xref: /freebsd/sys/fs/nullfs/null_vnops.c (revision c66ec88fed842fbaad62c30d510644ceb7bd2d71)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1992, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * John Heidemann of the UCLA Ficus project.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)null_vnops.c	8.6 (Berkeley) 5/27/95
35  *
36  * Ancestors:
37  *	@(#)lofs_vnops.c	1.2 (Berkeley) 6/18/92
38  *	...and...
39  *	@(#)null_vnodeops.c 1.20 92/07/07 UCLA Ficus project
40  *
41  * $FreeBSD$
42  */
43 
44 /*
45  * Null Layer
46  *
47  * (See mount_nullfs(8) for more information.)
48  *
49  * The null layer duplicates a portion of the filesystem
50  * name space under a new name.  In this respect, it is
51  * similar to the loopback filesystem.  It differs from
52  * the loopback fs in two respects:  it is implemented using
53  * a stackable layers techniques, and its "null-node"s stack above
54  * all lower-layer vnodes, not just over directory vnodes.
55  *
56  * The null layer has two purposes.  First, it serves as a demonstration
57  * of layering by proving a layer which does nothing.  (It actually
58  * does everything the loopback filesystem does, which is slightly
59  * more than nothing.)  Second, the null layer can serve as a prototype
60  * layer.  Since it provides all necessary layer framework,
61  * new filesystem layers can be created very easily be starting
62  * with a null layer.
63  *
64  * The remainder of this man page examines the null layer as a basis
65  * for constructing new layers.
66  *
67  *
68  * INSTANTIATING NEW NULL LAYERS
69  *
70  * New null layers are created with mount_nullfs(8).
71  * Mount_nullfs(8) takes two arguments, the pathname
72  * of the lower vfs (target-pn) and the pathname where the null
73  * layer will appear in the namespace (alias-pn).  After
74  * the null layer is put into place, the contents
75  * of target-pn subtree will be aliased under alias-pn.
76  *
77  *
78  * OPERATION OF A NULL LAYER
79  *
80  * The null layer is the minimum filesystem layer,
81  * simply bypassing all possible operations to the lower layer
82  * for processing there.  The majority of its activity centers
83  * on the bypass routine, through which nearly all vnode operations
84  * pass.
85  *
86  * The bypass routine accepts arbitrary vnode operations for
87  * handling by the lower layer.  It begins by examing vnode
88  * operation arguments and replacing any null-nodes by their
89  * lower-layer equivlants.  It then invokes the operation
90  * on the lower layer.  Finally, it replaces the null-nodes
91  * in the arguments and, if a vnode is return by the operation,
92  * stacks a null-node on top of the returned vnode.
93  *
94  * Although bypass handles most operations, vop_getattr, vop_lock,
95  * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not
96  * bypassed. Vop_getattr must change the fsid being returned.
97  * Vop_lock and vop_unlock must handle any locking for the
98  * current vnode as well as pass the lock request down.
99  * Vop_inactive and vop_reclaim are not bypassed so that
100  * they can handle freeing null-layer specific data. Vop_print
101  * is not bypassed to avoid excessive debugging information.
102  * Also, certain vnode operations change the locking state within
103  * the operation (create, mknod, remove, link, rename, mkdir, rmdir,
104  * and symlink). Ideally these operations should not change the
105  * lock state, but should be changed to let the caller of the
106  * function unlock them. Otherwise all intermediate vnode layers
107  * (such as union, umapfs, etc) must catch these functions to do
108  * the necessary locking at their layer.
109  *
110  *
111  * INSTANTIATING VNODE STACKS
112  *
113  * Mounting associates the null layer with a lower layer,
114  * effect stacking two VFSes.  Vnode stacks are instead
115  * created on demand as files are accessed.
116  *
117  * The initial mount creates a single vnode stack for the
118  * root of the new null layer.  All other vnode stacks
119  * are created as a result of vnode operations on
120  * this or other null vnode stacks.
121  *
122  * New vnode stacks come into existence as a result of
123  * an operation which returns a vnode.
124  * The bypass routine stacks a null-node above the new
125  * vnode before returning it to the caller.
126  *
127  * For example, imagine mounting a null layer with
128  * "mount_nullfs /usr/include /dev/layer/null".
129  * Changing directory to /dev/layer/null will assign
130  * the root null-node (which was created when the null layer was mounted).
131  * Now consider opening "sys".  A vop_lookup would be
132  * done on the root null-node.  This operation would bypass through
133  * to the lower layer which would return a vnode representing
134  * the UFS "sys".  Null_bypass then builds a null-node
135  * aliasing the UFS "sys" and returns this to the caller.
