44 */
45
46/*
47 * Null Layer
48 *
49 * (See mount_nullfs(8) for more information.)
50 *
51 * The null layer duplicates a portion of the filesystem
52 * name space under a new name. In this respect, it is
53 * similar to the loopback filesystem. It differs from
54 * the loopback fs in two respects: it is implemented using
55 * a stackable layers techniques, and its "null-node"s stack above
56 * all lower-layer vnodes, not just over directory vnodes.
57 *
58 * The null layer has two purposes. First, it serves as a demonstration
59 * of layering by proving a layer which does nothing. (It actually
60 * does everything the loopback filesystem does, which is slightly
61 * more than nothing.) Second, the null layer can serve as a prototype
62 * layer. Since it provides all necessary layer framework,
63 * new filesystem layers can be created very easily be starting
64 * with a null layer.
65 *
66 * The remainder of this man page examines the null layer as a basis
67 * for constructing new layers.
68 *
69 *
70 * INSTANTIATING NEW NULL LAYERS
71 *
72 * New null layers are created with mount_nullfs(8).
73 * Mount_nullfs(8) takes two arguments, the pathname
74 * of the lower vfs (target-pn) and the pathname where the null
75 * layer will appear in the namespace (alias-pn). After
76 * the null layer is put into place, the contents
77 * of target-pn subtree will be aliased under alias-pn.
78 *
79 *
80 * OPERATION OF A NULL LAYER
81 *
82 * The null layer is the minimum filesystem layer,
83 * simply bypassing all possible operations to the lower layer
84 * for processing there. The majority of its activity centers
85 * on the bypass routine, through which nearly all vnode operations
86 * pass.
87 *
88 * The bypass routine accepts arbitrary vnode operations for
89 * handling by the lower layer. It begins by examing vnode
90 * operation arguments and replacing any null-nodes by their
91 * lower-layer equivlants. It then invokes the operation
92 * on the lower layer. Finally, it replaces the null-nodes
93 * in the arguments and, if a vnode is return by the operation,
94 * stacks a null-node on top of the returned vnode.
95 *
96 * Although bypass handles most operations, vop_getattr, vop_lock,
97 * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not
98 * bypassed. Vop_getattr must change the fsid being returned.
99 * Vop_lock and vop_unlock must handle any locking for the
100 * current vnode as well as pass the lock request down.
101 * Vop_inactive and vop_reclaim are not bypassed so that
102 * they can handle freeing null-layer specific data. Vop_print
103 * is not bypassed to avoid excessive debugging information.
104 * Also, certain vnode operations change the locking state within
105 * the operation (create, mknod, remove, link, rename, mkdir, rmdir,
106 * and symlink). Ideally these operations should not change the
107 * lock state, but should be changed to let the caller of the
108 * function unlock them. Otherwise all intermediate vnode layers
109 * (such as union, umapfs, etc) must catch these functions to do
110 * the necessary locking at their layer.
111 *
112 *
113 * INSTANTIATING VNODE STACKS
114 *
115 * Mounting associates the null layer with a lower layer,
116 * effect stacking two VFSes. Vnode stacks are instead
117 * created on demand as files are accessed.
118 *
119 * The initial mount creates a single vnode stack for the
120 * root of the new null layer. All other vnode stacks
121 * are created as a result of vnode operations on
122 * this or other null vnode stacks.
123 *
124 * New vnode stacks come into existance as a result of
125 * an operation which returns a vnode.
126 * The bypass routine stacks a null-node above the new
127 * vnode before returning it to the caller.
128 *
129 * For example, imagine mounting a null layer with
130 * "mount_nullfs /usr/include /dev/layer/null".
131 * Changing directory to /dev/layer/null will assign
132 * the root null-node (which was created when the null layer was mounted).
133 * Now consider opening "sys". A vop_lookup would be
134 * done on the root null-node. This operation would bypass through
135 * to the lower layer which would return a vnode representing
136 * the UFS "sys". Null_bypass then builds a null-node
137 * aliasing the UFS "sys" and returns this to the caller.
138 * Later operations on the null-node "sys" will repeat this
139 * process when constructing other vnode stacks.
140 *
141 *
142 * CREATING OTHER FILE SYSTEM LAYERS
143 *
144 * One of the easiest ways to construct new filesystem layers is to make
145 * a copy of the null layer, rename all files and variables, and
146 * then begin modifing the copy. Sed can be used to easily rename
147 * all variables.
148 *
149 * The umap layer is an example of a layer descended from the
150 * null layer.
151 *
152 *
153 * INVOKING OPERATIONS ON LOWER LAYERS
154 *
155 * There are two techniques to invoke operations on a lower layer
156 * when the operation cannot be completely bypassed. Each method
157 * is appropriate in different situations. In both cases,
158 * it is the responsibility of the aliasing layer to make
159 * the operation arguments "correct" for the lower layer
| 44 */
45
46/*
47 * Null Layer
48 *
49 * (See mount_nullfs(8) for more information.)
50 *
51 * The null layer duplicates a portion of the filesystem
52 * name space under a new name. In this respect, it is
53 * similar to the loopback filesystem. It differs from
54 * the loopback fs in two respects: it is implemented using
55 * a stackable layers techniques, and its "null-node"s stack above
56 * all lower-layer vnodes, not just over directory vnodes.
57 *
58 * The null layer has two purposes. First, it serves as a demonstration
59 * of layering by proving a layer which does nothing. (It actually
60 * does everything the loopback filesystem does, which is slightly
61 * more than nothing.) Second, the null layer can serve as a prototype
62 * layer. Since it provides all necessary layer framework,
63 * new filesystem layers can be created very easily be starting
64 * with a null layer.
65 *
66 * The remainder of this man page examines the null layer as a basis
67 * for constructing new layers.
68 *
69 *
70 * INSTANTIATING NEW NULL LAYERS
71 *
72 * New null layers are created with mount_nullfs(8).
73 * Mount_nullfs(8) takes two arguments, the pathname
74 * of the lower vfs (target-pn) and the pathname where the null
75 * layer will appear in the namespace (alias-pn). After
76 * the null layer is put into place, the contents
77 * of target-pn subtree will be aliased under alias-pn.
78 *
79 *
80 * OPERATION OF A NULL LAYER
81 *
82 * The null layer is the minimum filesystem layer,
83 * simply bypassing all possible operations to the lower layer
84 * for processing there. The majority of its activity centers
85 * on the bypass routine, through which nearly all vnode operations
86 * pass.
87 *
88 * The bypass routine accepts arbitrary vnode operations for
89 * handling by the lower layer. It begins by examing vnode
90 * operation arguments and replacing any null-nodes by their
91 * lower-layer equivlants. It then invokes the operation
92 * on the lower layer. Finally, it replaces the null-nodes
93 * in the arguments and, if a vnode is return by the operation,
94 * stacks a null-node on top of the returned vnode.
95 *
96 * Although bypass handles most operations, vop_getattr, vop_lock,
97 * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not
98 * bypassed. Vop_getattr must change the fsid being returned.
99 * Vop_lock and vop_unlock must handle any locking for the
100 * current vnode as well as pass the lock request down.
