary_writes <= 0)
1799 		wakeup(&mp->mnt_secondary_writes);
1800 	MNT_IUNLOCK(mp);
1801 }
1802 
1803 
1804 
1805 /*
1806  * Request a filesystem to suspend write operations.
1807  */
1808 int
1809 vfs_write_suspend(struct mount *mp, int flags)
1810 {
1811 	int error;
1812 
1813 	MPASS(vn_suspendable_mp(mp));
1814 
1815 	MNT_ILOCK(mp);
1816 	if (mp->mnt_susp_owner == curthread) {
1817 		MNT_IUNLOCK(mp);
1818 		return (EALREADY);
1819 	}
1820 	while (mp->mnt_kern_flag & MNTK_SUSPEND)
1821 		msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1822 
1823 	/*
1824 	 * Unmount holds a write reference on the mount point.  If we
1825 	 * own busy reference and drain for writers, we deadlock with
1826 	 * the reference draining in the unmount path.  Callers of
1827 	 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1828 	 * vfs_busy() reference is owned and caller is not in the
1829 	 * unmount context.
1830 	 */
1831 	if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1832 	    (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1833 		MNT_IUNLOCK(mp);
1834 		return (EBUSY);
1835 	}
1836 
1837 	mp->mnt_kern_flag |= MNTK_SUSPEND;
1838 	mp->mnt_susp_owner = curthread;
1839 	if (mp->mnt_writeopcount > 0)
1840 		(void) msleep(&mp->mnt_writeopcount,
1841 		    MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1842 	else
1843 		MNT_IUNLOCK(mp);
1844 	if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1845 		vfs_write_resume(mp, 0);
1846 	return (error);
1847 }
1848 
1849 /*
1850  * Request a filesystem to resume write operations.
1851  */
1852 void
1853 vfs_write_resume(struct mount *mp, int flags)
1854 {
1855 
1856 	MPASS(vn_suspendable_mp(mp));
1857 
1858 	MNT_ILOCK(mp);
1859 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1860 		KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1861 		mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1862 				       MNTK_SUSPENDED);
1863 		mp->mnt_susp_owner = NULL;
1864 		wakeup(&mp->mnt_writeopcount);
1865 		wakeup(&mp->mnt_flag);
1866 		curthread->td_pflags &= ~TDP_IGNSUSP;
1867 		if ((flags & VR_START_WRITE) != 0) {
1868 			MNT_REF(mp);
1869 			mp->mnt_writeopcount++;
1870 		}
1871 		MNT_IUNLOCK(mp);
1872 		if ((flags & VR_NO_SUSPCLR) == 0)
1873 			VFS_SUSP_CLEAN(mp);
1874 	} else if ((flags & VR_START_WRITE) != 0) {
1875 		MNT_REF(mp);
1876 		vn_start_write_locked(mp, 0);
1877 	} else {
1878 		MNT_IUNLOCK(mp);
1879 	}
1880 }
1881 
1882 /*
1883  * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1884  * methods.
1885  */
1886 int
1887 vfs_write_suspend_umnt(struct mount *mp)
1888 {
1889 	int error;
1890 
1891 	MPASS(vn_suspendable_mp(mp));
1892 	KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1893 	    ("vfs_write_suspend_umnt: recursed"));
1894 
1895 	/* dounmount() already called vn_start_write(). */
1896 	for (;;) {
1897 		vn_finished_write(mp);
1898 		error = vfs_write_suspend(mp, 0);
1899 		if (error != 0) {
1900 			vn_start_write(NULL, &mp, V_WAIT);
1901 			return (error);
1902 		}
1903 		MNT_ILOCK(mp);
1904 		if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1905 			break;
1906 		MNT_IUNLOCK(mp);
1907 		vn_start_write(NULL, &mp, V_WAIT);
1908 	}
1909 	mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1910 	wakeup(&mp->mnt_flag);
1911 	MNT_IUNLOCK(mp);
1912 	curthread->td_pflags |= TDP_IGNSUSP;
1913 	return (0);
1914 }
1915 
1916 /*
1917  * Implement kqueues for files by translating it to vnode operation.
1918  */
1919 static int
1920 vn_kqfilter(struct file *fp, struct knote *kn)
1921 {
1922 
1923 	return (VOP_KQFILTER(fp->f_vnode, kn));
1924 }
1925 
1926 /*
1927  * Simplified in-kernel wrapper calls for extended attribute access.
1928  * Both calls pass in a NULL credential, authorizing as "kernel" access.