136  * Later operations on the null-node "sys" will repeat this
137  * process when constructing other vnode stacks.
138  *
139  *
140  * CREATING OTHER FILE SYSTEM LAYERS
141  *
142  * One of the easiest ways to construct new filesystem layers is to make
143  * a copy of the null layer, rename all files and variables, and
144  * then begin modifing the copy.  Sed can be used to easily rename
145  * all variables.
146  *
147  * The umap layer is an example of a layer descended from the
148  * null layer.
149  *
150  *
151  * INVOKING OPERATIONS ON LOWER LAYERS
152  *
153  * There are two techniques to invoke operations on a lower layer
154  * when the operation cannot be completely bypassed.  Each method
155  * is appropriate in different situations.  In both cases,
156  * it is the responsibility of the aliasing layer to make
157  * the operation arguments "correct" for the lower layer
158  * by mapping a vnode arguments to the lower layer.
159  *
160  * The first approach is to call the aliasing layer's bypass routine.
161  * This method is most suitable when you wish to invoke the operation
162  * currently being handled on the lower layer.  It has the advantage
163  * that the bypass routine already must do argument mapping.
164  * An example of this is null_getattrs in the null layer.
165  *
166  * A second approach is to directly invoke vnode operations on
167  * the lower layer with the VOP_OPERATIONNAME interface.
168  * The advantage of this method is that it is easy to invoke
169  * arbitrary operations on the lower layer.  The disadvantage
170  * is that vnode arguments must be manualy mapped.
171  *
172  */
173 
174 #include <sys/param.h>
175 #include <sys/systm.h>
176 #include <sys/conf.h>
177 #include <sys/kernel.h>
178 #include <sys/lock.h>
179 #include <sys/malloc.h>
180 #include <sys/mount.h>
181 #include <sys/mutex.h>
182 #include <sys/namei.h>
183 #include <sys/sysctl.h>
184 #include <sys/vnode.h>
185 #include <sys/stat.h>
186 
187 #include <fs/nullfs/null.h>
188 
189 #include <vm/vm.h>
190 #include <vm/vm_extern.h>
191 #include <vm/vm_object.h>
192 #include <vm/vnode_pager.h>
193 
194 static int null_bug_bypass = 0;   /* for debugging: enables bypass printf'ing */
195 SYSCTL_INT(_debug, OID_AUTO, nullfs_bug_bypass, CTLFLAG_RW,
196 	&null_bug_bypass, 0, "");
197 
198 /*
199  * This is the 10-Apr-92 bypass routine.
200  *    This version has been optimized for speed, throwing away some
201  * safety checks.  It should still always work, but it's not as
202  * robust to programmer errors.
203  *
204  * In general, we map all vnodes going down and unmap them on the way back.
205  * As an exception to this, vnodes can be marked "unmapped" by setting
206  * the Nth bit in operation's vdesc_flags.
207  *
208  * Also, some BSD vnode operations have the side effect of vrele'ing
209  * their arguments.  With stacking, the reference counts are held
210  * by the upper node, not the lower one, so we must handle these
211  * side-effects here.  This is not of concern in Sun-derived systems
212  * since there are no such side-effects.
213  *
214  * This makes the following assumptions:
215  * - only one returned vpp
216  * - no INOUT vpp's (Sun's vop_open has one of these)
217  * - the vnode operation vector of the first vnode should be used
218  *   to determine what implementation of the op should be invoked
219  * - all mapped vnodes are of our vnode-type (NEEDSWORK:
220  *   problems on rmdir'ing mount points and renaming?)
221  */
222 int
223 null_bypass(struct vop_generic_args *ap)
224 {
225 	struct vnode **this_vp_p;
226 	int error;
227 	struct vnode *old_vps[VDESC_MAX_VPS];
228 	struct vnode **vps_p[VDESC_MAX_VPS];
229 	struct vnode ***vppp;
230 	struct vnode *lvp;
231 	struct vnodeop_desc *descp = ap->a_desc;
232 	int reles, i;
233 
234 	if (null_bug_bypass)
235 		printf ("null_bypass: %s\n", descp->vdesc_name);
236 
237 #ifdef DIAGNOSTIC
238 	/*
239 	 * We require at least one vp.
240 	 */
241 	if (descp->vdesc_vp_offsets == NULL ||
242 	    descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET)
243 		panic ("null_bypass: no vp's in map");
244 #endif
245 
246 	/*
247 	 * Map the vnodes going in.
