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 *
| 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 *
|
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 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]), 0);
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
332static int
333null_add_writecount(struct vop_add_writecount_args *ap)
334{
335 struct vnode *lvp, *vp;
336 int error;
337
338 vp = ap->a_vp;
339 lvp = NULLVPTOLOWERVP(vp);
340 KASSERT(vp->v_writecount + ap->a_inc >= 0, ("wrong writecount inc"));
341 if (vp->v_writecount > 0 && vp->v_writecount + ap->a_inc == 0)
342 error = VOP_ADD_WRITECOUNT(lvp, -1);
343 else if (vp->v_writecount == 0 && vp->v_writecount + ap->a_inc > 0)
344 error = VOP_ADD_WRITECOUNT(lvp, 1);
345 else
346 error = 0;
347 if (error == 0)
348 vp->v_writecount += ap->a_inc;
349 return (error);
350}
351
352/*
353 * We have to carry on the locking protocol on the null layer vnodes
354 * as we progress through the tree. We also have to enforce read-only
355 * if this layer is mounted read-only.
356 */
357static int
358null_lookup(struct vop_lookup_args *ap)
359{
360 struct componentname *cnp = ap->a_cnp;
361 struct vnode *dvp = ap->a_dvp;
362 int flags = cnp->cn_flags;
363 struct vnode *vp, *ldvp, *lvp;
364 struct mount *mp;
365 int error;
366
367 mp = dvp->v_mount;
368 if ((flags & ISLASTCN) != 0 && (mp->mnt_flag & MNT_RDONLY) != 0 &&
369 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
370 return (EROFS);
371 /*
372 * Although it is possible to call null_bypass(), we'll do
373 * a direct call to reduce overhead
374 */
375 ldvp = NULLVPTOLOWERVP(dvp);
376 vp = lvp = NULL;
377 KASSERT((ldvp->v_vflag & VV_ROOT) == 0 ||
378 ((dvp->v_vflag & VV_ROOT) != 0 && (flags & ISDOTDOT) == 0),
379 ("ldvp %p fl %#x dvp %p fl %#x flags %#x", ldvp, ldvp->v_vflag,
380 dvp, dvp->v_vflag, flags));
381
382 /*
383 * Hold ldvp. The reference on it, owned by dvp, is lost in
384 * case of dvp reclamation, and we need ldvp to move our lock
385 * from ldvp to dvp.
386 */
387 vhold(ldvp);
388
389 error = VOP_LOOKUP(ldvp, &lvp, cnp);
390
391 /*
392 * VOP_LOOKUP() on lower vnode may unlock ldvp, which allows
393 * dvp to be reclaimed due to shared v_vnlock. Check for the
394 * doomed state and return error.
395 */
396 if ((error == 0 || error == EJUSTRETURN) &&
397 (dvp->v_iflag & VI_DOOMED) != 0) {
398 error = ENOENT;
399 if (lvp != NULL)
400 vput(lvp);
401
402 /*
403 * If vgone() did reclaimed dvp before curthread
404 * relocked ldvp, the locks of dvp and ldpv are no
405 * longer shared. In this case, relock of ldvp in
406 * lower fs VOP_LOOKUP() does not restore the locking
407 * state of dvp. Compensate for this by unlocking
408 * ldvp and locking dvp, which is also correct if the
409 * locks are still shared.
410 */
411 VOP_UNLOCK(ldvp, 0);
412 vn_lock(dvp, LK_EXCLUSIVE | LK_RETRY);
413 }
414 vdrop(ldvp);
415
416 if (error == EJUSTRETURN && (flags & ISLASTCN) != 0 &&
417 (mp->mnt_flag & MNT_RDONLY) != 0 &&
418 (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME))
419 error = EROFS;
420
421 if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) {
422 if (ldvp == lvp) {
423 *ap->a_vpp = dvp;
424 VREF(dvp);
425 vrele(lvp);
426 } else {
427 error = null_nodeget(mp, lvp, &vp);
428 if (error == 0)
429 *ap->a_vpp = vp;
430 }
431 }
432 return (error);
433}
434
435static int
436null_open(struct vop_open_args *ap)
437{
438 int retval;
439 struct vnode *vp, *ldvp;
440
441 vp = ap->a_vp;
442 ldvp = NULLVPTOLOWERVP(vp);
443 retval = null_bypass(&ap->a_gen);
444 if (retval == 0)
445 vp->v_object = ldvp->v_object;
446 return (retval);
447}
448
449/*
450 * Setattr call. Disallow write attempts if the layer is mounted read-only.
451 */
452static int
453null_setattr(struct vop_setattr_args *ap)
454{
455 struct vnode *vp = ap->a_vp;
456 struct vattr *vap = ap->a_vap;
457
458 if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL ||
459 vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL ||
460 vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) &&
461 (vp->v_mount->mnt_flag & MNT_RDONLY))
462 return (EROFS);
463 if (vap->va_size != VNOVAL) {
464 switch (vp->v_type) {
465 case VDIR:
466 return (EISDIR);
467 case VCHR:
468 case VBLK:
469 case VSOCK:
470 case VFIFO:
471 if (vap->va_flags != VNOVAL)
472 return (EOPNOTSUPP);
473 return (0);
474 case VREG:
475 case VLNK:
476 default:
477 /*
478 * Disallow write attempts if the filesystem is
479 * mounted read-only.