1929  * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1930  */
1931 int
1932 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1933     const char *attrname, int *buflen, char *buf, struct thread *td)
1934 {
1935 	struct uio	auio;
1936 	struct iovec	iov;
1937 	int	error;
1938 
1939 	iov.iov_len = *buflen;
1940 	iov.iov_base = buf;
1941 
1942 	auio.uio_iov = &iov;
1943 	auio.uio_iovcnt = 1;
1944 	auio.uio_rw = UIO_READ;
1945 	auio.uio_segflg = UIO_SYSSPACE;
1946 	auio.uio_td = td;
1947 	auio.uio_offset = 0;
1948 	auio.uio_resid = *buflen;
1949 
1950 	if ((ioflg & IO_NODELOCKED) == 0)
1951 		vn_lock(vp, LK_SHARED | LK_RETRY);
1952 
1953 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1954 
1955 	/* authorize attribute retrieval as kernel */
1956 	error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
1957 	    td);
1958 
1959 	if ((ioflg & IO_NODELOCKED) == 0)
1960 		VOP_UNLOCK(vp, 0);
1961 
1962 	if (error == 0) {
1963 		*buflen = *buflen - auio.uio_resid;
1964 	}
1965 
1966 	return (error);
1967 }
1968 
1969 /*
1970  * XXX failure mode if partially written?
1971  */
1972 int
1973 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
1974     const char *attrname, int buflen, char *buf, struct thread *td)
1975 {
1976 	struct uio	auio;
1977 	struct iovec	iov;
1978 	struct mount	*mp;
1979 	int	error;
1980 
1981 	iov.iov_len = buflen;
1982 	iov.iov_base = buf;
1983 
1984 	auio.uio_iov = &iov;
1985 	auio.uio_iovcnt = 1;
1986 	auio.uio_rw = UIO_WRITE;
1987 	auio.uio_segflg = UIO_SYSSPACE;
1988 	auio.uio_td = td;
1989 	auio.uio_offset = 0;
1990 	auio.uio_resid = buflen;
1991 
1992 	if ((ioflg & IO_NODELOCKED) == 0) {
1993 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
1994 			return (error);
1995 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1996 	}
1997 
1998 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1999 
2000 	/* authorize attribute setting as kernel */
2001 	error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2002 
2003 	if ((ioflg & IO_NODELOCKED) == 0) {
2004 		vn_finished_write(mp);
2005 		VOP_UNLOCK(vp, 0);
2006 	}
2007 
2008 	return (error);
2009 }
2010 
2011 int
2012 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2013     const char *attrname, struct thread *td)
2014 {
2015 	struct mount	*mp;
2016 	int	error;
2017 
2018 	if ((ioflg & IO_NODELOCKED) == 0) {
2019 		if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2020 			return (error);
2021 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2022 	}
2023 
2024 	ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2025 
2026 	/* authorize attribute removal as kernel */
2027 	error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2028 	if (error == EOPNOTSUPP)
2029 		error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2030 		    NULL, td);
2031 
2032 	if ((ioflg & IO_NODELOCKED) == 0) {
2033 		vn_finished_write(mp);
2034 		VOP_UNLOCK(vp, 0);
2035 	}
2036 
2037 	return (error);
2038 }
2039 
2040 static int
2041 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2042     struct vnode **rvp)
2043 {
2044 
2045 	return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2046 }
2047 
2048 int
2049 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2050 {
2051 
2052 	return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2053 	    lkflags, rvp));
2054 }
2055 
2056 int
2057 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2058     int lkflags, struct vnode **rvp)
2059 {
2060 	struct mount *mp;
2061 	int ltype, error;
2062 
2063 	ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2064 	mp = vp->v_mount;
2065 	ltype = VOP_ISLOCKED(vp);
2066 	KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2067 	    ("vn_vget_ino: vp not locked"));
2068 	error = vfs_busy(mp, MBF_NOWAIT);
2069 	if (error != 0) {
2070 		vfs_ref(mp);
2071 		VOP_UNLOCK(vp, 0);
2072 		error = vfs_busy(mp, 0);
2073 		vn_lock(vp, ltype | LK_RETRY);
2074 		vfs_rel(mp);
2075 		if (error != 0)
2076 			return (ENOENT);
2077 		if (vp->v_iflag & VI_DOOMED) {
2078 			vfs_unbusy(mp);
2079 			return (ENOENT);
2080 		}
2081 	}
2082 	VOP_UNLOCK(vp, 0);