248 	 * Later, we'll invoke the operation based on
249 	 * the first mapped vnode's operation vector.
250 	 */
251 	reles = descp->vdesc_flags;
252 	for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
253 		if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
254 			break;   /* bail out at end of list */
255 		vps_p[i] = this_vp_p =
256 			VOPARG_OFFSETTO(struct vnode**,descp->vdesc_vp_offsets[i],ap);
257 		/*
258 		 * We're not guaranteed that any but the first vnode
259 		 * are of our type.  Check for and don't map any
260 		 * that aren't.  (We must always map first vp or vclean fails.)
261 		 */
262 		if (i && (*this_vp_p == NULLVP ||
263 		    (*this_vp_p)->v_op != &null_vnodeops)) {
264 			old_vps[i] = NULLVP;
265 		} else {
266 			old_vps[i] = *this_vp_p;
267 			*(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p);
268 			/*
269 			 * XXX - Several operations have the side effect
270 			 * of vrele'ing their vp's.  We must account for
271 			 * that.  (This should go away in the future.)
272 			 */
273 			if (reles & VDESC_VP0_WILLRELE)
274 				VREF(*this_vp_p);
275 		}
276 	}
277 
278 	/*
279 	 * Call the operation on the lower layer
280 	 * with the modified argument structure.
281 	 */
282 	if (vps_p[0] && *vps_p[0])
283 		error = VCALL(ap);
284 	else {
285 		printf("null_bypass: no map for %s\n", descp->vdesc_name);
286 		error = EINVAL;
287 	}
288 
289 	/*
290 	 * Maintain the illusion of call-by-value
291 	 * by restoring vnodes in the argument structure
292 	 * to their original value.
293 	 */
294 	reles = descp->vdesc_flags;
295 	for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
296 		if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
297 			break;   /* bail out at end of list */
298 		if (old_vps[i]) {
299 			lvp = *(vps_p[i]);
300 
301 			/*
302 			 * If lowervp was unlocked during VOP
303 			 * operation, nullfs upper vnode could have
304 			 * been reclaimed, which changes its v_vnlock
305 			 * back to private v_lock.  In this case we
306 			 * must move lock ownership from lower to
307 			 * upper (reclaimed) vnode.
308 			 */
309 			if (lvp != NULLVP &&
310 			    VOP_ISLOCKED(lvp) == LK_EXCLUSIVE &&
311 			    old_vps[i]->v_vnlock != lvp->v_vnlock) {
312 				VOP_UNLOCK(lvp);
313 				VOP_LOCK(old_vps[i], LK_EXCLUSIVE | LK_RETRY);
314 			}
315 
316 			*(vps_p[i]) = old_vps[i];
317 #if 0
318 			if (reles & VDESC_VP0_WILLUNLOCK)
319 				VOP_UNLOCK(*(vps_p[i]), 0);
320 #endif
321 			if (reles & VDESC_VP0_WILLRELE)
322 				vrele(*(vps_p[i]));
323 		}
324 	}
325 
326 	/*
327 	 * Map the possible out-going vpp
328 	 * (Assumes that the lower layer always returns
329 	 * a VREF'ed vpp unless it gets an error.)
330 	 */
331 	if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET && !error) {
332 		/*
333 		 * XXX - even though some ops have vpp returned vp's,
334 		 * several ops actually vrele this before returning.
335 		 * We must avoid these ops.
336 		 * (This should go away when these ops are regularized.)
337 		 */
338 		vppp = VOPARG_OFFSETTO(struct vnode***,
339 				 descp->vdesc_vpp_offset,ap);
340 		if (*vppp)
341 			error = null_nodeget(old_vps[0]->v_mount, **vppp, *vppp);
342 	}
343 
344 	return (error);
345 }
346 
347 static int
348 null_add_writecount(struct vop_add_writecount_args *ap)
349 {
350 	struct vnode *lvp, *vp;
351 	int error;
352 
353 	vp = ap->a_vp;
354 	lvp = NULLVPTOLOWERVP(vp);
355 	VI_LOCK(vp);
356 	/* text refs are bypassed to lowervp */
357 	VNASSERT(vp->v_writecount >= 0, vp, ("wrong null writecount"));
358 	VNASSERT(vp->v_writecount + ap->a_inc >= 0, vp,
359 	    ("wrong writecount inc %d", ap->a_inc));
360 	error = VOP_ADD_WRITECOUNT(lvp, ap->a_inc);
361 	if (error == 0)
362 		vp->v_writecount += ap->a_inc;
363 	VI_UNLOCK(vp);
364 	return (error);
365 }
366 
367 /*
368  * We have to carry on the locking protocol on the null layer vnodes
369  * as we progress through the tree. We also have to enforce read-only
370  * if this layer is mounted read-only.
