1/*-
2 * Copyright (c) 1992, 1993
3 * The Regents of the University of California. All rights reserved.
4 *
5 * This code is derived from software contributed to Berkeley by
6 * John Heidemann of the UCLA Ficus project.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 *
32 * @(#)null_vnops.c 8.6 (Berkeley) 5/27/95
33 *
34 * Ancestors:
35 * @(#)lofs_vnops.c 1.2 (Berkeley) 6/18/92
36 * ...and...
37 * @(#)null_vnodeops.c 1.20 92/07/07 UCLA Ficus project
38 *
39 * $FreeBSD: head/sys/fs/nullfs/null_vnops.c 140728 2005-01-24 11:49:41Z phk $
40 */
41
42/*
43 * Null Layer
44 *
45 * (See mount_nullfs(8) for more information.)
46 *
47 * The null layer duplicates a portion of the filesystem
48 * name space under a new name. In this respect, it is
49 * similar to the loopback filesystem. It differs from
50 * the loopback fs in two respects: it is implemented using
51 * a stackable layers techniques, and its "null-node"s stack above
52 * all lower-layer vnodes, not just over directory vnodes.
53 *
54 * The null layer has two purposes. First, it serves as a demonstration
55 * of layering by proving a layer which does nothing. (It actually
56 * does everything the loopback filesystem does, which is slightly
57 * more than nothing.) Second, the null layer can serve as a prototype
58 * layer. Since it provides all necessary layer framework,
59 * new filesystem layers can be created very easily be starting
60 * with a null layer.
61 *
62 * The remainder of this man page examines the null layer as a basis
63 * for constructing new layers.
64 *
65 *
66 * INSTANTIATING NEW NULL LAYERS
67 *
68 * New null layers are created with mount_nullfs(8).
69 * Mount_nullfs(8) takes two arguments, the pathname
70 * of the lower vfs (target-pn) and the pathname where the null
71 * layer will appear in the namespace (alias-pn). After
72 * the null layer is put into place, the contents
73 * of target-pn subtree will be aliased under alias-pn.
74 *
75 *
76 * OPERATION OF A NULL LAYER
77 *
78 * The null layer is the minimum filesystem layer,
79 * simply bypassing all possible operations to the lower layer
80 * for processing there. The majority of its activity centers
81 * on the bypass routine, through which nearly all vnode operations
82 * pass.
83 *
84 * The bypass routine accepts arbitrary vnode operations for
85 * handling by the lower layer. It begins by examing vnode
86 * operation arguments and replacing any null-nodes by their
87 * lower-layer equivlants. It then invokes the operation
88 * on the lower layer. Finally, it replaces the null-nodes
89 * in the arguments and, if a vnode is return by the operation,
90 * stacks a null-node on top of the returned vnode.
91 *
92 * Although bypass handles most operations, vop_getattr, vop_lock,
93 * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not
94 * bypassed. Vop_getattr must change the fsid being returned.
95 * Vop_lock and vop_unlock must handle any locking for the
96 * current vnode as well as pass the lock request down.
97 * Vop_inactive and vop_reclaim are not bypassed so that
98 * they can handle freeing null-layer specific data. Vop_print
99 * is not bypassed to avoid excessive debugging information.
100 * Also, certain vnode operations change the locking state within
101 * the operation (create, mknod, remove, link, rename, mkdir, rmdir,
102 * and symlink). Ideally these operations should not change the
103 * lock state, but should be changed to let the caller of the
104 * function unlock them. Otherwise all intermediate vnode layers
105 * (such as union, umapfs, etc) must catch these functions to do
106 * the necessary locking at their layer.
107 *
108 *
109 * INSTANTIATING VNODE STACKS
110 *
111 * Mounting associates the null layer with a lower layer,
112 * effect stacking two VFSes. Vnode stacks are instead
113 * created on demand as files are accessed.
114 *
115 * The initial mount creates a single vnode stack for the
116 * root of the new null layer. All other vnode stacks
117 * are created as a result of vnode operations on
118 * this or other null vnode stacks.
119 *
120 * New vnode stacks come into existance as a result of
121 * an operation which returns a vnode.
122 * The bypass routine stacks a null-node above the new
123 * vnode before returning it to the caller.
124 *
125 * For example, imagine mounting a null layer with
126 * "mount_nullfs /usr/include /dev/layer/null".
127 * Changing directory to /dev/layer/null will assign
128 * the root null-node (which was created when the null layer was mounted).
129 * Now consider opening "sys". A vop_lookup would be
130 * done on the root null-node. This operation would bypass through
131 * to the lower layer which would return a vnode representing
132 * the UFS "sys". Null_bypass then builds a null-node
133 * aliasing the UFS "sys" and returns this to the caller.
134 * Later operations on the null-node "sys" will repeat this
135 * process when constructing other vnode stacks.
136 *
137 *
138 * CREATING OTHER FILE SYSTEM LAYERS
139 *
140 * One of the easiest ways to construct new filesystem layers is to make
141 * a copy of the null layer, rename all files and variables, and
142 * then begin modifing the copy. Sed can be used to easily rename
143 * all variables.
144 *
145 * The umap layer is an example of a layer descended from the
146 * null layer.
147 *
148 *
149 * INVOKING OPERATIONS ON LOWER LAYERS
150 *
151 * There are two techniques to invoke operations on a lower layer
152 * when the operation cannot be completely bypassed. Each method
153 * is appropriate in different situations. In both cases,
154 * it is the responsibility of the aliasing layer to make
155 * the operation arguments "correct" for the lower layer
156 * by mapping a vnode arguments to the lower layer.
157 *
158 * The first approach is to call the aliasing layer's bypass routine.
159 * This method is most suitable when you wish to invoke the operation
160 * currently being handled on the lower layer. It has the advantage
161 * that the bypass routine already must do argument mapping.
162 * An example of this is null_getattrs in the null layer.
163 *
164 * A second approach is to directly invoke vnode operations on
165 * the lower layer with the VOP_OPERATIONNAME interface.
166 * The advantage of this method is that it is easy to invoke
167 * arbitrary operations on the lower layer. The disadvantage
168 * is that vnode arguments must be manualy mapped.
169 *
170 */
171
172#include <sys/param.h>
173#include <sys/systm.h>
174#include <sys/conf.h>
175#include <sys/kernel.h>
176#include <sys/lock.h>
177#include <sys/malloc.h>
178#include <sys/mount.h>
179#include <sys/mutex.h>
180#include <sys/namei.h>
181#include <sys/sysctl.h>
182#include <sys/vnode.h>
183
184#include <fs/nullfs/null.h>
185
186#include <vm/vm.h>
187#include <vm/vm_extern.h>
188#include <vm/vm_object.h>
189#include <vm/vnode_pager.h>
190
191static int null_bug_bypass = 0; /* for debugging: enables bypass printf'ing */
192SYSCTL_INT(_debug, OID_AUTO, nullfs_bug_bypass, CTLFLAG_RW,
193 &null_bug_bypass, 0, "");
194
195/*
196 * This is the 10-Apr-92 bypass routine.