480 */
481 if (vp->v_mount->mnt_flag & MNT_RDONLY)
482 return (EROFS);
483 }
484 }
485
486 return (null_bypass((struct vop_generic_args *)ap));
487}
488
489/*
490 * We handle getattr only to change the fsid.
491 */
492static int
493null_getattr(struct vop_getattr_args *ap)
494{
495 int error;
496
497 if ((error = null_bypass((struct vop_generic_args *)ap)) != 0)
498 return (error);
499
500 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
501 return (0);
502}
503
504/*
505 * Handle to disallow write access if mounted read-only.
506 */
507static int
508null_access(struct vop_access_args *ap)
509{
510 struct vnode *vp = ap->a_vp;
511 accmode_t accmode = ap->a_accmode;
512
513 /*
514 * Disallow write attempts on read-only layers;
515 * unless the file is a socket, fifo, or a block or
516 * character device resident on the filesystem.
517 */
518 if (accmode & VWRITE) {
519 switch (vp->v_type) {
520 case VDIR:
521 case VLNK:
522 case VREG:
523 if (vp->v_mount->mnt_flag & MNT_RDONLY)
524 return (EROFS);
525 break;
526 default:
527 break;
528 }
529 }
530 return (null_bypass((struct vop_generic_args *)ap));
531}
532
533static int
534null_accessx(struct vop_accessx_args *ap)
535{
536 struct vnode *vp = ap->a_vp;
537 accmode_t accmode = ap->a_accmode;
538
539 /*
540 * Disallow write attempts on read-only layers;
541 * unless the file is a socket, fifo, or a block or
542 * character device resident on the filesystem.
543 */
544 if (accmode & VWRITE) {
545 switch (vp->v_type) {
546 case VDIR:
547 case VLNK:
548 case VREG:
549 if (vp->v_mount->mnt_flag & MNT_RDONLY)
550 return (EROFS);
551 break;
552 default:
553 break;
554 }
555 }
556 return (null_bypass((struct vop_generic_args *)ap));
557}
558
559/*
560 * Increasing refcount of lower vnode is needed at least for the case
561 * when lower FS is NFS to do sillyrename if the file is in use.
562 * Unfortunately v_usecount is incremented in many places in
563 * the kernel and, as such, there may be races that result in
564 * the NFS client doing an extraneous silly rename, but that seems
565 * preferable to not doing a silly rename when it is needed.
566 */
567static int
568null_remove(struct vop_remove_args *ap)
569{
570 int retval, vreleit;
571 struct vnode *lvp, *vp;
572
573 vp = ap->a_vp;
574 if (vrefcnt(vp) > 1) {
575 lvp = NULLVPTOLOWERVP(vp);
576 VREF(lvp);
577 vreleit = 1;
578 } else
579 vreleit = 0;
580 VTONULL(vp)->null_flags |= NULLV_DROP;
581 retval = null_bypass(&ap->a_gen);
582 if (vreleit != 0)
583 vrele(lvp);
584 return (retval);
585}
586
587/*
588 * We handle this to eliminate null FS to lower FS
589 * file moving. Don't know why we don't allow this,
590 * possibly we should.
591 */
592static int
593null_rename(struct vop_rename_args *ap)
594{
595 struct vnode *tdvp = ap->a_tdvp;
596 struct vnode *fvp = ap->a_fvp;
597 struct vnode *fdvp = ap->a_fdvp;
598 struct vnode *tvp = ap->a_tvp;
599 struct null_node *tnn;
600
601 /* Check for cross-device rename. */
602 if ((fvp->v_mount != tdvp->v_mount) ||
603 (tvp && (fvp->v_mount != tvp->v_mount))) {
604 if (tdvp == tvp)
605 vrele(tdvp);
606 else
607 vput(tdvp);
608 if (tvp)
609 vput(tvp);
610 vrele(fdvp);
611 vrele(fvp);
612 return (EXDEV);
613 }
614
615 if (tvp != NULL) {
616 tnn = VTONULL(tvp);
617 tnn->null_flags |= NULLV_DROP;
618 }
619 return (null_bypass((struct vop_generic_args *)ap));
620}
621
622static int
623null_rmdir(struct vop_rmdir_args *ap)
624{
625
626 VTONULL(ap->a_vp)->null_flags |= NULLV_DROP;
627 return (null_bypass(&ap->a_gen));
628}
629
630/*
631 * We need to process our own vnode lock and then clear the
632 * interlock flag as it applies only to our vnode, not the
633 * vnodes below us on the stack.
634 */
635static int
636null_lock(struct vop_lock1_args *ap)
637{
638 struct vnode *vp = ap->a_vp;
639 int flags = ap->a_flags;
640 struct null_node *nn;
641 struct vnode *lvp;
642 int error;
643
644
645 if ((flags & LK_INTERLOCK) == 0) {
646 VI_LOCK(vp);
647 ap->a_flags = flags |= LK_INTERLOCK;
648 }
649 nn = VTONULL(vp);
650 /*
651 * If we're still active we must ask the lower layer to
652 * lock as ffs has special lock considerations in it's
653 * vop lock.