371  */
372 static int
373 null_lookup(struct vop_lookup_args *ap)
374 {
375 	struct componentname *cnp = ap->a_cnp;
376 	struct vnode *dvp = ap->a_dvp;
377 	int flags = cnp->cn_flags;
378 	struct vnode *vp, *ldvp, *lvp;
379 	struct mount *mp;
380 	int error;
381 
382 	mp = dvp->v_mount;
383 	if ((flags & ISLASTCN) != 0 && (mp->mnt_flag & MNT_RDONLY) != 0 &&
384 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
385 		return (EROFS);
386 	/*
387 	 * Although it is possible to call null_bypass(), we'll do
388 	 * a direct call to reduce overhead
389 	 */
390 	ldvp = NULLVPTOLOWERVP(dvp);
391 	vp = lvp = NULL;
392 	KASSERT((ldvp->v_vflag & VV_ROOT) == 0 ||
393 	    ((dvp->v_vflag & VV_ROOT) != 0 && (flags & ISDOTDOT) == 0),
394 	    ("ldvp %p fl %#x dvp %p fl %#x flags %#x", ldvp, ldvp->v_vflag,
395 	     dvp, dvp->v_vflag, flags));
396 
397 	/*
398 	 * Hold ldvp.  The reference on it, owned by dvp, is lost in
399 	 * case of dvp reclamation, and we need ldvp to move our lock
400 	 * from ldvp to dvp.
401 	 */
402 	vhold(ldvp);
403 
404 	error = VOP_LOOKUP(ldvp, &lvp, cnp);
405 
406 	/*
407 	 * VOP_LOOKUP() on lower vnode may unlock ldvp, which allows
408 	 * dvp to be reclaimed due to shared v_vnlock.  Check for the
409 	 * doomed state and return error.
410 	 */
411 	if ((error == 0 || error == EJUSTRETURN) &&
412 	    VN_IS_DOOMED(dvp)) {
413 		error = ENOENT;
414 		if (lvp != NULL)
415 			vput(lvp);
416 
417 		/*
418 		 * If vgone() did reclaimed dvp before curthread
419 		 * relocked ldvp, the locks of dvp and ldpv are no
420 		 * longer shared.  In this case, relock of ldvp in
421 		 * lower fs VOP_LOOKUP() does not restore the locking
422 		 * state of dvp.  Compensate for this by unlocking
423 		 * ldvp and locking dvp, which is also correct if the
424 		 * locks are still shared.
425 		 */
426 		VOP_UNLOCK(ldvp);
427 		vn_lock(dvp, LK_EXCLUSIVE | LK_RETRY);
428 	}
429 	vdrop(ldvp);
430 
431 	if (error == EJUSTRETURN && (flags & ISLASTCN) != 0 &&
432 	    (mp->mnt_flag & MNT_RDONLY) != 0 &&
433 	    (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME))
434 		error = EROFS;
435 
436 	if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) {
437 		if (ldvp == lvp) {
438 			*ap->a_vpp = dvp;
439 			VREF(dvp);
440 			vrele(lvp);
441 		} else {
442 			error = null_nodeget(mp, lvp, &vp);
443 			if (error == 0)
444 				*ap->a_vpp = vp;
445 		}
446 	}
447 	return (error);
448 }
449 
450 static int
451 null_open(struct vop_open_args *ap)
452 {
453 	int retval;
454 	struct vnode *vp, *ldvp;
455 
456 	vp = ap->a_vp;
457 	ldvp = NULLVPTOLOWERVP(vp);
458 	retval = null_bypass(&ap->a_gen);
459 	if (retval == 0) {
460 		vp->v_object = ldvp->v_object;
461 		if ((vn_irflag_read(ldvp) & VIRF_PGREAD) != 0) {
462 			MPASS(vp->v_object != NULL);
463 			if ((vn_irflag_read(vp) & VIRF_PGREAD) == 0) {
464 				vn_irflag_set_cond(vp, VIRF_PGREAD);
465 			}
466 		}
467 	}
468 	return (retval);
469 }
470 
471 /*
472  * Setattr call. Disallow write attempts if the layer is mounted read-only.