197 * This version has been optimized for speed, throwing away some
198 * safety checks. It should still always work, but it's not as
199 * robust to programmer errors.
200 *
201 * In general, we map all vnodes going down and unmap them on the way back.
202 * As an exception to this, vnodes can be marked "unmapped" by setting
203 * the Nth bit in operation's vdesc_flags.
204 *
205 * Also, some BSD vnode operations have the side effect of vrele'ing
206 * their arguments. With stacking, the reference counts are held
207 * by the upper node, not the lower one, so we must handle these
208 * side-effects here. This is not of concern in Sun-derived systems
209 * since there are no such side-effects.
210 *
211 * This makes the following assumptions:
212 * - only one returned vpp
213 * - no INOUT vpp's (Sun's vop_open has one of these)
214 * - the vnode operation vector of the first vnode should be used
215 * to determine what implementation of the op should be invoked
216 * - all mapped vnodes are of our vnode-type (NEEDSWORK:
217 * problems on rmdir'ing mount points and renaming?)
218 */
219int
220null_bypass(struct vop_generic_args *ap)
221{
222 register struct vnode **this_vp_p;
223 int error;
224 struct vnode *old_vps[VDESC_MAX_VPS];
225 struct vnode **vps_p[VDESC_MAX_VPS];
226 struct vnode ***vppp;
227 struct vnodeop_desc *descp = ap->a_desc;
228 int reles, i;
229
230 if (null_bug_bypass)
231 printf ("null_bypass: %s\n", descp->vdesc_name);
232
233#ifdef DIAGNOSTIC
234 /*
235 * We require at least one vp.
236 */
237 if (descp->vdesc_vp_offsets == NULL ||
238 descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET)
239 panic ("null_bypass: no vp's in map");
240#endif
241
242 /*
243 * Map the vnodes going in.
244 * Later, we'll invoke the operation based on
245 * the first mapped vnode's operation vector.
246 */
247 reles = descp->vdesc_flags;
248 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
249 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
250 break; /* bail out at end of list */
251 vps_p[i] = this_vp_p =
252 VOPARG_OFFSETTO(struct vnode**,descp->vdesc_vp_offsets[i],ap);
253 /*
254 * We're not guaranteed that any but the first vnode
255 * are of our type. Check for and don't map any
256 * that aren't. (We must always map first vp or vclean fails.)
257 */
258 if (i && (*this_vp_p == NULLVP ||
259 (*this_vp_p)->v_op != &null_vnodeops)) {
260 old_vps[i] = NULLVP;
261 } else {
262 old_vps[i] = *this_vp_p;
263 *(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p);
264 /*
265 * XXX - Several operations have the side effect
266 * of vrele'ing their vp's. We must account for
267 * that. (This should go away in the future.)
268 */
269 if (reles & VDESC_VP0_WILLRELE)
270 VREF(*this_vp_p);
271 }
272
273 }
274
275 /*
276 * Call the operation on the lower layer
277 * with the modified argument structure.
278 */
279 if (vps_p[0] && *vps_p[0])
280 error = VCALL(ap);
281 else {
282 printf("null_bypass: no map for %s\n", descp->vdesc_name);
283 error = EINVAL;
284 }
285
286 /*
287 * Maintain the illusion of call-by-value
288 * by restoring vnodes in the argument structure
289 * to their original value.
290 */
291 reles = descp->vdesc_flags;
292 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
293 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
294 break; /* bail out at end of list */
295 if (old_vps[i]) {
296 *(vps_p[i]) = old_vps[i];
297#if 0
298 if (reles & VDESC_VP0_WILLUNLOCK)
299 VOP_UNLOCK(*(vps_p[i]), LK_THISLAYER, curthread);
300#endif
301 if (reles & VDESC_VP0_WILLRELE)
302 vrele(*(vps_p[i]));
303 }
304 }
305
306 /*
307 * Map the possible out-going vpp
308 * (Assumes that the lower layer always returns
309 * a VREF'ed vpp unless it gets an error.)
310 */
311 if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET &&
312 !(descp->vdesc_flags & VDESC_NOMAP_VPP) &&
313 !error) {
314 /*
315 * XXX - even though some ops have vpp returned vp's,
316 * several ops actually vrele this before returning.
317 * We must avoid these ops.
318 * (This should go away when these ops are regularized.)
319 */
320 if (descp->vdesc_flags & VDESC_VPP_WILLRELE)
321 goto out;
322 vppp = VOPARG_OFFSETTO(struct vnode***,
323 descp->vdesc_vpp_offset,ap);
324 if (*vppp)
325 error = null_nodeget(old_vps[0]->v_mount, **vppp, *vppp);
326 }
327
328 out:
329 return (error);
330}
331
332/*
333 * We have to carry on the locking protocol on the null layer vnodes
334 * as we progress through the tree. We also have to enforce read-only
335 * if this layer is mounted read-only.
336 */
337static int
338null_lookup(struct vop_lookup_args *ap)
339{
340 struct componentname *cnp = ap->a_cnp;
341 struct vnode *dvp = ap->a_dvp;
342 struct thread *td = cnp->cn_thread;
343 int flags = cnp->cn_flags;
344 struct vnode *vp, *ldvp, *lvp;
345 int error;
346
347 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
348 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
349 return (EROFS);
350 /*
351 * Although it is possible to call null_bypass(), we'll do
352 * a direct call to reduce overhead
353 */
354 ldvp = NULLVPTOLOWERVP(dvp);
355 vp = lvp = NULL;
356 error = VOP_LOOKUP(ldvp, &lvp, cnp);
357 if (error == EJUSTRETURN && (flags & ISLASTCN) &&
358 (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
359 (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME))
360 error = EROFS;
361
362 /*
363 * Rely only on the PDIRUNLOCK flag which should be carefully
364 * tracked by underlying filesystem.
365 */
366 if ((cnp->cn_flags & PDIRUNLOCK) && dvp->v_vnlock != ldvp->v_vnlock)
367 VOP_UNLOCK(dvp, LK_THISLAYER, td);
368 if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) {
369 if (ldvp == lvp) {
370 *ap->a_vpp = dvp;
371 VREF(dvp);
372 vrele(lvp);
373 } else {
374 error = null_nodeget(dvp->v_mount, lvp, &vp);
375 if (error) {
376 /* XXX Cleanup needed... */
377 panic("null_nodeget failed");
378 }
379 *ap->a_vpp = vp;
380 }
381 }
382 return (error);
383}
384
385/*
386 * Setattr call. Disallow write attempts if the layer is mounted read-only.