101 * Vop_inactive and vop_reclaim are not bypassed so that
102 * they can handle freeing null-layer specific data. Vop_print
103 * is not bypassed to avoid excessive debugging information.
104 * Also, certain vnode operations change the locking state within
105 * the operation (create, mknod, remove, link, rename, mkdir, rmdir,
106 * and symlink). Ideally these operations should not change the
107 * lock state, but should be changed to let the caller of the
108 * function unlock them. Otherwise all intermediate vnode layers
109 * (such as union, umapfs, etc) must catch these functions to do
110 * the necessary locking at their layer.
111 *
112 *
113 * INSTANTIATING VNODE STACKS
114 *
115 * Mounting associates the null layer with a lower layer,
116 * effect stacking two VFSes. Vnode stacks are instead
117 * created on demand as files are accessed.
118 *
119 * The initial mount creates a single vnode stack for the
120 * root of the new null layer. All other vnode stacks
121 * are created as a result of vnode operations on
122 * this or other null vnode stacks.
123 *
124 * New vnode stacks come into existance as a result of
125 * an operation which returns a vnode.
126 * The bypass routine stacks a null-node above the new
127 * vnode before returning it to the caller.
128 *
129 * For example, imagine mounting a null layer with
130 * "mount_nullfs /usr/include /dev/layer/null".
131 * Changing directory to /dev/layer/null will assign
132 * the root null-node (which was created when the null layer was mounted).
133 * Now consider opening "sys". A vop_lookup would be
134 * done on the root null-node. This operation would bypass through
135 * to the lower layer which would return a vnode representing
136 * the UFS "sys". Null_bypass then builds a null-node
137 * aliasing the UFS "sys" and returns this to the caller.
138 * Later operations on the null-node "sys" will repeat this
139 * process when constructing other vnode stacks.
140 *
141 *
142 * CREATING OTHER FILE SYSTEM LAYERS
143 *
144 * One of the easiest ways to construct new filesystem layers is to make
145 * a copy of the null layer, rename all files and variables, and
146 * then begin modifing the copy. Sed can be used to easily rename
147 * all variables.
148 *
149 * The umap layer is an example of a layer descended from the
150 * null layer.
151 *
152 *
153 * INVOKING OPERATIONS ON LOWER LAYERS
154 *
155 * There are two techniques to invoke operations on a lower layer
156 * when the operation cannot be completely bypassed. Each method
157 * is appropriate in different situations. In both cases,
158 * it is the responsibility of the aliasing layer to make
159 * the operation arguments "correct" for the lower layer
|
161 *
162 * The first approach is to call the aliasing layer's bypass routine.
163 * This method is most suitable when you wish to invoke the operation
164 * currently being handled on the lower layer. It has the advantage
165 * that the bypass routine already must do argument mapping.
166 * An example of this is null_getattrs in the null layer.
167 *
168 * A second approach is to directly invoke vnode operations on
169 * the lower layer with the VOP_OPERATIONNAME interface.
170 * The advantage of this method is that it is easy to invoke
171 * arbitrary operations on the lower layer. The disadvantage
172 * is that vnode arguments must be manualy mapped.
173 *
174 */
175
176#include <sys/param.h>
177#include <sys/systm.h>
178#include <sys/conf.h>
179#include <sys/kernel.h>
180#include <sys/lock.h>
181#include <sys/malloc.h>
182#include <sys/mount.h>
183#include <sys/mutex.h>
184#include <sys/namei.h>
185#include <sys/sysctl.h>
186#include <sys/vnode.h>
187
188#include <fs/nullfs/null.h>
189
190#include <vm/vm.h>
191#include <vm/vm_extern.h>
192#include <vm/vm_object.h>
193#include <vm/vnode_pager.h>
194
195static int null_bug_bypass = 0; /* for debugging: enables bypass printf'ing */
196SYSCTL_INT(_debug, OID_AUTO, nullfs_bug_bypass, CTLFLAG_RW,
197 &null_bug_bypass, 0, "");
198
199static int null_access(struct vop_access_args *ap);
200static int null_createvobject(struct vop_createvobject_args *ap);
201static int null_destroyvobject(struct vop_destroyvobject_args *ap);
202static int null_getattr(struct vop_getattr_args *ap);
203static int null_getvobject(struct vop_getvobject_args *ap);
204static int null_inactive(struct vop_inactive_args *ap);
205static int null_islocked(struct vop_islocked_args *ap);
206static int null_lock(struct vop_lock_args *ap);
207static int null_lookup(struct vop_lookup_args *ap);
208static int null_open(struct vop_open_args *ap);
209static int null_print(struct vop_print_args *ap);
210static int null_reclaim(struct vop_reclaim_args *ap);
211static int null_rename(struct vop_rename_args *ap);
212static int null_setattr(struct vop_setattr_args *ap);
213static int null_unlock(struct vop_unlock_args *ap);
214
215/*
216 * This is the 10-Apr-92 bypass routine.
217 * This version has been optimized for speed, throwing away some
218 * safety checks. It should still always work, but it's not as
219 * robust to programmer errors.
220 *
221 * In general, we map all vnodes going down and unmap them on the way back.
222 * As an exception to this, vnodes can be marked "unmapped" by setting
223 * the Nth bit in operation's vdesc_flags.
224 *
225 * Also, some BSD vnode operations have the side effect of vrele'ing
226 * their arguments. With stacking, the reference counts are held
227 * by the upper node, not the lower one, so we must handle these
228 * side-effects here. This is not of concern in Sun-derived systems
229 * since there are no such side-effects.
230 *
231 * This makes the following assumptions:
232 * - only one returned vpp
233 * - no INOUT vpp's (Sun's vop_open has one of these)
234 * - the vnode operation vector of the first vnode should be used
235 * to determine what implementation of the op should be invoked
236 * - all mapped vnodes are of our vnode-type (NEEDSWORK:
237 * problems on rmdir'ing mount points and renaming?)
238 */
239int
240null_bypass(ap)
241 struct vop_generic_args /* {
242 struct vnodeop_desc *a_desc;
243 <other random data follows, presumably>
244 } */ *ap;
245{
246 register struct vnode **this_vp_p;
247 int error;
248 struct vnode *old_vps[VDESC_MAX_VPS];
249 struct vnode **vps_p[VDESC_MAX_VPS];
250 struct vnode ***vppp;
251 struct vnodeop_desc *descp = ap->a_desc;
252 int reles, i;
253
254 if (null_bug_bypass)
255 printf ("null_bypass: %s\n", descp->vdesc_name);
256
257#ifdef DIAGNOSTIC
258 /*
259 * We require at least one vp.
260 */
261 if (descp->vdesc_vp_offsets == NULL ||
262 descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET)
263 panic ("null_bypass: no vp's in map");
264#endif
265
266 /*
267 * Map the vnodes going in.
268 * Later, we'll invoke the operation based on
269 * the first mapped vnode's operation vector.
270 */
271 reles = descp->vdesc_flags;
272 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
273 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
274 break; /* bail out at end of list */
275 vps_p[i] = this_vp_p =
276 VOPARG_OFFSETTO(struct vnode**,descp->vdesc_vp_offsets[i],ap);
277 /*
278 * We're not guaranteed that any but the first vnode
279 * are of our type. Check for and don't map any
280 * that aren't. (We must always map first vp or vclean fails.)