473  */
474 static int
475 null_setattr(struct vop_setattr_args *ap)
476 {
477 	struct vnode *vp = ap->a_vp;
478 	struct vattr *vap = ap->a_vap;
479 
480   	if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL ||
481 	    vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL ||
482 	    vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) &&
483 	    (vp->v_mount->mnt_flag & MNT_RDONLY))
484 		return (EROFS);
485 	if (vap->va_size != VNOVAL) {
486  		switch (vp->v_type) {
487  		case VDIR:
488  			return (EISDIR);
489  		case VCHR:
490  		case VBLK:
491  		case VSOCK:
492  		case VFIFO:
493 			if (vap->va_flags != VNOVAL)
494 				return (EOPNOTSUPP);
495 			return (0);
496 		case VREG:
497 		case VLNK:
498  		default:
499 			/*
500 			 * Disallow write attempts if the filesystem is
501 			 * mounted read-only.
502 			 */
503 			if (vp->v_mount->mnt_flag & MNT_RDONLY)
504 				return (EROFS);
505 		}
506 	}
507 
508 	return (null_bypass((struct vop_generic_args *)ap));
509 }
510 
511 /*
512  *  We handle stat and getattr only to change the fsid.
513  */
514 static int
515 null_stat(struct vop_stat_args *ap)
516 {
517 	int error;
518 
519 	if ((error = null_bypass((struct vop_generic_args *)ap)) != 0)
520 		return (error);
521 
522 	ap->a_sb->st_dev = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
523 	return (0);
524 }
525 
526 static int
527 null_getattr(struct vop_getattr_args *ap)
528 {
529 	int error;
530 
531 	if ((error = null_bypass((struct vop_generic_args *)ap)) != 0)
532 		return (error);
533 
534 	ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
535 	return (0);
536 }
537 
538 /*
539  * Handle to disallow write access if mounted read-only.
540  */
541 static int
542 null_access(struct vop_access_args *ap)
543 {
544 	struct vnode *vp = ap->a_vp;
545 	accmode_t accmode = ap->a_accmode;
546 
547 	/*
548 	 * Disallow write attempts on read-only layers;
549 	 * unless the file is a socket, fifo, or a block or
550 	 * character device resident on the filesystem.
551 	 */
552 	if (accmode & VWRITE) {
553 		switch (vp->v_type) {
554 		case VDIR:
555 		case VLNK:
556 		case VREG:
557 			if (vp->v_mount->mnt_flag & MNT_RDONLY)
558 				return (EROFS);
559 			break;
560 		default:
561 			break;
562 		}
563 	}
564 	return (null_bypass((struct vop_generic_args *)ap));
565 }
566 
567 static int
568 null_accessx(struct vop_accessx_args *ap)
569 {
570 	struct vnode *vp = ap->a_vp;
571 	accmode_t accmode = ap->a_accmode;
572 
573 	/*
574 	 * Disallow write attempts on read-only layers;
575 	 * unless the file is a socket, fifo, or a block or
576 	 * character device resident on the filesystem.
577 	 */
578 	if (accmode & VWRITE) {
579 		switch (vp->v_type) {
580 		case VDIR:
581 		case VLNK:
582 		case VREG:
583 			if (vp->v_mount->mnt_flag & MNT_RDONLY)
584 				return (EROFS);
585 			break;
586 		default:
587 			break;
588 		}
589 	}
590 	return (null_bypass((struct vop_generic_args *)ap));
591 }
592 
593 /*
594  * Increasing refcount of lower vnode is needed at least for the case
595  * when lower FS is NFS to do sillyrename if the file is in use.
596  * Unfortunately v_usecount is incremented in many places in
597  * the kernel and, as such, there may be races that result in
598  * the NFS client doing an extraneous silly rename, but that seems
599  * preferable to not doing a silly rename when it is needed.
600  */
601 static int
602 null_remove(struct vop_remove_args *ap)
603 {
604 	int retval, vreleit;
605 	struct vnode *lvp, *vp;
606 
607 	vp = ap->a_vp;
608 	if (vrefcnt(vp) > 1) {
609 		lvp = NULLVPTOLOWERVP(vp);
610 		VREF(lvp);
611 		vreleit = 1;
612 	} else
613 		vreleit = 0;
614 	VTONULL(vp)->null_flags |= NULLV_DROP;
615 	retval = null_bypass(&ap->a_gen);
616 	if (vreleit != 0)
617 		vrele(lvp);
618 	return (retval);
619 }
620 
621 /*
622  * We handle this to eliminate null FS to lower FS
623  * file moving. Don't know why we don't allow this,
624  * possibly we should.