387 */
388static int
389null_setattr(struct vop_setattr_args *ap)
390{
391 struct vnode *vp = ap->a_vp;
392 struct vattr *vap = ap->a_vap;
393
394 if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL ||
395 vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL ||
396 vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) &&
397 (vp->v_mount->mnt_flag & MNT_RDONLY))
398 return (EROFS);
399 if (vap->va_size != VNOVAL) {
400 switch (vp->v_type) {
401 case VDIR:
402 return (EISDIR);
403 case VCHR:
404 case VBLK:
405 case VSOCK:
406 case VFIFO:
407 if (vap->va_flags != VNOVAL)
408 return (EOPNOTSUPP);
409 return (0);
410 case VREG:
411 case VLNK:
412 default:
413 /*
414 * Disallow write attempts if the filesystem is
415 * mounted read-only.
416 */
417 if (vp->v_mount->mnt_flag & MNT_RDONLY)
418 return (EROFS);
419 }
420 }
421
422 return (null_bypass((struct vop_generic_args *)ap));
423}
424
425/*
426 * We handle getattr only to change the fsid.
427 */
428static int
429null_getattr(struct vop_getattr_args *ap)
430{
431 int error;
432
433 if ((error = null_bypass((struct vop_generic_args *)ap)) != 0)
434 return (error);
435
436 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
437 return (0);
438}
439
440/*
441 * Handle to disallow write access if mounted read-only.
442 */
443static int
444null_access(struct vop_access_args *ap)
445{
446 struct vnode *vp = ap->a_vp;
447 mode_t mode = ap->a_mode;
448
449 /*
450 * Disallow write attempts on read-only layers;
451 * unless the file is a socket, fifo, or a block or
452 * character device resident on the filesystem.
453 */
454 if (mode & VWRITE) {
455 switch (vp->v_type) {
456 case VDIR:
457 case VLNK:
458 case VREG:
459 if (vp->v_mount->mnt_flag & MNT_RDONLY)
460 return (EROFS);
461 break;
462 default:
463 break;
464 }
465 }
466 return (null_bypass((struct vop_generic_args *)ap));
467}
468
469/*
470 * We handle this to eliminate null FS to lower FS
471 * file moving. Don't know why we don't allow this,
472 * possibly we should.
473 */
474static int
475null_rename(struct vop_rename_args *ap)
476{
477 struct vnode *tdvp = ap->a_tdvp;
478 struct vnode *fvp = ap->a_fvp;
479 struct vnode *fdvp = ap->a_fdvp;
480 struct vnode *tvp = ap->a_tvp;
481
482 /* Check for cross-device rename. */
483 if ((fvp->v_mount != tdvp->v_mount) ||
484 (tvp && (fvp->v_mount != tvp->v_mount))) {
485 if (tdvp == tvp)
486 vrele(tdvp);
487 else
488 vput(tdvp);
489 if (tvp)
490 vput(tvp);
491 vrele(fdvp);
492 vrele(fvp);
493 return (EXDEV);
494 }
495
496 return (null_bypass((struct vop_generic_args *)ap));
497}
498
499/*
500 * We need to process our own vnode lock and then clear the
501 * interlock flag as it applies only to our vnode, not the
502 * vnodes below us on the stack.
503 */
504static int
505null_lock(struct vop_lock_args *ap)
506{
507 struct vnode *vp = ap->a_vp;
508 int flags = ap->a_flags;
509 struct thread *td = ap->a_td;
510 struct vnode *lvp;
511 int error;
512 struct null_node *nn;
513
514 if (flags & LK_THISLAYER) {
515 if (vp->v_vnlock != NULL) {
516 /* lock is shared across layers */
517 if (flags & LK_INTERLOCK)
518 mtx_unlock(&vp->v_interlock);
519 return 0;
520 }
521 error = lockmgr(&vp->v_lock, flags & ~LK_THISLAYER,
522 &vp->v_interlock, td);
523 return (error);
524 }
525
526 if (vp->v_vnlock != NULL) {
527 /*
528 * The lower level has exported a struct lock to us. Use
529 * it so that all vnodes in the stack lock and unlock
530 * simultaneously. Note: we don't DRAIN the lock as DRAIN
531 * decommissions the lock - just because our vnode is
532 * going away doesn't mean the struct lock below us is.
533 * LK_EXCLUSIVE is fine.
534 */
535 if ((flags & LK_INTERLOCK) == 0) {
536 VI_LOCK(vp);
537 flags |= LK_INTERLOCK;
538 }
539 nn = VTONULL(vp);
540 if ((flags & LK_TYPE_MASK) == LK_DRAIN) {
541 NULLFSDEBUG("null_lock: avoiding LK_DRAIN\n");
542 /*
543 * Emulate lock draining by waiting for all other
544 * pending locks to complete. Afterwards the
545 * lockmgr call might block, but no other threads
546 * will attempt to use this nullfs vnode due to the
547 * VI_XLOCK flag.
548 */
549 while (nn->null_pending_locks > 0) {
550 nn->null_drain_wakeup = 1;
551 msleep(&nn->null_pending_locks,
552 VI_MTX(vp),
553 PVFS,
554 "nuldr", 0);
555 }
556 error = lockmgr(vp->v_vnlock,
557 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE,
558 VI_MTX(vp), td);
559 return error;
560 }
561 nn->null_pending_locks++;
562 error = lockmgr(vp->v_vnlock, flags, &vp->v_interlock, td);
563 VI_LOCK(vp);
564 /*
565 * If we're called from vrele then v_usecount can have been 0
566 * and another process might have initiated a recycle
567 * operation. When that happens, just back out.
568 */
569 if (error == 0 && (vp->v_iflag & VI_XLOCK) != 0 &&
570 td != vp->v_vxthread) {
571 lockmgr(vp->v_vnlock,
572 (flags & ~LK_TYPE_MASK) | LK_RELEASE,
573 VI_MTX(vp), td);
574 VI_LOCK(vp);
575 error = ENOENT;
576 }
577 nn->null_pending_locks--;
578 /*
579 * Wakeup the process draining the vnode after all
580 * pending lock attempts has been failed.
581 */
582 if (nn->null_pending_locks == 0 &&
583 nn->null_drain_wakeup != 0) {
584 nn->null_drain_wakeup = 0;
585 wakeup(&nn->null_pending_locks);
586 }
587 if (error == ENOENT && (vp->v_iflag & VI_XLOCK) != 0 &&
588 vp->v_vxthread != curthread) {
589 vp->v_iflag |= VI_XWANT;
590 msleep(vp, VI_MTX(vp), PINOD, "nulbo", 0);
591 }
592 VI_UNLOCK(vp);
593 return error;
594 } else {
595 /*
596 * To prevent race conditions involving doing a lookup
597 * on "..", we have to lock the lower node, then lock our
598 * node. Most of the time it won't matter that we lock our
599 * node (as any locking would need the lower one locked
600 * first). But we can LK_DRAIN the upper lock as a step
601 * towards decomissioning it.