281 */
282 if (i && (*this_vp_p == NULLVP ||
283 (*this_vp_p)->v_op != null_vnodeop_p)) {
284 old_vps[i] = NULLVP;
285 } else {
286 old_vps[i] = *this_vp_p;
287 *(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p);
288 /*
289 * XXX - Several operations have the side effect
290 * of vrele'ing their vp's. We must account for
291 * that. (This should go away in the future.)
292 */
293 if (reles & VDESC_VP0_WILLRELE)
294 VREF(*this_vp_p);
295 }
296
297 }
298
299 /*
300 * Call the operation on the lower layer
301 * with the modified argument structure.
302 */
303 if (vps_p[0] && *vps_p[0])
304 error = VCALL(*(vps_p[0]), descp->vdesc_offset, ap);
305 else {
306 printf("null_bypass: no map for %s\n", descp->vdesc_name);
307 error = EINVAL;
308 }
309
310 /*
311 * Maintain the illusion of call-by-value
312 * by restoring vnodes in the argument structure
313 * to their original value.
314 */
315 reles = descp->vdesc_flags;
316 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
317 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
318 break; /* bail out at end of list */
319 if (old_vps[i]) {
320 *(vps_p[i]) = old_vps[i];
321#if 0
322 if (reles & VDESC_VP0_WILLUNLOCK)
323 VOP_UNLOCK(*(vps_p[i]), LK_THISLAYER, curthread);
324#endif
325 if (reles & VDESC_VP0_WILLRELE)
326 vrele(*(vps_p[i]));
327 }
328 }
329
330 /*
331 * Map the possible out-going vpp
332 * (Assumes that the lower layer always returns
333 * a VREF'ed vpp unless it gets an error.)
334 */
335 if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET &&
336 !(descp->vdesc_flags & VDESC_NOMAP_VPP) &&
337 !error) {
338 /*
339 * XXX - even though some ops have vpp returned vp's,
340 * several ops actually vrele this before returning.
341 * We must avoid these ops.
342 * (This should go away when these ops are regularized.)
343 */
344 if (descp->vdesc_flags & VDESC_VPP_WILLRELE)
345 goto out;
346 vppp = VOPARG_OFFSETTO(struct vnode***,
347 descp->vdesc_vpp_offset,ap);
348 if (*vppp)
349 error = null_nodeget(old_vps[0]->v_mount, **vppp, *vppp);
350 }
351
352 out:
353 return (error);
354}
355
356/*
357 * We have to carry on the locking protocol on the null layer vnodes
358 * as we progress through the tree. We also have to enforce read-only
359 * if this layer is mounted read-only.
360 */
361static int
362null_lookup(ap)
363 struct vop_lookup_args /* {
364 struct vnode * a_dvp;
365 struct vnode ** a_vpp;
366 struct componentname * a_cnp;
367 } */ *ap;
368{
369 struct componentname *cnp = ap->a_cnp;
370 struct vnode *dvp = ap->a_dvp;
371 struct thread *td = cnp->cn_thread;
372 int flags = cnp->cn_flags;
373 struct vnode *vp, *ldvp, *lvp;
374 int error;
375
376 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
377 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
378 return (EROFS);
379 /*
380 * Although it is possible to call null_bypass(), we'll do
381 * a direct call to reduce overhead
382 */
383 ldvp = NULLVPTOLOWERVP(dvp);
384 vp = lvp = NULL;
385 error = VOP_LOOKUP(ldvp, &lvp, cnp);
386 if (error == EJUSTRETURN && (flags & ISLASTCN) &&
387 (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
388 (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME))
389 error = EROFS;
390
391 /*
392 * Rely only on the PDIRUNLOCK flag which should be carefully
393 * tracked by underlying filesystem.
394 */
395 if (cnp->cn_flags & PDIRUNLOCK)
396 VOP_UNLOCK(dvp, LK_THISLAYER, td);
397 if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) {
398 if (ldvp == lvp) {
399 *ap->a_vpp = dvp;
400 VREF(dvp);
401 vrele(lvp);
402 } else {
403 error = null_nodeget(dvp->v_mount, lvp, &vp);
404 if (error) {
405 /* XXX Cleanup needed... */
406 panic("null_nodeget failed");
407 }
408 *ap->a_vpp = vp;
409 }
410 }
411 return (error);
412}
413
414/*
415 * Setattr call. Disallow write attempts if the layer is mounted read-only.
416 */
417static int
418null_setattr(ap)
419 struct vop_setattr_args /* {
420 struct vnodeop_desc *a_desc;
421 struct vnode *a_vp;
422 struct vattr *a_vap;
423 struct ucred *a_cred;
424 struct thread *a_td;
425 } */ *ap;
426{
427 struct vnode *vp = ap->a_vp;
428 struct vattr *vap = ap->a_vap;
429
430 if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL ||
431 vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL ||
432 vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) &&
433 (vp->v_mount->mnt_flag & MNT_RDONLY))
434 return (EROFS);
435 if (vap->va_size != VNOVAL) {
436 switch (vp->v_type) {
437 case VDIR:
438 return (EISDIR);
439 case VCHR:
440 case VBLK:
441 case VSOCK:
442 case VFIFO:
443 if (vap->va_flags != VNOVAL)
444 return (EOPNOTSUPP);
445 return (0);
446 case VREG:
447 case VLNK:
448 default:
449 /*
450 * Disallow write attempts if the filesystem is
451 * mounted read-only.
452 */
453 if (vp->v_mount->mnt_flag & MNT_RDONLY)
454 return (EROFS);
455 }
456 }
457
458 return (null_bypass((struct vop_generic_args *)ap));
459}
460
461/*
462 * We handle getattr only to change the fsid.
463 */
464static int
465null_getattr(ap)
466 struct vop_getattr_args /* {
467 struct vnode *a_vp;
468 struct vattr *a_vap;
469 struct ucred *a_cred;
470 struct thread *a_td;
471 } */ *ap;
472{
473 int error;
474
475 if ((error = null_bypass((struct vop_generic_args *)ap)) != 0)
476 return (error);
477
478 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
479 return (0);
480}
481
482/*
483 * Handle to disallow write access if mounted read-only.
484 */
485static int
486null_access(ap)
487 struct vop_access_args /* {
488 struct vnode *a_vp;
489 int a_mode;
490 struct ucred *a_cred;
491 struct thread *a_td;
492 } */ *ap;
493{
494 struct vnode *vp = ap->a_vp;
495 mode_t mode = ap->a_mode;
496
497 /*
498 * Disallow write attempts on read-only layers;
499 * unless the file is a socket, fifo, or a block or
500 * character device resident on the filesystem.
501 */
502 if (mode & VWRITE) {
503 switch (vp->v_type) {
504 case VDIR:
505 case VLNK:
506 case VREG:
507 if (vp->v_mount->mnt_flag & MNT_RDONLY)
508 return (EROFS);
509 break;
510 default:
511 break;
512 }
513 }
514 return (null_bypass((struct vop_generic_args *)ap));
515}
516
517/*
518 * We must handle open to be able to catch MNT_NODEV and friends.