625  */
626 static int
627 null_rename(struct vop_rename_args *ap)
628 {
629 	struct vnode *tdvp = ap->a_tdvp;
630 	struct vnode *fvp = ap->a_fvp;
631 	struct vnode *fdvp = ap->a_fdvp;
632 	struct vnode *tvp = ap->a_tvp;
633 	struct null_node *tnn;
634 
635 	/* Check for cross-device rename. */
636 	if ((fvp->v_mount != tdvp->v_mount) ||
637 	    (tvp && (fvp->v_mount != tvp->v_mount))) {
638 		if (tdvp == tvp)
639 			vrele(tdvp);
640 		else
641 			vput(tdvp);
642 		if (tvp)
643 			vput(tvp);
644 		vrele(fdvp);
645 		vrele(fvp);
646 		return (EXDEV);
647 	}
648 
649 	if (tvp != NULL) {
650 		tnn = VTONULL(tvp);
651 		tnn->null_flags |= NULLV_DROP;
652 	}
653 	return (null_bypass((struct vop_generic_args *)ap));
654 }
655 
656 static int
657 null_rmdir(struct vop_rmdir_args *ap)
658 {
659 
660 	VTONULL(ap->a_vp)->null_flags |= NULLV_DROP;
661 	return (null_bypass(&ap->a_gen));
662 }
663 
664 /*
665  * We need to process our own vnode lock and then clear the
666  * interlock flag as it applies only to our vnode, not the
667  * vnodes below us on the stack.
668  */
669 static int
670 null_lock(struct vop_lock1_args *ap)
671 {
672 	struct vnode *vp = ap->a_vp;
673 	int flags;
674 	struct null_node *nn;
675 	struct vnode *lvp;
676 	int error;
677 
678 	if ((ap->a_flags & LK_INTERLOCK) == 0)
679 		VI_LOCK(vp);
680 	else
681 		ap->a_flags &= ~LK_INTERLOCK;
682 	flags = ap->a_flags;
683 	nn = VTONULL(vp);
684 	/*
685 	 * If we're still active we must ask the lower layer to
686 	 * lock as ffs has special lock considerations in its
687 	 * vop lock.
688 	 */
689 	if (nn != NULL && (lvp = NULLVPTOLOWERVP(vp)) != NULL) {
690 		/*
691 		 * We have to hold the vnode here to solve a potential
692 		 * reclaim race.  If we're forcibly vgone'd while we
693 		 * still have refs, a thread could be sleeping inside
694 		 * the lowervp's vop_lock routine.  When we vgone we will
695 		 * drop our last ref to the lowervp, which would allow it
696 		 * to be reclaimed.  The lowervp could then be recycled,
697 		 * in which case it is not legal to be sleeping in its VOP.
698 		 * We prevent it from being recycled by holding the vnode
699 		 * here.
700 		 */
701 		vholdnz(lvp);
702 		VI_UNLOCK(vp);
703 		error = VOP_LOCK(lvp, flags);
704 
705 		/*
706 		 * We might have slept to get the lock and someone might have
707 		 * clean our vnode already, switching vnode lock from one in
708 		 * lowervp to v_lock in our own vnode structure.  Handle this
709 		 * case by reacquiring correct lock in requested mode.
710 		 */
711 		if (VTONULL(vp) == NULL && error == 0) {
712 			ap->a_flags &= ~LK_TYPE_MASK;
713 			switch (flags & LK_TYPE_MASK) {
714 			case LK_SHARED:
715 				ap->a_flags |= LK_SHARED;
716 				break;
717 			case LK_UPGRADE:
718 			case LK_EXCLUSIVE:
719 				ap->a_flags |= LK_EXCLUSIVE;
720 				break;
721 			default:
722 				panic("Unsupported lock request %d\n",
723 				    ap->a_flags);
724 			}
725 			VOP_UNLOCK(lvp);
726 			error = vop_stdlock(ap);
727 		}
728 		vdrop(lvp);
729 	} else {
730 		VI_UNLOCK(vp);
731 		error = vop_stdlock(ap);
732 	}
733 
734 	return (error);
735 }
736 
737 /*
738  * We need to process our own vnode unlock and then clear the
739  * interlock flag as it applies only to our vnode, not the
740  * vnodes below us on the stack.