602 */
603 lvp = NULLVPTOLOWERVP(vp);
604 if (lvp == NULL)
605 return (lockmgr(&vp->v_lock, flags, &vp->v_interlock, td));
606 if (flags & LK_INTERLOCK) {
607 mtx_unlock(&vp->v_interlock);
608 flags &= ~LK_INTERLOCK;
609 }
610 if ((flags & LK_TYPE_MASK) == LK_DRAIN) {
611 error = VOP_LOCK(lvp,
612 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, td);
613 } else
614 error = VOP_LOCK(lvp, flags, td);
615 if (error)
616 return (error);
617 error = lockmgr(&vp->v_lock, flags, &vp->v_interlock, td);
618 if (error)
619 VOP_UNLOCK(lvp, 0, td);
620 return (error);
621 }
622}
623
624/*
625 * We need to process our own vnode unlock and then clear the
626 * interlock flag as it applies only to our vnode, not the
627 * vnodes below us on the stack.
628 */
629static int
630null_unlock(struct vop_unlock_args *ap)
631{
632 struct vnode *vp = ap->a_vp;
633 int flags = ap->a_flags;
634 struct thread *td = ap->a_td;
635 struct vnode *lvp;
636
637 if (vp->v_vnlock != NULL) {
638 if (flags & LK_THISLAYER)
639 return 0; /* the lock is shared across layers */
640 flags &= ~LK_THISLAYER;
641 return (lockmgr(vp->v_vnlock, flags | LK_RELEASE,
642 &vp->v_interlock, td));
643 }
644 lvp = NULLVPTOLOWERVP(vp);
645 if (lvp == NULL)
646 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td));
647 if ((flags & LK_THISLAYER) == 0) {
648 if (flags & LK_INTERLOCK) {
649 mtx_unlock(&vp->v_interlock);
650 flags &= ~LK_INTERLOCK;
651 }
652 VOP_UNLOCK(lvp, flags & ~LK_INTERLOCK, td);
653 } else
654 flags &= ~LK_THISLAYER;
655 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td));
656}
657
658static int
659null_islocked(struct vop_islocked_args *ap)
660{
661 struct vnode *vp = ap->a_vp;
662 struct thread *td = ap->a_td;
663
664 if (vp->v_vnlock != NULL)
665 return (lockstatus(vp->v_vnlock, td));
666 return (lockstatus(&vp->v_lock, td));
667}
668
669/*
670 * There is no way to tell that someone issued remove/rmdir operation
671 * on the underlying filesystem. For now we just have to release lowevrp
672 * as soon as possible.
673 *
674 * Note, we can't release any resources nor remove vnode from hash before
675 * appropriate VXLOCK stuff is is done because other process can find this
676 * vnode in hash during inactivation and may be sitting in vget() and waiting
677 * for null_inactive to unlock vnode. Thus we will do all those in VOP_RECLAIM.
678 */
679static int
680null_inactive(struct vop_inactive_args *ap)
681{
682 struct vnode *vp = ap->a_vp;
683 struct thread *td = ap->a_td;
684
685 VOP_UNLOCK(vp, 0, td);
686
687 /*
688 * If this is the last reference, then free up the vnode
689 * so as not to tie up the lower vnodes.
690 */
691 vrecycle(vp, NULL, td);
692
693 return (0);
694}
695
696/*
697 * Now, the VXLOCK is in force and we're free to destroy the null vnode.
698 */
699static int
700null_reclaim(struct vop_reclaim_args *ap)
701{
702 struct vnode *vp = ap->a_vp;
703 struct null_node *xp = VTONULL(vp);
704 struct vnode *lowervp = xp->null_lowervp;
705
706 if (lowervp) {
707 null_hashrem(xp);
708
709 vrele(lowervp);
710 vrele(lowervp);
711 }
712
713 vp->v_data = NULL;
714 vp->v_vnlock = &vp->v_lock;
715 FREE(xp, M_NULLFSNODE);
716
717 return (0);
718}
719
720static int
721null_print(struct vop_print_args *ap)
722{
723 register struct vnode *vp = ap->a_vp;
724 printf("\tvp=%p, lowervp=%p\n", vp, NULLVPTOLOWERVP(vp));
725 return (0);
726}
727
728/*
729 * Let an underlying filesystem do the work
730 */
731static int
732null_createvobject(struct vop_createvobject_args *ap)
733{
734 struct vnode *vp = ap->a_vp;
735 struct vnode *lowervp = VTONULL(vp) ? NULLVPTOLOWERVP(vp) : NULL;
736 int error;
737
738 if (vp->v_type == VNON || lowervp == NULL)
739 return 0;
740 error = VOP_CREATEVOBJECT(lowervp, ap->a_cred, ap->a_td);
741 if (error)
742 return (error);
743 vp->v_vflag |= VV_OBJBUF;
744 return (0);
745}
746
747/*
748 * We have nothing to destroy and this operation shouldn't be bypassed.
749 */
750static int
751null_destroyvobject(struct vop_destroyvobject_args *ap)
752{
753 struct vnode *vp = ap->a_vp;
754
755 vp->v_vflag &= ~VV_OBJBUF;
756 return (0);
757}
758
759static int
760null_getvobject(struct vop_getvobject_args *ap)
761{
762 struct vnode *lvp = NULLVPTOLOWERVP(ap->a_vp);
763
764 if (lvp == NULL)
765 return EINVAL;
766 return (VOP_GETVOBJECT(lvp, ap->a_objpp));
767}
768
769/*
770 * Global vfs data structures
771 */
772struct vop_vector null_vnodeops = {
773 .vop_bypass = null_bypass,
774
775 .vop_access = null_access,
776 .vop_bmap = VOP_EOPNOTSUPP,
777 .vop_createvobject = null_createvobject,
778 .vop_destroyvobject = null_destroyvobject,
779 .vop_getattr = null_getattr,
780 .vop_getvobject = null_getvobject,
781 .vop_getwritemount = vop_stdgetwritemount,
782 .vop_inactive = null_inactive,
783 .vop_islocked = null_islocked,
784 .vop_lock = null_lock,
785 .vop_lookup = null_lookup,
786 .vop_print = null_print,
787 .vop_reclaim = null_reclaim,
788 .vop_rename = null_rename,
789 .vop_setattr = null_setattr,
790 .vop_strategy = VOP_EOPNOTSUPP,
791 .vop_unlock = null_unlock,
792};
2 * Copyright (c) 1992, 1993
3 * The Regents of the University of California. All rights reserved.