654 */
655 if (nn != NULL && (lvp = NULLVPTOLOWERVP(vp)) != NULL) {
656 VI_LOCK_FLAGS(lvp, MTX_DUPOK);
657 VI_UNLOCK(vp);
658 /*
659 * We have to hold the vnode here to solve a potential
660 * reclaim race. If we're forcibly vgone'd while we
661 * still have refs, a thread could be sleeping inside
662 * the lowervp's vop_lock routine. When we vgone we will
663 * drop our last ref to the lowervp, which would allow it
664 * to be reclaimed. The lowervp could then be recycled,
665 * in which case it is not legal to be sleeping in it's VOP.
666 * We prevent it from being recycled by holding the vnode
667 * here.
668 */
669 vholdl(lvp);
670 error = VOP_LOCK(lvp, flags);
671
672 /*
673 * We might have slept to get the lock and someone might have
674 * clean our vnode already, switching vnode lock from one in
675 * lowervp to v_lock in our own vnode structure. Handle this
676 * case by reacquiring correct lock in requested mode.
677 */
678 if (VTONULL(vp) == NULL && error == 0) {
679 ap->a_flags &= ~(LK_TYPE_MASK | LK_INTERLOCK);
680 switch (flags & LK_TYPE_MASK) {
681 case LK_SHARED:
682 ap->a_flags |= LK_SHARED;
683 break;
684 case LK_UPGRADE:
685 case LK_EXCLUSIVE:
686 ap->a_flags |= LK_EXCLUSIVE;
687 break;
688 default:
689 panic("Unsupported lock request %d\n",
690 ap->a_flags);
691 }
692 VOP_UNLOCK(lvp, 0);
693 error = vop_stdlock(ap);
694 }
695 vdrop(lvp);
696 } else
697 error = vop_stdlock(ap);
698
699 return (error);
700}
701
702/*
703 * We need to process our own vnode unlock and then clear the
704 * interlock flag as it applies only to our vnode, not the
705 * vnodes below us on the stack.
706 */
707static int
708null_unlock(struct vop_unlock_args *ap)
709{
710 struct vnode *vp = ap->a_vp;
711 int flags = ap->a_flags;
712 int mtxlkflag = 0;
713 struct null_node *nn;
714 struct vnode *lvp;
715 int error;
716
717 if ((flags & LK_INTERLOCK) != 0)
718 mtxlkflag = 1;
719 else if (mtx_owned(VI_MTX(vp)) == 0) {
720 VI_LOCK(vp);
721 mtxlkflag = 2;
722 }
723 nn = VTONULL(vp);
724 if (nn != NULL && (lvp = NULLVPTOLOWERVP(vp)) != NULL) {
725 VI_LOCK_FLAGS(lvp, MTX_DUPOK);
726 flags |= LK_INTERLOCK;
727 vholdl(lvp);
728 VI_UNLOCK(vp);
729 error = VOP_UNLOCK(lvp, flags);
730 vdrop(lvp);
731 if (mtxlkflag == 0)
732 VI_LOCK(vp);
733 } else {
734 if (mtxlkflag == 2)
735 VI_UNLOCK(vp);
736 error = vop_stdunlock(ap);
737 }
738
739 return (error);
740}
741
742/*
743 * Do not allow the VOP_INACTIVE to be passed to the lower layer,
744 * since the reference count on the lower vnode is not related to
745 * ours.
746 */
747static int
748null_inactive(struct vop_inactive_args *ap __unused)
749{
750 struct vnode *vp, *lvp;
751 struct null_node *xp;
752 struct mount *mp;
753 struct null_mount *xmp;
754
755 vp = ap->a_vp;
756 xp = VTONULL(vp);
757 lvp = NULLVPTOLOWERVP(vp);
758 mp = vp->v_mount;
759 xmp = MOUNTTONULLMOUNT(mp);
760 if ((xmp->nullm_flags & NULLM_CACHE) == 0 ||
761 (xp->null_flags & NULLV_DROP) != 0 ||
762 (lvp->v_vflag & VV_NOSYNC) != 0) {
763 /*
764 * If this is the last reference and caching of the
765 * nullfs vnodes is not enabled, or the lower vnode is
766 * deleted, then free up the vnode so as not to tie up
767 * the lower vnodes.
768 */
769 vp->v_object = NULL;
770 vrecycle(vp);
771 }
772 return (0);
773}
774
775/*
776 * Now, the nullfs vnode and, due to the sharing lock, the lower
777 * vnode, are exclusively locked, and we shall destroy the null vnode.
778 */
779static int
780null_reclaim(struct vop_reclaim_args *ap)
781{
782 struct vnode *vp;
783 struct null_node *xp;
784 struct vnode *lowervp;
785
786 vp = ap->a_vp;
787 xp = VTONULL(vp);
788 lowervp = xp->null_lowervp;
789
790 KASSERT(lowervp != NULL && vp->v_vnlock != &vp->v_lock,
791 ("Reclaiming incomplete null vnode %p", vp));
792
793 null_hashrem(xp);
794 /*
795 * Use the interlock to protect the clearing of v_data to
796 * prevent faults in null_lock().