519 */
520static int
521null_open(ap)
522 struct vop_open_args /* {
523 struct vnode *a_vp;
524 int a_mode;
525 struct ucred *a_cred;
526 struct thread *a_td;
527 } */ *ap;
528{
529 struct vnode *vp = ap->a_vp;
530 struct vnode *lvp = NULLVPTOLOWERVP(ap->a_vp);
531
532 if ((vp->v_mount->mnt_flag & MNT_NODEV) &&
533 (lvp->v_type == VBLK || lvp->v_type == VCHR))
534 return ENXIO;
535
536 return (null_bypass((struct vop_generic_args *)ap));
537}
538
539/*
540 * We handle this to eliminate null FS to lower FS
541 * file moving. Don't know why we don't allow this,
542 * possibly we should.
543 */
544static int
545null_rename(ap)
546 struct vop_rename_args /* {
547 struct vnode *a_fdvp;
548 struct vnode *a_fvp;
549 struct componentname *a_fcnp;
550 struct vnode *a_tdvp;
551 struct vnode *a_tvp;
552 struct componentname *a_tcnp;
553 } */ *ap;
554{
555 struct vnode *tdvp = ap->a_tdvp;
556 struct vnode *fvp = ap->a_fvp;
557 struct vnode *fdvp = ap->a_fdvp;
558 struct vnode *tvp = ap->a_tvp;
559
560 /* Check for cross-device rename. */
561 if ((fvp->v_mount != tdvp->v_mount) ||
562 (tvp && (fvp->v_mount != tvp->v_mount))) {
563 if (tdvp == tvp)
564 vrele(tdvp);
565 else
566 vput(tdvp);
567 if (tvp)
568 vput(tvp);
569 vrele(fdvp);
570 vrele(fvp);
571 return (EXDEV);
572 }
573
574 return (null_bypass((struct vop_generic_args *)ap));
575}
576
577/*
578 * We need to process our own vnode lock and then clear the
579 * interlock flag as it applies only to our vnode, not the
580 * vnodes below us on the stack.
581 */
582static int
583null_lock(ap)
584 struct vop_lock_args /* {
585 struct vnode *a_vp;
586 int a_flags;
587 struct thread *a_td;
588 } */ *ap;
589{
590 struct vnode *vp = ap->a_vp;
591 int flags = ap->a_flags;
592 struct thread *td = ap->a_td;
593 struct vnode *lvp;
594 int error;
595
596 if (flags & LK_THISLAYER) {
597 if (vp->v_vnlock != NULL) {
598 /* lock is shared across layers */
599 if (flags & LK_INTERLOCK)
600 mtx_unlock(&vp->v_interlock);
601 return 0;
602 }
603 error = lockmgr(&vp->v_lock, flags & ~LK_THISLAYER,
604 &vp->v_interlock, td);
605 return (error);
606 }
607
608 if (vp->v_vnlock != NULL) {
609 /*
610 * The lower level has exported a struct lock to us. Use
611 * it so that all vnodes in the stack lock and unlock
612 * simultaneously. Note: we don't DRAIN the lock as DRAIN
613 * decommissions the lock - just because our vnode is
614 * going away doesn't mean the struct lock below us is.
615 * LK_EXCLUSIVE is fine.
616 */
617 if ((flags & LK_TYPE_MASK) == LK_DRAIN) {
618 NULLFSDEBUG("null_lock: avoiding LK_DRAIN\n");
619 return(lockmgr(vp->v_vnlock,
620 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE,
621 &vp->v_interlock, td));
622 }
623 return(lockmgr(vp->v_vnlock, flags, &vp->v_interlock, td));
624 } else {
625 /*
626 * To prevent race conditions involving doing a lookup
627 * on "..", we have to lock the lower node, then lock our
628 * node. Most of the time it won't matter that we lock our
629 * node (as any locking would need the lower one locked
630 * first). But we can LK_DRAIN the upper lock as a step
631 * towards decomissioning it.
632 */
633 lvp = NULLVPTOLOWERVP(vp);
634 if (lvp == NULL)
635 return (lockmgr(&vp->v_lock, flags, &vp->v_interlock, td));
636 if (flags & LK_INTERLOCK) {
637 mtx_unlock(&vp->v_interlock);
638 flags &= ~LK_INTERLOCK;
639 }
640 if ((flags & LK_TYPE_MASK) == LK_DRAIN) {
641 error = VOP_LOCK(lvp,
642 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, td);
643 } else
644 error = VOP_LOCK(lvp, flags, td);
645 if (error)
646 return (error);
647 error = lockmgr(&vp->v_lock, flags, &vp->v_interlock, td);
648 if (error)
649 VOP_UNLOCK(lvp, 0, td);
650 return (error);
651 }
652}
653
654/*
655 * We need to process our own vnode unlock and then clear the
656 * interlock flag as it applies only to our vnode, not the
657 * vnodes below us on the stack.
658 */
659static int
660null_unlock(ap)
661 struct vop_unlock_args /* {
662 struct vnode *a_vp;
663 int a_flags;
664 struct thread *a_td;
665 } */ *ap;
666{
667 struct vnode *vp = ap->a_vp;
668 int flags = ap->a_flags;
669 struct thread *td = ap->a_td;
670 struct vnode *lvp;
671
672 if (vp->v_vnlock != NULL) {
673 if (flags & LK_THISLAYER)
674 return 0; /* the lock is shared across layers */
675 flags &= ~LK_THISLAYER;
676 return (lockmgr(vp->v_vnlock, flags | LK_RELEASE,
677 &vp->v_interlock, td));
678 }
679 lvp = NULLVPTOLOWERVP(vp);
680 if (lvp == NULL)
681 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td));
682 if ((flags & LK_THISLAYER) == 0) {
683 if (flags & LK_INTERLOCK) {
684 mtx_unlock(&vp->v_interlock);
685 flags &= ~LK_INTERLOCK;
686 }
687 VOP_UNLOCK(lvp, flags & ~LK_INTERLOCK, td);
688 } else
689 flags &= ~LK_THISLAYER;
690 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td));
691}
692
693static int
694null_islocked(ap)
695 struct vop_islocked_args /* {
696 struct vnode *a_vp;
697 struct thread *a_td;
698 } */ *ap;
699{
700 struct vnode *vp = ap->a_vp;
701 struct thread *td = ap->a_td;
702
703 if (vp->v_vnlock != NULL)
704 return (lockstatus(vp->v_vnlock, td));
705 return (lockstatus(&vp->v_lock, td));
706}
707
708/*
709 * There is no way to tell that someone issued remove/rmdir operation
710 * on the underlying filesystem. For now we just have to release lowevrp
711 * as soon as possible.
712 *
713 * Note, we can't release any resources nor remove vnode from hash before
714 * appropriate VXLOCK stuff is is done because other process can find this
715 * vnode in hash during inactivation and may be sitting in vget() and waiting
716 * for null_inactive to unlock vnode. Thus we will do all those in VOP_RECLAIM.
717 */
718static int
719null_inactive(ap)
720 struct vop_inactive_args /* {
721 struct vnode *a_vp;
722 struct thread *a_td;
723 } */ *ap;
724{
725 struct vnode *vp = ap->a_vp;
726 struct thread *td = ap->a_td;
727
728 VOP_UNLOCK(vp, 0, td);
729
730 /*
731 * If this is the last reference, then free up the vnode
732 * so as not to tie up the lower vnodes.