4 *
5 * This code is derived from software contributed to Berkeley by
6 * John Heidemann of the UCLA Ficus project.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 *
32 * @(#)null_vnops.c 8.6 (Berkeley) 5/27/95
33 *
34 * Ancestors:
35 * @(#)lofs_vnops.c 1.2 (Berkeley) 6/18/92
36 * ...and...
37 * @(#)null_vnodeops.c 1.20 92/07/07 UCLA Ficus project
38 *
39 * $FreeBSD: head/sys/fs/nullfs/null_vnops.c 140728 2005-01-24 11:49:41Z phk $
40 */
41
42/*
43 * Null Layer
44 *
45 * (See mount_nullfs(8) for more information.)
46 *
47 * The null layer duplicates a portion of the filesystem
48 * name space under a new name. In this respect, it is
49 * similar to the loopback filesystem. It differs from
50 * the loopback fs in two respects: it is implemented using
51 * a stackable layers techniques, and its "null-node"s stack above
52 * all lower-layer vnodes, not just over directory vnodes.
53 *
54 * The null layer has two purposes. First, it serves as a demonstration
55 * of layering by proving a layer which does nothing. (It actually
56 * does everything the loopback filesystem does, which is slightly
57 * more than nothing.) Second, the null layer can serve as a prototype
58 * layer. Since it provides all necessary layer framework,
59 * new filesystem layers can be created very easily be starting
60 * with a null layer.
61 *
62 * The remainder of this man page examines the null layer as a basis
63 * for constructing new layers.
64 *
65 *
66 * INSTANTIATING NEW NULL LAYERS
67 *
68 * New null layers are created with mount_nullfs(8).
69 * Mount_nullfs(8) takes two arguments, the pathname
70 * of the lower vfs (target-pn) and the pathname where the null
71 * layer will appear in the namespace (alias-pn). After
72 * the null layer is put into place, the contents
73 * of target-pn subtree will be aliased under alias-pn.
74 *
75 *
76 * OPERATION OF A NULL LAYER
77 *
78 * The null layer is the minimum filesystem layer,
79 * simply bypassing all possible operations to the lower layer
80 * for processing there. The majority of its activity centers
81 * on the bypass routine, through which nearly all vnode operations
82 * pass.
83 *
84 * The bypass routine accepts arbitrary vnode operations for
85 * handling by the lower layer. It begins by examing vnode
86 * operation arguments and replacing any null-nodes by their
87 * lower-layer equivlants. It then invokes the operation
88 * on the lower layer. Finally, it replaces the null-nodes
89 * in the arguments and, if a vnode is return by the operation,
90 * stacks a null-node on top of the returned vnode.
91 *
92 * Although bypass handles most operations, vop_getattr, vop_lock,
93 * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not
94 * bypassed. Vop_getattr must change the fsid being returned.
95 * Vop_lock and vop_unlock must handle any locking for the
96 * current vnode as well as pass the lock request down.
97 * Vop_inactive and vop_reclaim are not bypassed so that
98 * they can handle freeing null-layer specific data. Vop_print
99 * is not bypassed to avoid excessive debugging information.
100 * Also, certain vnode operations change the locking state within
101 * the operation (create, mknod, remove, link, rename, mkdir, rmdir,
102 * and symlink). Ideally these operations should not change the
103 * lock state, but should be changed to let the caller of the
104 * function unlock them. Otherwise all intermediate vnode layers
105 * (such as union, umapfs, etc) must catch these functions to do
106 * the necessary locking at their layer.
107 *
108 *
109 * INSTANTIATING VNODE STACKS
110 *
111 * Mounting associates the null layer with a lower layer,
112 * effect stacking two VFSes. Vnode stacks are instead
113 * created on demand as files are accessed.
114 *
115 * The initial mount creates a single vnode stack for the
116 * root of the new null layer. All other vnode stacks
117 * are created as a result of vnode operations on
118 * this or other null vnode stacks.
119 *
120 * New vnode stacks come into existance as a result of
121 * an operation which returns a vnode.
122 * The bypass routine stacks a null-node above the new
123 * vnode before returning it to the caller.
124 *
125 * For example, imagine mounting a null layer with
126 * "mount_nullfs /usr/include /dev/layer/null".
127 * Changing directory to /dev/layer/null will assign
128 * the root null-node (which was created when the null layer was mounted).
129 * Now consider opening "sys". A vop_lookup would be
130 * done on the root null-node. This operation would bypass through
131 * to the lower layer which would return a vnode representing
132 * the UFS "sys". Null_bypass then builds a null-node
133 * aliasing the UFS "sys" and returns this to the caller.
134 * Later operations on the null-node "sys" will repeat this
135 * process when constructing other vnode stacks.
136 *
137 *
138 * CREATING OTHER FILE SYSTEM LAYERS
139 *
140 * One of the easiest ways to construct new filesystem layers is to make
141 * a copy of the null layer, rename all files and variables, and
142 * then begin modifing the copy. Sed can be used to easily rename
143 * all variables.
144 *
145 * The umap layer is an example of a layer descended from the
146 * null layer.
147 *
148 *
149 * INVOKING OPERATIONS ON LOWER LAYERS
150 *
151 * There are two techniques to invoke operations on a lower layer
152 * when the operation cannot be completely bypassed. Each method
153 * is appropriate in different situations. In both cases,
154 * it is the responsibility of the aliasing layer to make
155 * the operation arguments "correct" for the lower layer
156 * by mapping a vnode arguments to the lower layer.
157 *
158 * The first approach is to call the aliasing layer's bypass routine.
159 * This method is most suitable when you wish to invoke the operation
160 * currently being handled on the lower layer. It has the advantage
161 * that the bypass routine already must do argument mapping.
162 * An example of this is null_getattrs in the null layer.
163 *
164 * A second approach is to directly invoke vnode operations on
165 * the lower layer with the VOP_OPERATIONNAME interface.
166 * The advantage of this method is that it is easy to invoke
167 * arbitrary operations on the lower layer. The disadvantage
168 * is that vnode arguments must be manualy mapped.
169 *
170 */
171
172#include <sys/param.h>
173#include <sys/systm.h>
174#include <sys/conf.h>
175#include <sys/kernel.h>
176#include <sys/lock.h>
177#include <sys/malloc.h>
178#include <sys/mount.h>
179#include <sys/mutex.h>
180#include <sys/namei.h>
181#include <sys/sysctl.h>
182#include <sys/vnode.h>
183
184#include <fs/nullfs/null.h>
185
186#include <vm/vm.h>
187#include <vm/vm_extern.h>
188#include <vm/vm_object.h>
189#include <vm/vnode_pager.h>
190
191static int null_bug_bypass = 0; /* for debugging: enables bypass printf'ing */
192SYSCTL_INT(_debug, OID_AUTO, nullfs_bug_bypass, CTLFLAG_RW,
193 &null_bug_bypass, 0, "");
194
195/*
196 * This is the 10-Apr-92 bypass routine.