797 */
798 lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
799 VI_LOCK(vp);
800 vp->v_data = NULL;
801 vp->v_object = NULL;
802 vp->v_vnlock = &vp->v_lock;
803 VI_UNLOCK(vp);
804
805 /*
806 * If we were opened for write, we leased one write reference
807 * to the lower vnode. If this is a reclamation due to the
808 * forced unmount, undo the reference now.
809 */
810 if (vp->v_writecount > 0)
811 VOP_ADD_WRITECOUNT(lowervp, -1);
812 if ((xp->null_flags & NULLV_NOUNLOCK) != 0)
813 vunref(lowervp);
814 else
815 vput(lowervp);
816 free(xp, M_NULLFSNODE);
817
818 return (0);
819}
820
821static int
822null_print(struct vop_print_args *ap)
823{
824 struct vnode *vp = ap->a_vp;
825
826 printf("\tvp=%p, lowervp=%p\n", vp, VTONULL(vp)->null_lowervp);
827 return (0);
828}
829
830/* ARGSUSED */
831static int
832null_getwritemount(struct vop_getwritemount_args *ap)
833{
834 struct null_node *xp;
835 struct vnode *lowervp;
836 struct vnode *vp;
837
838 vp = ap->a_vp;
839 VI_LOCK(vp);
840 xp = VTONULL(vp);
841 if (xp && (lowervp = xp->null_lowervp)) {
842 VI_LOCK_FLAGS(lowervp, MTX_DUPOK);
843 VI_UNLOCK(vp);
844 vholdl(lowervp);
845 VI_UNLOCK(lowervp);
846 VOP_GETWRITEMOUNT(lowervp, ap->a_mpp);
847 vdrop(lowervp);
848 } else {
849 VI_UNLOCK(vp);
850 *(ap->a_mpp) = NULL;
851 }
852 return (0);
853}
854
855static int
856null_vptofh(struct vop_vptofh_args *ap)
857{
858 struct vnode *lvp;
859
860 lvp = NULLVPTOLOWERVP(ap->a_vp);
861 return VOP_VPTOFH(lvp, ap->a_fhp);
862}
863
864static int
865null_vptocnp(struct vop_vptocnp_args *ap)
866{
867 struct vnode *vp = ap->a_vp;
868 struct vnode **dvp = ap->a_vpp;
869 struct vnode *lvp, *ldvp;
870 struct ucred *cred = ap->a_cred;
871 int error, locked;
872
873 if (vp->v_type == VDIR)
874 return (vop_stdvptocnp(ap));
875
876 locked = VOP_ISLOCKED(vp);
877 lvp = NULLVPTOLOWERVP(vp);
878 vhold(lvp);
879 VOP_UNLOCK(vp, 0); /* vp is held by vn_vptocnp_locked that called us */
880 ldvp = lvp;
881 vref(lvp);
882 error = vn_vptocnp(&ldvp, cred, ap->a_buf, ap->a_buflen);
883 vdrop(lvp);
884 if (error != 0) {
885 vn_lock(vp, locked | LK_RETRY);
886 return (ENOENT);
887 }
888
889 /*
890 * Exclusive lock is required by insmntque1 call in
891 * null_nodeget()
892 */
893 error = vn_lock(ldvp, LK_EXCLUSIVE);
894 if (error != 0) {
895 vrele(ldvp);
896 vn_lock(vp, locked | LK_RETRY);
897 return (ENOENT);
898 }
899 vref(ldvp);
900 error = null_nodeget(vp->v_mount, ldvp, dvp);
901 if (error == 0) {
902#ifdef DIAGNOSTIC
903 NULLVPTOLOWERVP(*dvp);
904#endif
905 VOP_UNLOCK(*dvp, 0); /* keep reference on *dvp */
906 }
907 vn_lock(vp, locked | LK_RETRY);
908 return (error);
909}
910
911/*
912 * Global vfs data structures
913 */
914struct vop_vector null_vnodeops = {
915 .vop_bypass = null_bypass,
916 .vop_access = null_access,
917 .vop_accessx = null_accessx,
918 .vop_advlockpurge = vop_stdadvlockpurge,
919 .vop_bmap = VOP_EOPNOTSUPP,
920 .vop_getattr = null_getattr,
921 .vop_getwritemount = null_getwritemount,
922 .vop_inactive = null_inactive,
923 .vop_islocked = vop_stdislocked,
924 .vop_lock1 = null_lock,
925 .vop_lookup = null_lookup,
926 .vop_open = null_open,
927 .vop_print = null_print,
928 .vop_reclaim = null_reclaim,
929 .vop_remove = null_remove,
930 .vop_rename = null_rename,
931 .vop_rmdir = null_rmdir,
932 .vop_setattr = null_setattr,
933 .vop_strategy = VOP_EOPNOTSUPP,
934 .vop_unlock = null_unlock,
935 .vop_vptocnp = null_vptocnp,
936 .vop_vptofh = null_vptofh,
937 .vop_add_writecount = null_add_writecount,
938};
| 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 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]), 0);
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
332static int
333null_add_writecount(struct vop_add_writecount_args *ap)
334{
335 struct vnode *lvp, *vp;
336 int error;
337
338 vp = ap->a_vp;
339 lvp = NULLVPTOLOWERVP(vp);
340 KASSERT(vp->v_writecount + ap->a_inc >= 0, ("wrong writecount inc"));
341 if (vp->v_writecount > 0 && vp->v_writecount + ap->a_inc == 0)
342 error = VOP_ADD_WRITECOUNT(lvp, -1);
343 else if (vp->v_writecount == 0 && vp->v_writecount + ap->a_inc > 0)
344 error = VOP_ADD_WRITECOUNT(lvp, 1);
345 else
346 error = 0;
347 if (error == 0)
348 vp->v_writecount += ap->a_inc;
349 return (error);
350}
351
352/*
353 * We have to carry on the locking protocol on the null layer vnodes
354 * as we progress through the tree. We also have to enforce read-only
355 * if this layer is mounted read-only.