2083 	error = alloc(mp, alloc_arg, lkflags, rvp);
2084 	vfs_unbusy(mp);
2085 	if (*rvp != vp)
2086 		vn_lock(vp, ltype | LK_RETRY);
2087 	if (vp->v_iflag & VI_DOOMED) {
2088 		if (error == 0) {
2089 			if (*rvp == vp)
2090 				vunref(vp);
2091 			else
2092 				vput(*rvp);
2093 		}
2094 		error = ENOENT;
2095 	}
2096 	return (error);
2097 }
2098 
2099 int
2100 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2101     const struct thread *td)
2102 {
2103 
2104 	if (vp->v_type != VREG || td == NULL)
2105 		return (0);
2106 	PROC_LOCK(td->td_proc);
2107 	if ((uoff_t)uio->uio_offset + uio->uio_resid >
2108 	    lim_cur(td->td_proc, RLIMIT_FSIZE)) {
2109 		kern_psignal(td->td_proc, SIGXFSZ);
2110 		PROC_UNLOCK(td->td_proc);
2111 		return (EFBIG);
2112 	}
2113 	PROC_UNLOCK(td->td_proc);
2114 	return (0);
2115 }
2116 
2117 int
2118 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2119     struct thread *td)
2120 {
2121 	struct vnode *vp;
2122 
2123 	vp = fp->f_vnode;
2124 #ifdef AUDIT
2125 	vn_lock(vp, LK_SHARED | LK_RETRY);
2126 	AUDIT_ARG_VNODE1(vp);
2127 	VOP_UNLOCK(vp, 0);
2128 #endif
2129 	return (setfmode(td, active_cred, vp, mode));
2130 }
2131 
2132 int
2133 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2134     struct thread *td)
2135 {
2136 	struct vnode *vp;
2137 
2138 	vp = fp->f_vnode;
2139 #ifdef AUDIT
2140 	vn_lock(vp, LK_SHARED | LK_RETRY);
2141 	AUDIT_ARG_VNODE1(vp);
2142 	VOP_UNLOCK(vp, 0);
2143 #endif
2144 	return (setfown(td, active_cred, vp, uid, gid));
2145 }
2146 
2147 void
2148 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2149 {
2150 	vm_object_t object;
2151 
2152 	if ((object = vp->v_object) == NULL)
2153 		return;
2154 	VM_OBJECT_WLOCK(object);
2155 	vm_object_page_remove(object, start, end, 0);
2156 	VM_OBJECT_WUNLOCK(object);
2157 }
2158 
2159 int
2160 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2161 {
2162 	struct vattr va;
2163 	daddr_t bn, bnp;
2164 	uint64_t bsize;
2165 	off_t noff;
2166 	int error;
2167 
2168 	KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2169 	    ("Wrong command %lu", cmd));
2170 
2171 	if (vn_lock(vp, LK_SHARED) != 0)
2172 		return (EBADF);
2173 	if (vp->v_type != VREG) {
2174 		error = ENOTTY;
2175 		goto unlock;
2176 	}
2177 	error = VOP_GETATTR(vp, &va, cred);
2178 	if (error != 0)
2179 		goto unlock;
2180 	noff = *off;
2181 	if (noff >= va.va_size) {
2182 		error = ENXIO;
2183 		goto unlock;
2184 	}
2185 	bsize = vp->v_mount->mnt_stat.f_iosize;
2186 	for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize) {
2187 		error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2188 		if (error == EOPNOTSUPP) {
2189 			error = ENOTTY;
2190 			goto unlock;
2191 		}
2192 		if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2193 		    (bnp != -1 && cmd == FIOSEEKDATA)) {
2194 			noff = bn * bsize;
2195 			if (noff < *off)
2196 				noff = *off;
2197 			goto unlock;
2198 		}
2199 	}
2200 	if (noff > va.va_size)
2201 		noff = va.va_size;
2202 	/* noff == va.va_size. There is an implicit hole at the end of file. */
2203 	if (cmd == FIOSEEKDATA)
2204 		error = ENXIO;
2205 unlock:
2206 	VOP_UNLOCK(vp, 0);
2207 	if (error == 0)
2208 		*off = noff;
2209 	return (error);
2210 }
2211 
2212 int
2213 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2214 {
2215 	struct ucred *cred;
2216 	struct vnode *vp;
2217 	struct vattr vattr;
2218 	off_t foffset, size;
2219 	int error, noneg;
2220 
2221 	cred = td->td_ucred;
2222 	vp = fp->f_vnode;
2223 	foffset = foffset_lock(fp, 0);
2224 	noneg = (vp->v_type != VCHR);
2225 	error = 0;
2226 	switch (whence) {
2227 	case L_INCR:
2228 		if (noneg &&
2229 		    (foffset < 0 ||
2230 		    (offset > 0 && foffset > OFF_MAX - offset))) {
2231 			error = EOVERFLOW;
2232 			break;
2233 		}
2234 		offset += foffset;
2235 		break;
2236 	case L_XTND:
2237 		vn_lock(vp, LK_SHARED | LK_RETRY);
2238 		error = VOP_GETATTR(vp, &vattr, cred);
2239 		VOP_UNLOCK(vp, 0);
2240 		if (error)
2241 			break;
2242 
2243 		/*
2244 		 * If the file references a disk device, then fetch
2245 		 * the media size and use that to determine the ending
2246 		 * offset.