197 * This version has been optimized for speed, throwing away some
198 * safety checks. It should still always work, but it's not as
199 * robust to programmer errors.
200 *
201 * In general, we map all vnodes going down and unmap them on the way back.
202 * As an exception to this, vnodes can be marked "unmapped" by setting
203 * the Nth bit in operation's vdesc_flags.
204 *
205 * Also, some BSD vnode operations have the side effect of vrele'ing
206 * their arguments. With stacking, the reference counts are held
207 * by the upper node, not the lower one, so we must handle these
208 * side-effects here. This is not of concern in Sun-derived systems
209 * since there are no such side-effects.
210 *
211 * This makes the following assumptions:
212 * - only one returned vpp
213 * - no INOUT vpp's (Sun's vop_open has one of these)
214 * - the vnode operation vector of the first vnode should be used
215 * to determine what implementation of the op should be invoked
216 * - all mapped vnodes are of our vnode-type (NEEDSWORK:
217 * problems on rmdir'ing mount points and renaming?)
218 */
219int
220null_bypass(struct vop_generic_args *ap)
221{
222 register struct vnode **this_vp_p;
223 int error;
224 struct vnode *old_vps[VDESC_MAX_VPS];
225 struct vnode **vps_p[VDESC_MAX_VPS];
226 struct vnode ***vppp;
227 struct vnodeop_desc *descp = ap->a_desc;
228 int reles, i;
229
230 if (null_bug_bypass)
231 printf ("null_bypass: %s\n", descp->vdesc_name);
232
233#ifdef DIAGNOSTIC
234 /*
235 * We require at least one vp.
236 */
237 if (descp->vdesc_vp_offsets == NULL ||
238 descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET)
239 panic ("null_bypass: no vp's in map");
240#endif
241
242 /*
243 * Map the vnodes going in.
244 * Later, we'll invoke the operation based on
245 * the first mapped vnode's operation vector.
246 */
247 reles = descp->vdesc_flags;
248 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
249 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
250 break; /* bail out at end of list */
251 vps_p[i] = this_vp_p =
252 VOPARG_OFFSETTO(struct vnode**,descp->vdesc_vp_offsets[i],ap);
253 /*
254 * We're not guaranteed that any but the first vnode
255 * are of our type. Check for and don't map any
256 * that aren't. (We must always map first vp or vclean fails.)
257 */
258 if (i && (*this_vp_p == NULLVP ||
259 (*this_vp_p)->v_op != &null_vnodeops)) {
260 old_vps[i] = NULLVP;
261 } else {
262 old_vps[i] = *this_vp_p;
263 *(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p);
264 /*
265 * XXX - Several operations have the side effect
266 * of vrele'ing their vp's. We must account for
267 * that. (This should go away in the future.)
268 */
269 if (reles & VDESC_VP0_WILLRELE)
270 VREF(*this_vp_p);
271 }
272
273 }
274
275 /*
276 * Call the operation on the lower layer
277 * with the modified argument structure.
278 */
279 if (vps_p[0] && *vps_p[0])
280 error = VCALL(ap);
281 else {
282 printf("null_bypass: no map for %s\n", descp->vdesc_name);
283 error = EINVAL;
284 }
285
286 /*
287 * Maintain the illusion of call-by-value
288 * by restoring vnodes in the argument structure
289 * to their original value.
290 */
291 reles = descp->vdesc_flags;
292 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
293 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
294 break; /* bail out at end of list */
295 if (old_vps[i]) {
296 *(vps_p[i]) = old_vps[i];
297#if 0
298 if (reles & VDESC_VP0_WILLUNLOCK)
299 VOP_UNLOCK(*(vps_p[i]), LK_THISLAYER, curthread);
300#endif
301 if (reles & VDESC_VP0_WILLRELE)
302 vrele(*(vps_p[i]));
303 }
304 }
305
306 /*
307 * Map the possible out-going vpp
308 * (Assumes that the lower layer always returns
309 * a VREF'ed vpp unless it gets an error.)
310 */
311 if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET &&
312 !(descp->vdesc_flags & VDESC_NOMAP_VPP) &&
313 !error) {
314 /*
315 * XXX - even though some ops have vpp returned vp's,
316 * several ops actually vrele this before returning.
317 * We must avoid these ops.
318 * (This should go away when these ops are regularized.)
319 */
320 if (descp->vdesc_flags & VDESC_VPP_WILLRELE)
321 goto out;
322 vppp = VOPARG_OFFSETTO(struct vnode***,
323 descp->vdesc_vpp_offset,ap);
324 if (*vppp)
325 error = null_nodeget(old_vps[0]->v_mount, **vppp, *vppp);
326 }
327
328 out:
329 return (error);
330}
331
332/*
333 * We have to carry on the locking protocol on the null layer vnodes
334 * as we progress through the tree. We also have to enforce read-only
335 * if this layer is mounted read-only.
336 */
337static int
338null_lookup(struct vop_lookup_args *ap)
339{
340 struct componentname *cnp = ap->a_cnp;
341 struct vnode *dvp = ap->a_dvp;
342 struct thread *td = cnp->cn_thread;
343 int flags = cnp->cn_flags;
344 struct vnode *vp, *ldvp, *lvp;
345 int error;
346
347 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
348 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
349 return (EROFS);
350 /*
351 * Although it is possible to call null_bypass(), we'll do
352 * a direct call to reduce overhead
353 */
354 ldvp = NULLVPTOLOWERVP(dvp);
355 vp = lvp = NULL;
356 error = VOP_LOOKUP(ldvp, &lvp, cnp);
357 if (error == EJUSTRETURN && (flags & ISLASTCN) &&
358 (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
359 (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME))
360 error = EROFS;
361
362 /*
363 * Rely only on the PDIRUNLOCK flag which should be carefully
364 * tracked by underlying filesystem.
365 */
366 if ((cnp->cn_flags & PDIRUNLOCK) && dvp->v_vnlock != ldvp->v_vnlock)
367 VOP_UNLOCK(dvp, LK_THISLAYER, td);
368 if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) {
369 if (ldvp == lvp) {
370 *ap->a_vpp = dvp;
371 VREF(dvp);
372 vrele(lvp);
373 } else {
374 error = null_nodeget(dvp->v_mount, lvp, &vp);
375 if (error) {
376 /* XXX Cleanup needed... */
377 panic("null_nodeget failed");
378 }
379 *ap->a_vpp = vp;
380 }
381 }
382 return (error);
383}
384
385/*
386 * Setattr call. Disallow write attempts if the layer is mounted read-only.