356 */
357static int
358null_lookup(struct vop_lookup_args *ap)
359{
360 struct componentname *cnp = ap->a_cnp;
361 struct vnode *dvp = ap->a_dvp;
362 int flags = cnp->cn_flags;
363 struct vnode *vp, *ldvp, *lvp;
364 struct mount *mp;
365 int error;
366
367 mp = dvp->v_mount;
368 if ((flags & ISLASTCN) != 0 && (mp->mnt_flag & MNT_RDONLY) != 0 &&
369 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
370 return (EROFS);
371 /*
372 * Although it is possible to call null_bypass(), we'll do
373 * a direct call to reduce overhead
374 */
375 ldvp = NULLVPTOLOWERVP(dvp);
376 vp = lvp = NULL;
377 KASSERT((ldvp->v_vflag & VV_ROOT) == 0 ||
378 ((dvp->v_vflag & VV_ROOT) != 0 && (flags & ISDOTDOT) == 0),
379 ("ldvp %p fl %#x dvp %p fl %#x flags %#x", ldvp, ldvp->v_vflag,
380 dvp, dvp->v_vflag, flags));
381
382 /*
383 * Hold ldvp. The reference on it, owned by dvp, is lost in
384 * case of dvp reclamation, and we need ldvp to move our lock
385 * from ldvp to dvp.
386 */
387 vhold(ldvp);
388
389 error = VOP_LOOKUP(ldvp, &lvp, cnp);
390
391 /*
392 * VOP_LOOKUP() on lower vnode may unlock ldvp, which allows
393 * dvp to be reclaimed due to shared v_vnlock. Check for the
394 * doomed state and return error.
395 */
396 if ((error == 0 || error == EJUSTRETURN) &&
397 (dvp->v_iflag & VI_DOOMED) != 0) {
398 error = ENOENT;
399 if (lvp != NULL)
400 vput(lvp);
401
402 /*
403 * If vgone() did reclaimed dvp before curthread
404 * relocked ldvp, the locks of dvp and ldpv are no
405 * longer shared. In this case, relock of ldvp in
406 * lower fs VOP_LOOKUP() does not restore the locking
407 * state of dvp. Compensate for this by unlocking
408 * ldvp and locking dvp, which is also correct if the
409 * locks are still shared.
410 */
411 VOP_UNLOCK(ldvp, 0);
412 vn_lock(dvp, LK_EXCLUSIVE | LK_RETRY);
413 }
414 vdrop(ldvp);
415
416 if (error == EJUSTRETURN && (flags & ISLASTCN) != 0 &&
417 (mp->mnt_flag & MNT_RDONLY) != 0 &&
418 (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME))
419 error = EROFS;
420
421 if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) {
422 if (ldvp == lvp) {
423 *ap->a_vpp = dvp;
424 VREF(dvp);
425 vrele(lvp);
426 } else {
427 error = null_nodeget(mp, lvp, &vp);
428 if (error == 0)
429 *ap->a_vpp = vp;
430 }
431 }
432 return (error);
433}
434
435static int
436null_open(struct vop_open_args *ap)
437{
438 int retval;
439 struct vnode *vp, *ldvp;
440
441 vp = ap->a_vp;
442 ldvp = NULLVPTOLOWERVP(vp);
443 retval = null_bypass(&ap->a_gen);
444 if (retval == 0)
445 vp->v_object = ldvp->v_object;
446 return (retval);
447}
448
449/*
450 * Setattr call. Disallow write attempts if the layer is mounted read-only.
451 */
452static int
453null_setattr(struct vop_setattr_args *ap)
454{
455 struct vnode *vp = ap->a_vp;
456 struct vattr *vap = ap->a_vap;
457
458 if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL ||
459 vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL ||
460 vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) &&
461 (vp->v_mount->mnt_flag & MNT_RDONLY))
462 return (EROFS);
463 if (vap->va_size != VNOVAL) {
464 switch (vp->v_type) {
465 case VDIR:
466 return (EISDIR);
467 case VCHR:
468 case VBLK:
469 case VSOCK:
470 case VFIFO:
471 if (vap->va_flags != VNOVAL)
472 return (EOPNOTSUPP);
473 return (0);
474 case VREG:
475 case VLNK:
476 default:
477 /*
478 * Disallow write attempts if the filesystem is
479 * mounted read-only.