733 */
734 vrecycle(vp, NULL, td);
735
736 return (0);
737}
738
739/*
740 * Now, the VXLOCK is in force and we're free to destroy the null vnode.
741 */
742static int
743null_reclaim(ap)
744 struct vop_reclaim_args /* {
745 struct vnode *a_vp;
746 struct thread *a_td;
747 } */ *ap;
748{
749 struct vnode *vp = ap->a_vp;
750 struct null_node *xp = VTONULL(vp);
751 struct vnode *lowervp = xp->null_lowervp;
752
753 if (lowervp) {
754 null_hashrem(xp);
755
756 vrele(lowervp);
757 vrele(lowervp);
758 }
759
760 vp->v_data = NULL;
761 vp->v_vnlock = &vp->v_lock;
762 FREE(xp, M_NULLFSNODE);
763
764 return (0);
765}
766
767static int
768null_print(ap)
769 struct vop_print_args /* {
770 struct vnode *a_vp;
771 } */ *ap;
772{
773 register struct vnode *vp = ap->a_vp;
774 printf("\ttag %s, vp=%p, lowervp=%p\n", vp->v_tag, vp,
775 NULLVPTOLOWERVP(vp));
776 return (0);
777}
778
779/*
780 * Let an underlying filesystem do the work
781 */
782static int
783null_createvobject(ap)
784 struct vop_createvobject_args /* {
785 struct vnode *vp;
786 struct ucred *cred;
787 struct thread *td;
788 } */ *ap;
789{
790 struct vnode *vp = ap->a_vp;
791 struct vnode *lowervp = VTONULL(vp) ? NULLVPTOLOWERVP(vp) : NULL;
792 int error;
793
794 if (vp->v_type == VNON || lowervp == NULL)
795 return 0;
796 error = VOP_CREATEVOBJECT(lowervp, ap->a_cred, ap->a_td);
797 if (error)
798 return (error);
799 vp->v_vflag |= VV_OBJBUF;
800 return (0);
801}
802
803/*
804 * We have nothing to destroy and this operation shouldn't be bypassed.
805 */
806static int
807null_destroyvobject(ap)
808 struct vop_destroyvobject_args /* {
809 struct vnode *vp;
810 } */ *ap;
811{
812 struct vnode *vp = ap->a_vp;
813
814 vp->v_vflag &= ~VV_OBJBUF;
815 return (0);
816}
817
818static int
819null_getvobject(ap)
820 struct vop_getvobject_args /* {
821 struct vnode *vp;
822 struct vm_object **objpp;
823 } */ *ap;
824{
825 struct vnode *lvp = NULLVPTOLOWERVP(ap->a_vp);
826
827 if (lvp == NULL)
828 return EINVAL;
829 return (VOP_GETVOBJECT(lvp, ap->a_objpp));
830}
831
832/*
833 * Global vfs data structures
834 */
835vop_t **null_vnodeop_p;
836static struct vnodeopv_entry_desc null_vnodeop_entries[] = {
837 { &vop_default_desc, (vop_t *) null_bypass },
838
839 { &vop_access_desc, (vop_t *) null_access },
840 { &vop_bmap_desc, (vop_t *) vop_eopnotsupp },
841 { &vop_createvobject_desc, (vop_t *) null_createvobject },
842 { &vop_destroyvobject_desc, (vop_t *) null_destroyvobject },
843 { &vop_getattr_desc, (vop_t *) null_getattr },
844 { &vop_getvobject_desc, (vop_t *) null_getvobject },
845 { &vop_getwritemount_desc, (vop_t *) vop_stdgetwritemount},
846 { &vop_inactive_desc, (vop_t *) null_inactive },
847 { &vop_islocked_desc, (vop_t *) null_islocked },
848 { &vop_lock_desc, (vop_t *) null_lock },
849 { &vop_lookup_desc, (vop_t *) null_lookup },
850 { &vop_open_desc, (vop_t *) null_open },
851 { &vop_print_desc, (vop_t *) null_print },
852 { &vop_reclaim_desc, (vop_t *) null_reclaim },
853 { &vop_rename_desc, (vop_t *) null_rename },
854 { &vop_setattr_desc, (vop_t *) null_setattr },
855 { &vop_strategy_desc, (vop_t *) vop_eopnotsupp },
856 { &vop_unlock_desc, (vop_t *) null_unlock },
857 { NULL, NULL }
858};
859static struct vnodeopv_desc null_vnodeop_opv_desc =
860 { &null_vnodeop_p, null_vnodeop_entries };
861
862VNODEOP_SET(null_vnodeop_opv_desc);
| 161 *
162 * The first approach is to call the aliasing layer's bypass routine.
163 * This method is most suitable when you wish to invoke the operation
164 * currently being handled on the lower layer. It has the advantage
165 * that the bypass routine already must do argument mapping.
166 * An example of this is null_getattrs in the null layer.
167 *
168 * A second approach is to directly invoke vnode operations on
169 * the lower layer with the VOP_OPERATIONNAME interface.
170 * The advantage of this method is that it is easy to invoke
171 * arbitrary operations on the lower layer. The disadvantage
172 * is that vnode arguments must be manualy mapped.
173 *
174 */
175
176#include <sys/param.h>
177#include <sys/systm.h>
178#include <sys/conf.h>
179#include <sys/kernel.h>
180#include <sys/lock.h>
181#include <sys/malloc.h>
182#include <sys/mount.h>
183#include <sys/mutex.h>
184#include <sys/namei.h>
185#include <sys/sysctl.h>
186#include <sys/vnode.h>
187
188#include <fs/nullfs/null.h>
189
190#include <vm/vm.h>
191#include <vm/vm_extern.h>
192#include <vm/vm_object.h>
193#include <vm/vnode_pager.h>
194
195static int null_bug_bypass = 0; /* for debugging: enables bypass printf'ing */
196SYSCTL_INT(_debug, OID_AUTO, nullfs_bug_bypass, CTLFLAG_RW,
197 &null_bug_bypass, 0, "");
198
199static int null_access(struct vop_access_args *ap);
200static int null_createvobject(struct vop_createvobject_args *ap);
201static int null_destroyvobject(struct vop_destroyvobject_args *ap);
202static int null_getattr(struct vop_getattr_args *ap);
203static int null_getvobject(struct vop_getvobject_args *ap);
204static int null_inactive(struct vop_inactive_args *ap);
205static int null_islocked(struct vop_islocked_args *ap);
206static int null_lock(struct vop_lock_args *ap);
207static int null_lookup(struct vop_lookup_args *ap);
208static int null_open(struct vop_open_args *ap);
209static int null_print(struct vop_print_args *ap);
210static int null_reclaim(struct vop_reclaim_args *ap);
211static int null_rename(struct vop_rename_args *ap);
212static int null_setattr(struct vop_setattr_args *ap);
213static int null_unlock(struct vop_unlock_args *ap);
214
215/*
216 * This is the 10-Apr-92 bypass routine.