387 */
388static int
389null_setattr(struct vop_setattr_args *ap)
390{
391 struct vnode *vp = ap->a_vp;
392 struct vattr *vap = ap->a_vap;
393
394 if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL ||
395 vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL ||
396 vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) &&
397 (vp->v_mount->mnt_flag & MNT_RDONLY))
398 return (EROFS);
399 if (vap->va_size != VNOVAL) {
400 switch (vp->v_type) {
401 case VDIR:
402 return (EISDIR);
403 case VCHR:
404 case VBLK:
405 case VSOCK:
406 case VFIFO:
407 if (vap->va_flags != VNOVAL)
408 return (EOPNOTSUPP);
409 return (0);
410 case VREG:
411 case VLNK:
412 default:
413 /*
414 * Disallow write attempts if the filesystem is
415 * mounted read-only.
416 */
417 if (vp->v_mount->mnt_flag & MNT_RDONLY)
418 return (EROFS);
419 }
420 }
421
422 return (null_bypass((struct vop_generic_args *)ap));
423}
424
425/*
426 * We handle getattr only to change the fsid.
427 */
428static int
429null_getattr(struct vop_getattr_args *ap)
430{
431 int error;
432
433 if ((error = null_bypass((struct vop_generic_args *)ap)) != 0)
434 return (error);
435
436 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
437 return (0);
438}
439
440/*
441 * Handle to disallow write access if mounted read-only.
442 */
443static int
444null_access(struct vop_access_args *ap)
445{
446 struct vnode *vp = ap->a_vp;
447 mode_t mode = ap->a_mode;
448
449 /*
450 * Disallow write attempts on read-only layers;
451 * unless the file is a socket, fifo, or a block or
452 * character device resident on the filesystem.
453 */
454 if (mode & VWRITE) {
455 switch (vp->v_type) {
456 case VDIR:
457 case VLNK:
458 case VREG:
459 if (vp->v_mount->mnt_flag & MNT_RDONLY)
460 return (EROFS);
461 break;
462 default:
463 break;
464 }
465 }
466 return (null_bypass((struct vop_generic_args *)ap));
467}
468
469/*
470 * We handle this to eliminate null FS to lower FS
471 * file moving. Don't know why we don't allow this,
472 * possibly we should.
473 */
474static int
475null_rename(struct vop_rename_args *ap)
476{
477 struct vnode *tdvp = ap->a_tdvp;
478 struct vnode *fvp = ap->a_fvp;
479 struct vnode *fdvp = ap->a_fdvp;
480 struct vnode *tvp = ap->a_tvp;
481
482 /* Check for cross-device rename. */
483 if ((fvp->v_mount != tdvp->v_mount) ||
484 (tvp && (fvp->v_mount != tvp->v_mount))) {
485 if (tdvp == tvp)
486 vrele(tdvp);
487 else
488 vput(tdvp);
489 if (tvp)
490 vput(tvp);
491 vrele(fdvp);
492 vrele(fvp);
493 return (EXDEV);
494 }
495
496 return (null_bypass((struct vop_generic_args *)ap));
497}
498
499/*
500 * We need to process our own vnode lock and then clear the
501 * interlock flag as it applies only to our vnode, not the
502 * vnodes below us on the stack.
503 */
504static int
505null_lock(struct vop_lock_args *ap)
506{
507 struct vnode *vp = ap->a_vp;
508 int flags = ap->a_flags;
509 struct thread *td = ap->a_td;
510 struct vnode *lvp;
511 int error;
512 struct null_node *nn;
513
514 if (flags & LK_THISLAYER) {
515 if (vp->v_vnlock != NULL) {
516 /* lock is shared across layers */
517 if (flags & LK_INTERLOCK)
518 mtx_unlock(&vp->v_interlock);
519 return 0;
520 }
521 error = lockmgr(&vp->v_lock, flags & ~LK_THISLAYER,
522 &vp->v_interlock, td);
523 return (error);
524 }
525
526 if (vp->v_vnlock != NULL) {
527 /*
528 * The lower level has exported a struct lock to us. Use
529 * it so that all vnodes in the stack lock and unlock
530 * simultaneously. Note: we don't DRAIN the lock as DRAIN
531 * decommissions the lock - just because our vnode is
532 * going away doesn't mean the struct lock below us is.
533 * LK_EXCLUSIVE is fine.
534 */
535 if ((flags & LK_INTERLOCK) == 0) {
536 VI_LOCK(vp);
537 flags |= LK_INTERLOCK;
538 }
539 nn = VTONULL(vp);
540 if ((flags & LK_TYPE_MASK) == LK_DRAIN) {
541 NULLFSDEBUG("null_lock: avoiding LK_DRAIN\n");
542 /*
543 * Emulate lock draining by waiting for all other
544 * pending locks to complete. Afterwards the
545 * lockmgr call might block, but no other threads
546 * will attempt to use this nullfs vnode due to the
547 * VI_XLOCK flag.
548 */
549 while (nn->null_pending_locks > 0) {
550 nn->null_drain_wakeup = 1;
551 msleep(&nn->null_pending_locks,
552 VI_MTX(vp),
553 PVFS,
554 "nuldr", 0);
555 }
556 error = lockmgr(vp->v_vnlock,
557 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE,
558 VI_MTX(vp), td);
559 return error;
560 }
561 nn->null_pending_locks++;
562 error = lockmgr(vp->v_vnlock, flags, &vp->v_interlock, td);
563 VI_LOCK(vp);
564 /*
565 * If we're called from vrele then v_usecount can have been 0
566 * and another process might have initiated a recycle
567 * operation. When that happens, just back out.
568 */
569 if (error == 0 && (vp->v_iflag & VI_XLOCK) != 0 &&
570 td != vp->v_vxthread) {
571 lockmgr(vp->v_vnlock,
572 (flags & ~LK_TYPE_MASK) | LK_RELEASE,
573 VI_MTX(vp), td);
574 VI_LOCK(vp);
575 error = ENOENT;
576 }
577 nn->null_pending_locks--;
578 /*
579 * Wakeup the process draining the vnode after all
580 * pending lock attempts has been failed.
581 */
582 if (nn->null_pending_locks == 0 &&
583 nn->null_drain_wakeup != 0) {
584 nn->null_drain_wakeup = 0;
585 wakeup(&nn->null_pending_locks);
586 }
587 if (error == ENOENT && (vp->v_iflag & VI_XLOCK) != 0 &&
588 vp->v_vxthread != curthread) {
589 vp->v_iflag |= VI_XWANT;
590 msleep(vp, VI_MTX(vp), PINOD, "nulbo", 0);
591 }
592 VI_UNLOCK(vp);
593 return error;
594 } else {
595 /*
596 * To prevent race conditions involving doing a lookup
597 * on "..", we have to lock the lower node, then lock our
598 * node. Most of the time it won't matter that we lock our
599 * node (as any locking would need the lower one locked
600 * first). But we can LK_DRAIN the upper lock as a step
601 * towards decomissioning it.