480 */
481 if (vp->v_mount->mnt_flag & MNT_RDONLY)
482 return (EROFS);
483 }
484 }
485
486 return (null_bypass((struct vop_generic_args *)ap));
487}
488
489/*
490 * We handle getattr only to change the fsid.
491 */
492static int
493null_getattr(struct vop_getattr_args *ap)
494{
495 int error;
496
497 if ((error = null_bypass((struct vop_generic_args *)ap)) != 0)
498 return (error);
499
500 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
501 return (0);
502}
503
504/*
505 * Handle to disallow write access if mounted read-only.
506 */
507static int
508null_access(struct vop_access_args *ap)
509{
510 struct vnode *vp = ap->a_vp;
511 accmode_t accmode = ap->a_accmode;
512
513 /*
514 * Disallow write attempts on read-only layers;
515 * unless the file is a socket, fifo, or a block or
516 * character device resident on the filesystem.
517 */
518 if (accmode & VWRITE) {
519 switch (vp->v_type) {
520 case VDIR:
521 case VLNK:
522 case VREG:
523 if (vp->v_mount->mnt_flag & MNT_RDONLY)
524 return (EROFS);
525 break;
526 default:
527 break;
528 }
529 }
530 return (null_bypass((struct vop_generic_args *)ap));
531}
532
533static int
534null_accessx(struct vop_accessx_args *ap)
535{
536 struct vnode *vp = ap->a_vp;
537 accmode_t accmode = ap->a_accmode;
538
539 /*
540 * Disallow write attempts on read-only layers;
541 * unless the file is a socket, fifo, or a block or
542 * character device resident on the filesystem.
543 */
544 if (accmode & VWRITE) {
545 switch (vp->v_type) {
546 case VDIR:
547 case VLNK:
548 case VREG:
549 if (vp->v_mount->mnt_flag & MNT_RDONLY)
550 return (EROFS);
551 break;
552 default:
553 break;
554 }
555 }
556 return (null_bypass((struct vop_generic_args *)ap));
557}
558
559/*
560 * Increasing refcount of lower vnode is needed at least for the case
561 * when lower FS is NFS to do sillyrename if the file is in use.
562 * Unfortunately v_usecount is incremented in many places in
563 * the kernel and, as such, there may be races that result in
564 * the NFS client doing an extraneous silly rename, but that seems
565 * preferable to not doing a silly rename when it is needed.
566 */
567static int
568null_remove(struct vop_remove_args *ap)
569{
570 int retval, vreleit;
571 struct vnode *lvp, *vp;
572
573 vp = ap->a_vp;
574 if (vrefcnt(vp) > 1) {
575 lvp = NULLVPTOLOWERVP(vp);
576 VREF(lvp);
577 vreleit = 1;
578 } else
579 vreleit = 0;
580 VTONULL(vp)->null_flags |= NULLV_DROP;
581 retval = null_bypass(&ap->a_gen);
582 if (vreleit != 0)
583 vrele(lvp);
584 return (retval);
585}
586
587/*
588 * We handle this to eliminate null FS to lower FS
589 * file moving. Don't know why we don't allow this,
590 * possibly we should.
591 */
592static int
593null_rename(struct vop_rename_args *ap)
594{
595 struct vnode *tdvp = ap->a_tdvp;
596 struct vnode *fvp = ap->a_fvp;
597 struct vnode *fdvp = ap->a_fdvp;
598 struct vnode *tvp = ap->a_tvp;
599 struct null_node *tnn;
600
601 /* Check for cross-device rename. */
602 if ((fvp->v_mount != tdvp->v_mount) ||
603 (tvp && (fvp->v_mount != tvp->v_mount))) {
604 if (tdvp == tvp)
605 vrele(tdvp);
606 else
607 vput(tdvp);
608 if (tvp)
609 vput(tvp);
610 vrele(fdvp);
611 vrele(fvp);
612 return (EXDEV);
613 }
614
615 if (tvp != NULL) {
616 tnn = VTONULL(tvp);
617 tnn->null_flags |= NULLV_DROP;
618 }
619 return (null_bypass((struct vop_generic_args *)ap));
620}
621
622static int
623null_rmdir(struct vop_rmdir_args *ap)
624{
625
626 VTONULL(ap->a_vp)->null_flags |= NULLV_DROP;
627 return (null_bypass(&ap->a_gen));
628}
629
630/*
631 * We need to process our own vnode lock and then clear the
632 * interlock flag as it applies only to our vnode, not the
633 * vnodes below us on the stack.
634 */
635static int
636null_lock(struct vop_lock1_args *ap)
637{
638 struct vnode *vp = ap->a_vp;
639 int flags = ap->a_flags;
640 struct null_node *nn;
641 struct vnode *lvp;
642 int error;
643
644
645 if ((flags & LK_INTERLOCK) == 0) {
646 VI_LOCK(vp);
647 ap->a_flags = flags |= LK_INTERLOCK;
648 }
649 nn = VTONULL(vp);
650 /*
651 * If we're still active we must ask the lower layer to
652 * lock as ffs has special lock considerations in it's
653 * vop lock.