217 * This version has been optimized for speed, throwing away some
218 * safety checks. It should still always work, but it's not as
219 * robust to programmer errors.
220 *
221 * In general, we map all vnodes going down and unmap them on the way back.
222 * As an exception to this, vnodes can be marked "unmapped" by setting
223 * the Nth bit in operation's vdesc_flags.
224 *
225 * Also, some BSD vnode operations have the side effect of vrele'ing
226 * their arguments. With stacking, the reference counts are held
227 * by the upper node, not the lower one, so we must handle these
228 * side-effects here. This is not of concern in Sun-derived systems
229 * since there are no such side-effects.
230 *
231 * This makes the following assumptions:
232 * - only one returned vpp
233 * - no INOUT vpp's (Sun's vop_open has one of these)
234 * - the vnode operation vector of the first vnode should be used
235 * to determine what implementation of the op should be invoked
236 * - all mapped vnodes are of our vnode-type (NEEDSWORK:
237 * problems on rmdir'ing mount points and renaming?)
238 */
239int
240null_bypass(ap)
241 struct vop_generic_args /* {
242 struct vnodeop_desc *a_desc;
243 <other random data follows, presumably>
244 } */ *ap;
245{
246 register struct vnode **this_vp_p;
247 int error;
248 struct vnode *old_vps[VDESC_MAX_VPS];
249 struct vnode **vps_p[VDESC_MAX_VPS];
250 struct vnode ***vppp;
251 struct vnodeop_desc *descp = ap->a_desc;
252 int reles, i;
253
254 if (null_bug_bypass)
255 printf ("null_bypass: %s\n", descp->vdesc_name);
256
257#ifdef DIAGNOSTIC
258 /*
259 * We require at least one vp.
260 */
261 if (descp->vdesc_vp_offsets == NULL ||
262 descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET)
263 panic ("null_bypass: no vp's in map");
264#endif
265
266 /*
267 * Map the vnodes going in.
268 * Later, we'll invoke the operation based on
269 * the first mapped vnode's operation vector.
270 */
271 reles = descp->vdesc_flags;
272 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
273 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
274 break; /* bail out at end of list */
275 vps_p[i] = this_vp_p =
276 VOPARG_OFFSETTO(struct vnode**,descp->vdesc_vp_offsets[i],ap);
277 /*
278 * We're not guaranteed that any but the first vnode
279 * are of our type. Check for and don't map any
280 * that aren't. (We must always map first vp or vclean fails.)
281 */
282 if (i && (*this_vp_p == NULLVP ||
283 (*this_vp_p)->v_op != null_vnodeop_p)) {
284 old_vps[i] = NULLVP;
285 } else {
286 old_vps[i] = *this_vp_p;
287 *(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p);
288 /*
289 * XXX - Several operations have the side effect
290 * of vrele'ing their vp's. We must account for
291 * that. (This should go away in the future.)
292 */
293 if (reles & VDESC_VP0_WILLRELE)
294 VREF(*this_vp_p);
295 }
296
297 }
298
299 /*
300 * Call the operation on the lower layer
301 * with the modified argument structure.
302 */
303 if (vps_p[0] && *vps_p[0])
304 error = VCALL(*(vps_p[0]), descp->vdesc_offset, ap);
305 else {
306 printf("null_bypass: no map for %s\n", descp->vdesc_name);
307 error = EINVAL;
308 }
309
310 /*
311 * Maintain the illusion of call-by-value
312 * by restoring vnodes in the argument structure
313 * to their original value.
314 */
315 reles = descp->vdesc_flags;
316 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
317 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
318 break; /* bail out at end of list */
319 if (old_vps[i]) {
320 *(vps_p[i]) = old_vps[i];
321#if 0
322 if (reles & VDESC_VP0_WILLUNLOCK)
323 VOP_UNLOCK(*(vps_p[i]), LK_THISLAYER, curthread);
324#endif
325 if (reles & VDESC_VP0_WILLRELE)
326 vrele(*(vps_p[i]));
327 }
328 }
329
330 /*
331 * Map the possible out-going vpp
332 * (Assumes that the lower layer always returns
333 * a VREF'ed vpp unless it gets an error.)
334 */
335 if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET &&
336 !(descp->vdesc_flags & VDESC_NOMAP_VPP) &&
337 !error) {
338 /*
339 * XXX - even though some ops have vpp returned vp's,
340 * several ops actually vrele this before returning.
341 * We must avoid these ops.
342 * (This should go away when these ops are regularized.)
343 */
344 if (descp->vdesc_flags & VDESC_VPP_WILLRELE)
345 goto out;
346 vppp = VOPARG_OFFSETTO(struct vnode***,
347 descp->vdesc_vpp_offset,ap);
348 if (*vppp)
349 error = null_nodeget(old_vps[0]->v_mount, **vppp, *vppp);
350 }
351
352 out:
353 return (error);
354}
355
356/*
357 * We have to carry on the locking protocol on the null layer vnodes
358 * as we progress through the tree. We also have to enforce read-only
359 * if this layer is mounted read-only.
360 */
361static int
362null_lookup(ap)
363 struct vop_lookup_args /* {
364 struct vnode * a_dvp;
365 struct vnode ** a_vpp;
366 struct componentname * a_cnp;
367 } */ *ap;
368{
369 struct componentname *cnp = ap->a_cnp;
370 struct vnode *dvp = ap->a_dvp;
371 struct thread *td = cnp->cn_thread;
372 int flags = cnp->cn_flags;
373 struct vnode *vp, *ldvp, *lvp;
374 int error;
375
376 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
377 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
378 return (EROFS);
379 /*
380 * Although it is possible to call null_bypass(), we'll do
381 * a direct call to reduce overhead
382 */
383 ldvp = NULLVPTOLOWERVP(dvp);
384 vp = lvp = NULL;
385 error = VOP_LOOKUP(ldvp, &lvp, cnp);
386 if (error == EJUSTRETURN && (flags & ISLASTCN) &&
387 (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
388 (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME))
389 error = EROFS;
390
391 /*
392 * Rely only on the PDIRUNLOCK flag which should be carefully
393 * tracked by underlying filesystem.
394 */
395 if (cnp->cn_flags & PDIRUNLOCK)
396 VOP_UNLOCK(dvp, LK_THISLAYER, td);
397 if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) {
398 if (ldvp == lvp) {
399 *ap->a_vpp = dvp;
400 VREF(dvp);
401 vrele(lvp);
402 } else {
403 error = null_nodeget(dvp->v_mount, lvp, &vp);
404 if (error) {
405 /* XXX Cleanup needed... */
406 panic("null_nodeget failed");
407 }
408 *ap->a_vpp = vp;
409 }
410 }
411 return (error);
412}
413
414/*
415 * Setattr call. Disallow write attempts if the layer is mounted read-only.