2247 		 */
2248 		if (vattr.va_size == 0 && vp->v_type == VCHR &&
2249 		    fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2250 			vattr.va_size = size;
2251 		if (noneg &&
2252 		    (vattr.va_size > OFF_MAX ||
2253 		    (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2254 			error = EOVERFLOW;
2255 			break;
2256 		}
2257 		offset += vattr.va_size;
2258 		break;
2259 	case L_SET:
2260 		break;
2261 	case SEEK_DATA:
2262 		error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2263 		break;
2264 	case SEEK_HOLE:
2265 		error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2266 		break;
2267 	default:
2268 		error = EINVAL;
2269 	}
2270 	if (error == 0 && noneg && offset < 0)
2271 		error = EINVAL;
2272 	if (error != 0)
2273 		goto drop;
2274 	VFS_KNOTE_UNLOCKED(vp, 0);
2275 	td->td_uretoff.tdu_off = offset;
2276 drop:
2277 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2278 	return (error);
2279 }
2280 
2281 int
2282 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2283     struct thread *td)
2284 {
2285 	int error;
2286 
2287 	/*
2288 	 * Grant permission if the caller is the owner of the file, or
2289 	 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2290 	 * on the file.  If the time pointer is null, then write
2291 	 * permission on the file is also sufficient.
2292 	 *
2293 	 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2294 	 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2295 	 * will be allowed to set the times [..] to the current
2296 	 * server time.
2297 	 */
2298 	error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2299 	if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2300 		error = VOP_ACCESS(vp, VWRITE, cred, td);
2301 	return (error);
2302 }
2303 
2304 int
2305 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
2306 {
2307 	struct vnode *vp;
2308 	int error;
2309 
2310 	if (fp->f_type == DTYPE_FIFO)
2311 		kif->kf_type = KF_TYPE_FIFO;
2312 	else
2313 		kif->kf_type = KF_TYPE_VNODE;
2314 	vp = fp->f_vnode;
2315 	vref(vp);
2316 	FILEDESC_SUNLOCK(fdp);
2317 	error = vn_fill_kinfo_vnode(vp, kif);
2318 	vrele(vp);
2319 	FILEDESC_SLOCK(fdp);
2320 	return (error);
2321 }
2322 
2323 int
2324 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif)
2325 {
2326 	struct vattr va;
2327 	char *fullpath, *freepath;
2328 	int error;
2329 
2330 	kif->kf_vnode_type = vntype_to_kinfo(vp->v_type);
2331 	freepath = NULL;
2332 	fullpath = "-";
2333 	error = vn_fullpath(curthread, vp, &fullpath, &freepath);
2334 	if (error == 0) {
2335 		strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path));
2336 	}
2337 	if (freepath != NULL)
2338 		free(freepath, M_TEMP);
2339 
2340 	/*
2341 	 * Retrieve vnode attributes.
2342 	 */
2343 	va.va_fsid = VNOVAL;
2344 	va.va_rdev = NODEV;
2345 	vn_lock(vp, LK_SHARED | LK_RETRY);
2346 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
2347 	VOP_UNLOCK(vp, 0);
2348 	if (error != 0)
2349 		return (error);
2350 	if (va.va_fsid != VNOVAL)
2351 		kif->kf_un.kf_file.kf_file_fsid = va.va_fsid;
2352 	else
2353 		kif->kf_un.kf_file.kf_file_fsid =
2354 		    vp->v_mount->mnt_stat.f_fsid.val[0];
2355 	kif->kf_un.kf_file.kf_file_fileid = va.va_fileid;
2356 	kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode);
2357 	kif->kf_un.kf_file.kf_file_size = va.va_size;
2358 	kif->kf_un.kf_file.kf_file_rdev = va.va_rdev;
2359 	return (0);
2360 }
2361