654 */
655 if (nn != NULL && (lvp = NULLVPTOLOWERVP(vp)) != NULL) {
656 VI_LOCK_FLAGS(lvp, MTX_DUPOK);
657 VI_UNLOCK(vp);
658 /*
659 * We have to hold the vnode here to solve a potential
660 * reclaim race. If we're forcibly vgone'd while we
661 * still have refs, a thread could be sleeping inside
662 * the lowervp's vop_lock routine. When we vgone we will
663 * drop our last ref to the lowervp, which would allow it
664 * to be reclaimed. The lowervp could then be recycled,
665 * in which case it is not legal to be sleeping in it's VOP.
666 * We prevent it from being recycled by holding the vnode
667 * here.
668 */
669 vholdl(lvp);
670 error = VOP_LOCK(lvp, flags);
671
672 /*
673 * We might have slept to get the lock and someone might have
674 * clean our vnode already, switching vnode lock from one in
675 * lowervp to v_lock in our own vnode structure. Handle this
676 * case by reacquiring correct lock in requested mode.
677 */
678 if (VTONULL(vp) == NULL && error == 0) {
679 ap->a_flags &= ~(LK_TYPE_MASK | LK_INTERLOCK);
680 switch (flags & LK_TYPE_MASK) {
681 case LK_SHARED:
682 ap->a_flags |= LK_SHARED;
683 break;
684 case LK_UPGRADE:
685 case LK_EXCLUSIVE:
686 ap->a_flags |= LK_EXCLUSIVE;
687 break;
688 default:
689 panic("Unsupported lock request %d\n",
690 ap->a_flags);
691 }
692 VOP_UNLOCK(lvp, 0);
693 error = vop_stdlock(ap);
694 }
695 vdrop(lvp);
696 } else
697 error = vop_stdlock(ap);
698
699 return (error);
700}
701
702/*
703 * We need to process our own vnode unlock and then clear the
704 * interlock flag as it applies only to our vnode, not the
705 * vnodes below us on the stack.
706 */
707static int
708null_unlock(struct vop_unlock_args *ap)
709{
710 struct vnode *vp = ap->a_vp;
711 int flags = ap->a_flags;
712 int mtxlkflag = 0;
713 struct null_node *nn;
714 struct vnode *lvp;
715 int error;
716
717 if ((flags & LK_INTERLOCK) != 0)
718 mtxlkflag = 1;
719 else if (mtx_owned(VI_MTX(vp)) == 0) {
720 VI_LOCK(vp);
721 mtxlkflag = 2;
722 }
723 nn = VTONULL(vp);
724 if (nn != NULL && (lvp = NULLVPTOLOWERVP(vp)) != NULL) {
725 VI_LOCK_FLAGS(lvp, MTX_DUPOK);
726 flags |= LK_INTERLOCK;
727 vholdl(lvp);
728 VI_UNLOCK(vp);
729 error = VOP_UNLOCK(lvp, flags);
730 vdrop(lvp);
731 if (mtxlkflag == 0)
732 VI_LOCK(vp);
733 } else {
734 if (mtxlkflag == 2)
735 VI_UNLOCK(vp);
736 error = vop_stdunlock(ap);
737 }
738
739 return (error);
740}
741
742/*
743 * Do not allow the VOP_INACTIVE to be passed to the lower layer,
744 * since the reference count on the lower vnode is not related to
745 * ours.
746 */
747static int
748null_inactive(struct vop_inactive_args *ap __unused)
749{
750 struct vnode *vp, *lvp;
751 struct null_node *xp;
752 struct mount *mp;
753 struct null_mount *xmp;
754
755 vp = ap->a_vp;
756 xp = VTONULL(vp);
757 lvp = NULLVPTOLOWERVP(vp);
758 mp = vp->v_mount;
759 xmp = MOUNTTONULLMOUNT(mp);
760 if ((xmp->nullm_flags & NULLM_CACHE) == 0 ||
761 (xp->null_flags & NULLV_DROP) != 0 ||
762 (lvp->v_vflag & VV_NOSYNC) != 0) {
763 /*
764 * If this is the last reference and caching of the
765 * nullfs vnodes is not enabled, or the lower vnode is
766 * deleted, then free up the vnode so as not to tie up
767 * the lower vnodes.
768 */
769 vp->v_object = NULL;
770 vrecycle(vp);
771 }
772 return (0);
773}
774
775/*
776 * Now, the nullfs vnode and, due to the sharing lock, the lower
777 * vnode, are exclusively locked, and we shall destroy the null vnode.
778 */
779static int
780null_reclaim(struct vop_reclaim_args *ap)
781{
782 struct vnode *vp;
783 struct null_node *xp;
784 struct vnode *lowervp;
785
786 vp = ap->a_vp;
787 xp = VTONULL(vp);
788 lowervp = xp->null_lowervp;
789
790 KASSERT(lowervp != NULL && vp->v_vnlock != &vp->v_lock,
791 ("Reclaiming incomplete null vnode %p", vp));
792
793 null_hashrem(xp);
794 /*
795 * Use the interlock to protect the clearing of v_data to
796 * prevent faults in null_lock().