416 */
417static int
418null_setattr(ap)
419 struct vop_setattr_args /* {
420 struct vnodeop_desc *a_desc;
421 struct vnode *a_vp;
422 struct vattr *a_vap;
423 struct ucred *a_cred;
424 struct thread *a_td;
425 } */ *ap;
426{
427 struct vnode *vp = ap->a_vp;
428 struct vattr *vap = ap->a_vap;
429
430 if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL ||
431 vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL ||
432 vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) &&
433 (vp->v_mount->mnt_flag & MNT_RDONLY))
434 return (EROFS);
435 if (vap->va_size != VNOVAL) {
436 switch (vp->v_type) {
437 case VDIR:
438 return (EISDIR);
439 case VCHR:
440 case VBLK:
441 case VSOCK:
442 case VFIFO:
443 if (vap->va_flags != VNOVAL)
444 return (EOPNOTSUPP);
445 return (0);
446 case VREG:
447 case VLNK:
448 default:
449 /*
450 * Disallow write attempts if the filesystem is
451 * mounted read-only.
452 */
453 if (vp->v_mount->mnt_flag & MNT_RDONLY)
454 return (EROFS);
455 }
456 }
457
458 return (null_bypass((struct vop_generic_args *)ap));
459}
460
461/*
462 * We handle getattr only to change the fsid.
463 */
464static int
465null_getattr(ap)
466 struct vop_getattr_args /* {
467 struct vnode *a_vp;
468 struct vattr *a_vap;
469 struct ucred *a_cred;
470 struct thread *a_td;
471 } */ *ap;
472{
473 int error;
474
475 if ((error = null_bypass((struct vop_generic_args *)ap)) != 0)
476 return (error);
477
478 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
479 return (0);
480}
481
482/*
483 * Handle to disallow write access if mounted read-only.
484 */
485static int
486null_access(ap)
487 struct vop_access_args /* {
488 struct vnode *a_vp;
489 int a_mode;
490 struct ucred *a_cred;
491 struct thread *a_td;
492 } */ *ap;
493{
494 struct vnode *vp = ap->a_vp;
495 mode_t mode = ap->a_mode;
496
497 /*
498 * Disallow write attempts on read-only layers;
499 * unless the file is a socket, fifo, or a block or
500 * character device resident on the filesystem.
501 */
502 if (mode & VWRITE) {
503 switch (vp->v_type) {
504 case VDIR:
505 case VLNK:
506 case VREG:
507 if (vp->v_mount->mnt_flag & MNT_RDONLY)
508 return (EROFS);
509 break;
510 default:
511 break;
512 }
513 }
514 return (null_bypass((struct vop_generic_args *)ap));
515}
516
517/*
518 * We must handle open to be able to catch MNT_NODEV and friends.
519 */
520static int
521null_open(ap)
522 struct vop_open_args /* {
523 struct vnode *a_vp;
524 int a_mode;
525 struct ucred *a_cred;
526 struct thread *a_td;
527 } */ *ap;
528{
529 struct vnode *vp = ap->a_vp;
530 struct vnode *lvp = NULLVPTOLOWERVP(ap->a_vp);
531
532 if ((vp->v_mount->mnt_flag & MNT_NODEV) &&
533 (lvp->v_type == VBLK || lvp->v_type == VCHR))
534 return ENXIO;
535
536 return (null_bypass((struct vop_generic_args *)ap));
537}
538
539/*
540 * We handle this to eliminate null FS to lower FS
541 * file moving. Don't know why we don't allow this,
542 * possibly we should.
543 */
544static int
545null_rename(ap)
546 struct vop_rename_args /* {
547 struct vnode *a_fdvp;
548 struct vnode *a_fvp;
549 struct componentname *a_fcnp;
550 struct vnode *a_tdvp;
551 struct vnode *a_tvp;
552 struct componentname *a_tcnp;
553 } */ *ap;
554{
555 struct vnode *tdvp = ap->a_tdvp;
556 struct vnode *fvp = ap->a_fvp;
557 struct vnode *fdvp = ap->a_fdvp;
558 struct vnode *tvp = ap->a_tvp;
559
560 /* Check for cross-device rename. */
561 if ((fvp->v_mount != tdvp->v_mount) ||
562 (tvp && (fvp->v_mount != tvp->v_mount))) {
563 if (tdvp == tvp)
564 vrele(tdvp);
565 else
566 vput(tdvp);
567 if (tvp)
568 vput(tvp);
569 vrele(fdvp);
570 vrele(fvp);
571 return (EXDEV);
572 }
573
574 return (null_bypass((struct vop_generic_args *)ap));
575}
576
577/*
578 * We need to process our own vnode lock and then clear the
579 * interlock flag as it applies only to our vnode, not the
580 * vnodes below us on the stack.
581 */
582static int
583null_lock(ap)
584 struct vop_lock_args /* {
585 struct vnode *a_vp;
586 int a_flags;
587 struct thread *a_td;
588 } */ *ap;
589{
590 struct vnode *vp = ap->a_vp;
591 int flags = ap->a_flags;
592 struct thread *td = ap->a_td;
593 struct vnode *lvp;
594 int error;
595
596 if (flags & LK_THISLAYER) {
597 if (vp->v_vnlock != NULL) {
598 /* lock is shared across layers */
599 if (flags & LK_INTERLOCK)
600 mtx_unlock(&vp->v_interlock);
601 return 0;
602 }
603 error = lockmgr(&vp->v_lock, flags & ~LK_THISLAYER,
604 &vp->v_interlock, td);
605 return (error);
606 }
607
608 if (vp->v_vnlock != NULL) {
609 /*
610 * The lower level has exported a struct lock to us. Use
611 * it so that all vnodes in the stack lock and unlock
612 * simultaneously. Note: we don't DRAIN the lock as DRAIN
613 * decommissions the lock - just because our vnode is
614 * going away doesn't mean the struct lock below us is.
615 * LK_EXCLUSIVE is fine.
616 */
617 if ((flags & LK_TYPE_MASK) == LK_DRAIN) {
618 NULLFSDEBUG("null_lock: avoiding LK_DRAIN\n");
619 return(lockmgr(vp->v_vnlock,
620 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE,
621 &vp->v_interlock, td));
622 }
623 return(lockmgr(vp->v_vnlock, flags, &vp->v_interlock, td));
624 } else {
625 /*
626 * To prevent race conditions involving doing a lookup
627 * on "..", we have to lock the lower node, then lock our
628 * node. Most of the time it won't matter that we lock our
629 * node (as any locking would need the lower one locked
630 * first). But we can LK_DRAIN the upper lock as a step
631 * towards decomissioning it.
632 */
633 lvp = NULLVPTOLOWERVP(vp);
634 if (lvp == NULL)
635 return (lockmgr(&vp->v_lock, flags, &vp->v_interlock, td));
636 if (flags & LK_INTERLOCK) {
637 mtx_unlock(&vp->v_interlock);
638 flags &= ~LK_INTERLOCK;
639 }
640 if ((flags & LK_TYPE_MASK) == LK_DRAIN) {
641 error = VOP_LOCK(lvp,
642 (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE, td);
643 } else
644 error = VOP_LOCK(lvp, flags, td);
645 if (error)
646 return (error);
647 error = lockmgr(&vp->v_lock, flags, &vp->v_interlock, td);
648 if (error)
649 VOP_UNLOCK(lvp, 0, td);
650 return (error);
651 }
652}
653
654/*
655 * We need to process our own vnode unlock and then clear the
656 * interlock flag as it applies only to our vnode, not the
657 * vnodes below us on the stack.