741  */
742 static int
743 null_unlock(struct vop_unlock_args *ap)
744 {
745 	struct vnode *vp = ap->a_vp;
746 	struct null_node *nn;
747 	struct vnode *lvp;
748 	int error;
749 
750 	nn = VTONULL(vp);
751 	if (nn != NULL && (lvp = NULLVPTOLOWERVP(vp)) != NULL) {
752 		vholdnz(lvp);
753 		error = VOP_UNLOCK(lvp);
754 		vdrop(lvp);
755 	} else {
756 		error = vop_stdunlock(ap);
757 	}
758 
759 	return (error);
760 }
761 
762 /*
763  * Do not allow the VOP_INACTIVE to be passed to the lower layer,
764  * since the reference count on the lower vnode is not related to
765  * ours.
766  */
767 static int
768 null_want_recycle(struct vnode *vp)
769 {
770 	struct vnode *lvp;
771 	struct null_node *xp;
772 	struct mount *mp;
773 	struct null_mount *xmp;
774 
775 	xp = VTONULL(vp);
776 	lvp = NULLVPTOLOWERVP(vp);
777 	mp = vp->v_mount;
778 	xmp = MOUNTTONULLMOUNT(mp);
779 	if ((xmp->nullm_flags & NULLM_CACHE) == 0 ||
780 	    (xp->null_flags & NULLV_DROP) != 0 ||
781 	    (lvp->v_vflag & VV_NOSYNC) != 0) {
782 		/*
783 		 * If this is the last reference and caching of the
784 		 * nullfs vnodes is not enabled, or the lower vnode is
785 		 * deleted, then free up the vnode so as not to tie up
786 		 * the lower vnodes.
787 		 */
788 		return (1);
789 	}
790 	return (0);
791 }
792 
793 static int
794 null_inactive(struct vop_inactive_args *ap)
795 {
796 	struct vnode *vp;
797 
798 	vp = ap->a_vp;
799 	if (null_want_recycle(vp)) {
800 		vp->v_object = NULL;
801 		vrecycle(vp);
802 	}
803 	return (0);
804 }
805 
806 static int
807 null_need_inactive(struct vop_need_inactive_args *ap)
808 {
809 
810 	return (null_want_recycle(ap->a_vp));
811 }
812 
813 /*
814  * Now, the nullfs vnode and, due to the sharing lock, the lower
815  * vnode, are exclusively locked, and we shall destroy the null vnode.
816  */
817 static int
818 null_reclaim(struct vop_reclaim_args *ap)
819 {
820 	struct vnode *vp;
821 	struct null_node *xp;
822 	struct vnode *lowervp;
823 
824 	vp = ap->a_vp;
825 	xp = VTONULL(vp);
826 	lowervp = xp->null_lowervp;
827 
828 	KASSERT(lowervp != NULL && vp->v_vnlock != &vp->v_lock,
829 	    ("Reclaiming incomplete null vnode %p", vp));
830 
831 	null_hashrem(xp);
832 	/*
833 	 * Use the interlock to protect the clearing of v_data to
834 	 * prevent faults in null_lock().
835 	 */
836 	lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
837 	VI_LOCK(vp);
838 	vp->v_data = NULL;
839 	vp->v_object = NULL;
840 	vp->v_vnlock = &vp->v_lock;
841 
842 	/*
843 	 * If we were opened for write, we leased the write reference
844 	 * to the lower vnode.  If this is a reclamation due to the
845 	 * forced unmount, undo the reference now.
846 	 */
847 	if (vp->v_writecount > 0)
848 		VOP_ADD_WRITECOUNT(lowervp, -vp->v_writecount);
849 	else if (vp->v_writecount < 0)
850 		vp->v_writecount = 0;
851 
852 	VI_UNLOCK(vp);
853 
854 	if ((xp->null_flags & NULLV_NOUNLOCK) != 0)
855 		vunref(lowervp);
856 	else
857 		vput(lowervp);
858 	free(xp, M_NULLFSNODE);
859 
860 	return (0);
861 }
862 
863 static int
864 null_print(struct vop_print_args *ap)
865 {
866 	struct vnode *vp = ap->a_vp;
867 
868 	printf("\tvp=%p, lowervp=%p\n", vp, VTONULL(vp)->null_lowervp);
869 	return (0);
870 }
871 
872 /* ARGSUSED */
873 static int