797 */
798 lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
799 VI_LOCK(vp);
800 vp->v_data = NULL;
801 vp->v_object = NULL;
802 vp->v_vnlock = &vp->v_lock;
803 VI_UNLOCK(vp);
804
805 /*
806 * If we were opened for write, we leased one write reference
807 * to the lower vnode. If this is a reclamation due to the
808 * forced unmount, undo the reference now.
809 */
810 if (vp->v_writecount > 0)
811 VOP_ADD_WRITECOUNT(lowervp, -1);
812 if ((xp->null_flags & NULLV_NOUNLOCK) != 0)
813 vunref(lowervp);
814 else
815 vput(lowervp);
816 free(xp, M_NULLFSNODE);
817
818 return (0);
819}
820
821static int
822null_print(struct vop_print_args *ap)
823{
824 struct vnode *vp = ap->a_vp;
825
826 printf("\tvp=%p, lowervp=%p\n", vp, VTONULL(vp)->null_lowervp);
827 return (0);
828}
829
830/* ARGSUSED */
831static int
832null_getwritemount(struct vop_getwritemount_args *ap)
833{
834 struct null_node *xp;
835 struct vnode *lowervp;
836 struct vnode *vp;
837
838 vp = ap->a_vp;
839 VI_LOCK(vp);
840 xp = VTONULL(vp);
841 if (xp && (lowervp = xp->null_lowervp)) {
842 VI_LOCK_FLAGS(lowervp, MTX_DUPOK);
843 VI_UNLOCK(vp);
844 vholdl(lowervp);
845 VI_UNLOCK(lowervp);
846 VOP_GETWRITEMOUNT(lowervp, ap->a_mpp);
847 vdrop(lowervp);
848 } else {
849 VI_UNLOCK(vp);
850 *(ap->a_mpp) = NULL;
851 }
852 return (0);
853}
854
855static int
856null_vptofh(struct vop_vptofh_args *ap)
857{
858 struct vnode *lvp;
859
860 lvp = NULLVPTOLOWERVP(ap->a_vp);
861 return VOP_VPTOFH(lvp, ap->a_fhp);
862}
863
864static int
865null_vptocnp(struct vop_vptocnp_args *ap)
866{
867 struct vnode *vp = ap->a_vp;
868 struct vnode **dvp = ap->a_vpp;
869 struct vnode *lvp, *ldvp;
870 struct ucred *cred = ap->a_cred;
871 int error, locked;
872
873 if (vp->v_type == VDIR)
874 return (vop_stdvptocnp(ap));
875
876 locked = VOP_ISLOCKED(vp);
877 lvp = NULLVPTOLOWERVP(vp);
878 vhold(lvp);
879 VOP_UNLOCK(vp, 0); /* vp is held by vn_vptocnp_locked that called us */
880 ldvp = lvp;
881 vref(lvp);
882 error = vn_vptocnp(&ldvp, cred, ap->a_buf, ap->a_buflen);
883 vdrop(lvp);
884 if (error != 0) {
885 vn_lock(vp, locked | LK_RETRY);
886 return (ENOENT);
887 }
888
889 /*
890 * Exclusive lock is required by insmntque1 call in
891 * null_nodeget()
892 */
893 error = vn_lock(ldvp, LK_EXCLUSIVE);
894 if (error != 0) {
895 vrele(ldvp);
896 vn_lock(vp, locked | LK_RETRY);
897 return (ENOENT);
898 }
899 vref(ldvp);
900 error = null_nodeget(vp->v_mount, ldvp, dvp);
901 if (error == 0) {
902#ifdef DIAGNOSTIC
903 NULLVPTOLOWERVP(*dvp);
904#endif
905 VOP_UNLOCK(*dvp, 0); /* keep reference on *dvp */
906 }
907 vn_lock(vp, locked | LK_RETRY);
908 return (error);
909}
910
911/*
912 * Global vfs data structures
913 */
914struct vop_vector null_vnodeops = {
915 .vop_bypass = null_bypass,
916 .vop_access = null_access,
917 .vop_accessx = null_accessx,
918 .vop_advlockpurge = vop_stdadvlockpurge,
919 .vop_bmap = VOP_EOPNOTSUPP,
920 .vop_getattr = null_getattr,
921 .vop_getwritemount = null_getwritemount,
922 .vop_inactive = null_inactive,
923 .vop_islocked = vop_stdislocked,
924 .vop_lock1 = null_lock,
925 .vop_lookup = null_lookup,
926 .vop_open = null_open,
927 .vop_print = null_print,
928 .vop_reclaim = null_reclaim,
929 .vop_remove = null_remove,
930 .vop_rename = null_rename,
931 .vop_rmdir = null_rmdir,
932 .vop_setattr = null_setattr,
933 .vop_strategy = VOP_EOPNOTSUPP,
934 .vop_unlock = null_unlock,
935 .vop_vptocnp = null_vptocnp,
936 .vop_vptofh = null_vptofh,
937 .vop_add_writecount = null_add_writecount,
938};
|