602 */
603 lvp = NULLVPTOLOWERVP(vp);
604 if (lvp == NULL)
605 return (lockmgr(&vp->v_lock, flags, &vp->v_interlock, td));
606 if (flags & LK_INTERLOCK) {
607 mtx_unlock(&vp->v_interlock);
608 flags &= ~LK_INTERLOCK;
609 }
610 if ((flags & LK_TYPE_MASK) == LK_DRAIN) {
611 error = VOP_LOCK(lvp,
612 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, td);
613 } else
614 error = VOP_LOCK(lvp, flags, td);
615 if (error)
616 return (error);
617 error = lockmgr(&vp->v_lock, flags, &vp->v_interlock, td);
618 if (error)
619 VOP_UNLOCK(lvp, 0, td);
620 return (error);
621 }
622}
623
624/*
625 * We need to process our own vnode unlock and then clear the
626 * interlock flag as it applies only to our vnode, not the
627 * vnodes below us on the stack.
628 */
629static int
630null_unlock(struct vop_unlock_args *ap)
631{
632 struct vnode *vp = ap->a_vp;
633 int flags = ap->a_flags;
634 struct thread *td = ap->a_td;
635 struct vnode *lvp;
636
637 if (vp->v_vnlock != NULL) {
638 if (flags & LK_THISLAYER)
639 return 0; /* the lock is shared across layers */
640 flags &= ~LK_THISLAYER;
641 return (lockmgr(vp->v_vnlock, flags | LK_RELEASE,
642 &vp->v_interlock, td));
643 }
644 lvp = NULLVPTOLOWERVP(vp);
645 if (lvp == NULL)
646 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td));
647 if ((flags & LK_THISLAYER) == 0) {
648 if (flags & LK_INTERLOCK) {
649 mtx_unlock(&vp->v_interlock);
650 flags &= ~LK_INTERLOCK;
651 }
652 VOP_UNLOCK(lvp, flags & ~LK_INTERLOCK, td);
653 } else
654 flags &= ~LK_THISLAYER;
655 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td));
656}
657
658static int
659null_islocked(struct vop_islocked_args *ap)
660{
661 struct vnode *vp = ap->a_vp;
662 struct thread *td = ap->a_td;
663
664 if (vp->v_vnlock != NULL)
665 return (lockstatus(vp->v_vnlock, td));
666 return (lockstatus(&vp->v_lock, td));
667}
668
669/*
670 * There is no way to tell that someone issued remove/rmdir operation
671 * on the underlying filesystem. For now we just have to release lowevrp
672 * as soon as possible.
673 *
674 * Note, we can't release any resources nor remove vnode from hash before
675 * appropriate VXLOCK stuff is is done because other process can find this
676 * vnode in hash during inactivation and may be sitting in vget() and waiting
677 * for null_inactive to unlock vnode. Thus we will do all those in VOP_RECLAIM.
678 */
679static int
680null_inactive(struct vop_inactive_args *ap)
681{
682 struct vnode *vp = ap->a_vp;
683 struct thread *td = ap->a_td;
684
685 VOP_UNLOCK(vp, 0, td);
686
687 /*
688 * If this is the last reference, then free up the vnode
689 * so as not to tie up the lower vnodes.
690 */
691 vrecycle(vp, NULL, td);
692
693 return (0);
694}
695
696/*
697 * Now, the VXLOCK is in force and we're free to destroy the null vnode.
698 */
699static int
700null_reclaim(struct vop_reclaim_args *ap)
701{
702 struct vnode *vp = ap->a_vp;
703 struct null_node *xp = VTONULL(vp);
704 struct vnode *lowervp = xp->null_lowervp;
705
706 if (lowervp) {
707 null_hashrem(xp);
708
709 vrele(lowervp);
710 vrele(lowervp);
711 }
712
713 vp->v_data = NULL;
714 vp->v_vnlock = &vp->v_lock;
715 FREE(xp, M_NULLFSNODE);
716
717 return (0);
718}
719
720static int
721null_print(struct vop_print_args *ap)
722{
723 register struct vnode *vp = ap->a_vp;
724 printf("\tvp=%p, lowervp=%p\n", vp, NULLVPTOLOWERVP(vp));
725 return (0);
726}
727
728/*
729 * Let an underlying filesystem do the work
730 */
731static int
732null_createvobject(struct vop_createvobject_args *ap)
733{
734 struct vnode *vp = ap->a_vp;
735 struct vnode *lowervp = VTONULL(vp) ? NULLVPTOLOWERVP(vp) : NULL;
736 int error;
737
738 if (vp->v_type == VNON || lowervp == NULL)
739 return 0;
740 error = VOP_CREATEVOBJECT(lowervp, ap->a_cred, ap->a_td);
741 if (error)
742 return (error);
743 vp->v_vflag |= VV_OBJBUF;
744 return (0);
745}
746
747/*
748 * We have nothing to destroy and this operation shouldn't be bypassed.
749 */
750static int
751null_destroyvobject(struct vop_destroyvobject_args *ap)
752{
753 struct vnode *vp = ap->a_vp;
754
755 vp->v_vflag &= ~VV_OBJBUF;
756 return (0);
757}
758
759static int
760null_getvobject(struct vop_getvobject_args *ap)
761{
762 struct vnode *lvp = NULLVPTOLOWERVP(ap->a_vp);
763
764 if (lvp == NULL)
765 return EINVAL;
766 return (VOP_GETVOBJECT(lvp, ap->a_objpp));
767}
768
769/*
770 * Global vfs data structures
771 */
772struct vop_vector null_vnodeops = {
773 .vop_bypass = null_bypass,
774
775 .vop_access = null_access,
776 .vop_bmap = VOP_EOPNOTSUPP,
777 .vop_createvobject = null_createvobject,
778 .vop_destroyvobject = null_destroyvobject,
779 .vop_getattr = null_getattr,
780 .vop_getvobject = null_getvobject,
781 .vop_getwritemount = vop_stdgetwritemount,
782 .vop_inactive = null_inactive,
783 .vop_islocked = null_islocked,
784 .vop_lock = null_lock,
785 .vop_lookup = null_lookup,
786 .vop_print = null_print,
787 .vop_reclaim = null_reclaim,
788 .vop_rename = null_rename,
789 .vop_setattr = null_setattr,
790 .vop_strategy = VOP_EOPNOTSUPP,
791 .vop_unlock = null_unlock,
792};