658 */
659static int
660null_unlock(ap)
661 struct vop_unlock_args /* {
662 struct vnode *a_vp;
663 int a_flags;
664 struct thread *a_td;
665 } */ *ap;
666{
667 struct vnode *vp = ap->a_vp;
668 int flags = ap->a_flags;
669 struct thread *td = ap->a_td;
670 struct vnode *lvp;
671
672 if (vp->v_vnlock != NULL) {
673 if (flags & LK_THISLAYER)
674 return 0; /* the lock is shared across layers */
675 flags &= ~LK_THISLAYER;
676 return (lockmgr(vp->v_vnlock, flags | LK_RELEASE,
677 &vp->v_interlock, td));
678 }
679 lvp = NULLVPTOLOWERVP(vp);
680 if (lvp == NULL)
681 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td));
682 if ((flags & LK_THISLAYER) == 0) {
683 if (flags & LK_INTERLOCK) {
684 mtx_unlock(&vp->v_interlock);
685 flags &= ~LK_INTERLOCK;
686 }
687 VOP_UNLOCK(lvp, flags & ~LK_INTERLOCK, td);
688 } else
689 flags &= ~LK_THISLAYER;
690 return (lockmgr(&vp->v_lock, flags | LK_RELEASE, &vp->v_interlock, td));
691}
692
693static int
694null_islocked(ap)
695 struct vop_islocked_args /* {
696 struct vnode *a_vp;
697 struct thread *a_td;
698 } */ *ap;
699{
700 struct vnode *vp = ap->a_vp;
701 struct thread *td = ap->a_td;
702
703 if (vp->v_vnlock != NULL)
704 return (lockstatus(vp->v_vnlock, td));
705 return (lockstatus(&vp->v_lock, td));
706}
707
708/*
709 * There is no way to tell that someone issued remove/rmdir operation
710 * on the underlying filesystem. For now we just have to release lowevrp
711 * as soon as possible.
712 *
713 * Note, we can't release any resources nor remove vnode from hash before
714 * appropriate VXLOCK stuff is is done because other process can find this
715 * vnode in hash during inactivation and may be sitting in vget() and waiting
716 * for null_inactive to unlock vnode. Thus we will do all those in VOP_RECLAIM.
717 */
718static int
719null_inactive(ap)
720 struct vop_inactive_args /* {
721 struct vnode *a_vp;
722 struct thread *a_td;
723 } */ *ap;
724{
725 struct vnode *vp = ap->a_vp;
726 struct thread *td = ap->a_td;
727
728 VOP_UNLOCK(vp, 0, td);
729
730 /*
731 * If this is the last reference, then free up the vnode
732 * so as not to tie up the lower vnodes.
733 */
734 vrecycle(vp, NULL, td);
735
736 return (0);
737}
738
739/*
740 * Now, the VXLOCK is in force and we're free to destroy the null vnode.
741 */
742static int
743null_reclaim(ap)
744 struct vop_reclaim_args /* {
745 struct vnode *a_vp;
746 struct thread *a_td;
747 } */ *ap;
748{
749 struct vnode *vp = ap->a_vp;
750 struct null_node *xp = VTONULL(vp);
751 struct vnode *lowervp = xp->null_lowervp;
752
753 if (lowervp) {
754 null_hashrem(xp);
755
756 vrele(lowervp);
757 vrele(lowervp);
758 }
759
760 vp->v_data = NULL;
761 vp->v_vnlock = &vp->v_lock;
762 FREE(xp, M_NULLFSNODE);
763
764 return (0);
765}
766
767static int
768null_print(ap)
769 struct vop_print_args /* {
770 struct vnode *a_vp;
771 } */ *ap;
772{
773 register struct vnode *vp = ap->a_vp;
774 printf("\ttag %s, vp=%p, lowervp=%p\n", vp->v_tag, vp,
775 NULLVPTOLOWERVP(vp));
776 return (0);
777}
778
779/*
780 * Let an underlying filesystem do the work
781 */
782static int
783null_createvobject(ap)
784 struct vop_createvobject_args /* {
785 struct vnode *vp;
786 struct ucred *cred;
787 struct thread *td;
788 } */ *ap;
789{
790 struct vnode *vp = ap->a_vp;
791 struct vnode *lowervp = VTONULL(vp) ? NULLVPTOLOWERVP(vp) : NULL;
792 int error;
793
794 if (vp->v_type == VNON || lowervp == NULL)
795 return 0;
796 error = VOP_CREATEVOBJECT(lowervp, ap->a_cred, ap->a_td);
797 if (error)
798 return (error);
799 vp->v_vflag |= VV_OBJBUF;
800 return (0);
801}
802
803/*
804 * We have nothing to destroy and this operation shouldn't be bypassed.
805 */
806static int
807null_destroyvobject(ap)
808 struct vop_destroyvobject_args /* {
809 struct vnode *vp;
810 } */ *ap;
811{
812 struct vnode *vp = ap->a_vp;
813
814 vp->v_vflag &= ~VV_OBJBUF;
815 return (0);
816}
817
818static int
819null_getvobject(ap)
820 struct vop_getvobject_args /* {
821 struct vnode *vp;
822 struct vm_object **objpp;
823 } */ *ap;
824{
825 struct vnode *lvp = NULLVPTOLOWERVP(ap->a_vp);
826
827 if (lvp == NULL)
828 return EINVAL;
829 return (VOP_GETVOBJECT(lvp, ap->a_objpp));
830}
831
832/*
833 * Global vfs data structures
834 */
835vop_t **null_vnodeop_p;
836static struct vnodeopv_entry_desc null_vnodeop_entries[] = {
837 { &vop_default_desc, (vop_t *) null_bypass },
838
839 { &vop_access_desc, (vop_t *) null_access },
840 { &vop_bmap_desc, (vop_t *) vop_eopnotsupp },
841 { &vop_createvobject_desc, (vop_t *) null_createvobject },
842 { &vop_destroyvobject_desc, (vop_t *) null_destroyvobject },
843 { &vop_getattr_desc, (vop_t *) null_getattr },
844 { &vop_getvobject_desc, (vop_t *) null_getvobject },
845 { &vop_getwritemount_desc, (vop_t *) vop_stdgetwritemount},
846 { &vop_inactive_desc, (vop_t *) null_inactive },
847 { &vop_islocked_desc, (vop_t *) null_islocked },
848 { &vop_lock_desc, (vop_t *) null_lock },
849 { &vop_lookup_desc, (vop_t *) null_lookup },
850 { &vop_open_desc, (vop_t *) null_open },
851 { &vop_print_desc, (vop_t *) null_print },
852 { &vop_reclaim_desc, (vop_t *) null_reclaim },
853 { &vop_rename_desc, (vop_t *) null_rename },
854 { &vop_setattr_desc, (vop_t *) null_setattr },
855 { &vop_strategy_desc, (vop_t *) vop_eopnotsupp },
856 { &vop_unlock_desc, (vop_t *) null_unlock },
857 { NULL, NULL }
858};
859static struct vnodeopv_desc null_vnodeop_opv_desc =
860 { &null_vnodeop_p, null_vnodeop_entries };
861
862VNODEOP_SET(null_vnodeop_opv_desc);
|