874 null_getwritemount(struct vop_getwritemount_args *ap)
875 {
876 	struct null_node *xp;
877 	struct vnode *lowervp;
878 	struct vnode *vp;
879 
880 	vp = ap->a_vp;
881 	VI_LOCK(vp);
882 	xp = VTONULL(vp);
883 	if (xp && (lowervp = xp->null_lowervp)) {
884 		vholdnz(lowervp);
885 		VI_UNLOCK(vp);
886 		VOP_GETWRITEMOUNT(lowervp, ap->a_mpp);
887 		vdrop(lowervp);
888 	} else {
889 		VI_UNLOCK(vp);
890 		*(ap->a_mpp) = NULL;
891 	}
892 	return (0);
893 }
894 
895 static int
896 null_vptofh(struct vop_vptofh_args *ap)
897 {
898 	struct vnode *lvp;
899 
900 	lvp = NULLVPTOLOWERVP(ap->a_vp);
901 	return VOP_VPTOFH(lvp, ap->a_fhp);
902 }
903 
904 static int
905 null_vptocnp(struct vop_vptocnp_args *ap)
906 {
907 	struct vnode *vp = ap->a_vp;
908 	struct vnode **dvp = ap->a_vpp;
909 	struct vnode *lvp, *ldvp;
910 	struct mount *mp;
911 	int error, locked;
912 
913 	locked = VOP_ISLOCKED(vp);
914 	lvp = NULLVPTOLOWERVP(vp);
915 	vhold(lvp);
916 	mp = vp->v_mount;
917 	vfs_ref(mp);
918 	VOP_UNLOCK(vp); /* vp is held by vn_vptocnp_locked that called us */
919 	ldvp = lvp;
920 	vref(lvp);
921 	error = vn_vptocnp(&ldvp, ap->a_buf, ap->a_buflen);
922 	vdrop(lvp);
923 	if (error != 0) {
924 		vn_lock(vp, locked | LK_RETRY);
925 		vfs_rel(mp);
926 		return (ENOENT);
927 	}
928 
929 	error = vn_lock(ldvp, LK_SHARED);
930 	if (error != 0) {
931 		vrele(ldvp);
932 		vn_lock(vp, locked | LK_RETRY);
933 		vfs_rel(mp);
934 		return (ENOENT);
935 	}
936 	error = null_nodeget(mp, ldvp, dvp);
937 	if (error == 0) {
938 #ifdef DIAGNOSTIC
939 		NULLVPTOLOWERVP(*dvp);
940 #endif
941 		VOP_UNLOCK(*dvp); /* keep reference on *dvp */
942 	}
943 	vn_lock(vp, locked | LK_RETRY);
944 	vfs_rel(mp);
945 	return (error);
946 }
947 
948 static int
949 null_read_pgcache(struct vop_read_pgcache_args *ap)
950 {
951 	struct vnode *lvp, *vp;
952 	struct null_node *xp;
953 	int error;
954 
955 	vp = ap->a_vp;
956 	VI_LOCK(vp);
957 	xp = VTONULL(vp);
958 	if (xp == NULL) {
959 		VI_UNLOCK(vp);
960 		return (EJUSTRETURN);
961 	}
962 	lvp = xp->null_lowervp;
963 	vref(lvp);
964 	VI_UNLOCK(vp);
965 	error = VOP_READ_PGCACHE(lvp, ap->a_uio, ap->a_ioflag, ap->a_cred);
966 	vrele(lvp);
967 	return (error);
968 }
969 
970 /*
971  * Global vfs data structures
972  */
973 struct vop_vector null_vnodeops = {
974 	.vop_bypass =		null_bypass,
975 	.vop_access =		null_access,
976 	.vop_accessx =		null_accessx,
977 	.vop_advlockpurge =	vop_stdadvlockpurge,
978 	.vop_bmap =		VOP_EOPNOTSUPP,
979 	.vop_stat =		null_stat,
980 	.vop_getattr =		null_getattr,
981 	.vop_getwritemount =	null_getwritemount,
982 	.vop_inactive =		null_inactive,
983 	.vop_need_inactive =	null_need_inactive,
984 	.vop_islocked =		vop_stdislocked,
985 	.vop_lock1 =		null_lock,
986 	.vop_lookup =		null_lookup,
987 	.vop_open =		null_open,
988 	.vop_print =		null_print,
989 	.vop_read_pgcache =	null_read_pgcache,
990 	.vop_reclaim =		null_reclaim,
991 	.vop_remove =		null_remove,
992 	.vop_rename =		null_rename,
993 	.vop_rmdir =		null_rmdir,
994 	.vop_setattr =		null_setattr,
995 	.vop_strategy =		VOP_EOPNOTSUPP,
996 	.vop_unlock =		null_unlock,
997 	.vop_vptocnp =		null_vptocnp,
998 	.vop_vptofh =		null_vptofh,
999 	.vop_add_writecount =	null_add_writecount,
1000 };
1001 VFS_VOP_VECTOR_REGISTER(null_vnodeops);
1002