1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1992, 1993
5 * The Regents of the University of California. All rights reserved.
6 *
7 * This code is derived from software contributed to Berkeley by
8 * John Heidemann of the UCLA Ficus project.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
34 * Ancestors:
35 * ...and...
36 */
37
38 /*
39 * Null Layer
40 *
41 * (See mount_nullfs(8) for more information.)
42 *
43 * The null layer duplicates a portion of the filesystem
44 * name space under a new name. In this respect, it is
45 * similar to the loopback filesystem. It differs from
46 * the loopback fs in two respects: it is implemented using
47 * a stackable layers techniques, and its "null-node"s stack above
48 * all lower-layer vnodes, not just over directory vnodes.
49 *
50 * The null layer has two purposes. First, it serves as a demonstration
51 * of layering by proving a layer which does nothing. (It actually
52 * does everything the loopback filesystem does, which is slightly
53 * more than nothing.) Second, the null layer can serve as a prototype
54 * layer. Since it provides all necessary layer framework,
55 * new filesystem layers can be created very easily be starting
56 * with a null layer.
57 *
58 * The remainder of this man page examines the null layer as a basis
59 * for constructing new layers.
60 *
61 *
62 * INSTANTIATING NEW NULL LAYERS
63 *
64 * New null layers are created with mount_nullfs(8).
65 * Mount_nullfs(8) takes two arguments, the pathname
66 * of the lower vfs (target-pn) and the pathname where the null
67 * layer will appear in the namespace (alias-pn). After
68 * the null layer is put into place, the contents
69 * of target-pn subtree will be aliased under alias-pn.
70 *
71 *
72 * OPERATION OF A NULL LAYER
73 *
74 * The null layer is the minimum filesystem layer,
75 * simply bypassing all possible operations to the lower layer
76 * for processing there. The majority of its activity centers
77 * on the bypass routine, through which nearly all vnode operations
78 * pass.
79 *
80 * The bypass routine accepts arbitrary vnode operations for
81 * handling by the lower layer. It begins by examining vnode
82 * operation arguments and replacing any null-nodes by their
83 * lower-layer equivlants. It then invokes the operation
84 * on the lower layer. Finally, it replaces the null-nodes
85 * in the arguments and, if a vnode is return by the operation,
86 * stacks a null-node on top of the returned vnode.
87 *
88 * Although bypass handles most operations, vop_getattr, vop_lock,
89 * vop_unlock, vop_inactive, vop_reclaim, and vop_print are not
90 * bypassed. Vop_getattr must change the fsid being returned.
91 * Vop_lock and vop_unlock must handle any locking for the
92 * current vnode as well as pass the lock request down.
93 * Vop_inactive and vop_reclaim are not bypassed so that
94 * they can handle freeing null-layer specific data. Vop_print
95 * is not bypassed to avoid excessive debugging information.
96 * Also, certain vnode operations change the locking state within
97 * the operation (create, mknod, remove, link, rename, mkdir, rmdir,
98 * and symlink). Ideally these operations should not change the
99 * lock state, but should be changed to let the caller of the
100 * function unlock them. Otherwise all intermediate vnode layers
101 * (such as union, umapfs, etc) must catch these functions to do
102 * the necessary locking at their layer.
103 *
104 *
105 * INSTANTIATING VNODE STACKS
106 *
107 * Mounting associates the null layer with a lower layer,
108 * effect stacking two VFSes. Vnode stacks are instead
109 * created on demand as files are accessed.
110 *
111 * The initial mount creates a single vnode stack for the
112 * root of the new null layer. All other vnode stacks
113 * are created as a result of vnode operations on
114 * this or other null vnode stacks.
115 *
116 * New vnode stacks come into existence as a result of
117 * an operation which returns a vnode.
118 * The bypass routine stacks a null-node above the new
119 * vnode before returning it to the caller.
120 *
121 * For example, imagine mounting a null layer with
122 * "mount_nullfs /usr/include /dev/layer/null".
123 * Changing directory to /dev/layer/null will assign
124 * the root null-node (which was created when the null layer was mounted).
125 * Now consider opening "sys". A vop_lookup would be
126 * done on the root null-node. This operation would bypass through
127 * to the lower layer which would return a vnode representing
128 * the UFS "sys". Null_bypass then builds a null-node
129 * aliasing the UFS "sys" and returns this to the caller.
130 * Later operations on the null-node "sys" will repeat this
131 * process when constructing other vnode stacks.
132 *
133 *
134 * CREATING OTHER FILE SYSTEM LAYERS
135 *
136 * One of the easiest ways to construct new filesystem layers is to make
137 * a copy of the null layer, rename all files and variables, and
138 * then begin modifying the copy. Sed can be used to easily rename
139 * all variables.
140 *
141 * The umap layer is an example of a layer descended from the
142 * null layer.
143 *
144 *
145 * INVOKING OPERATIONS ON LOWER LAYERS
146 *
147 * There are two techniques to invoke operations on a lower layer
148 * when the operation cannot be completely bypassed. Each method
149 * is appropriate in different situations. In both cases,
150 * it is the responsibility of the aliasing layer to make
151 * the operation arguments "correct" for the lower layer
152 * by mapping a vnode arguments to the lower layer.
153 *
154 * The first approach is to call the aliasing layer's bypass routine.
155 * This method is most suitable when you wish to invoke the operation
156 * currently being handled on the lower layer. It has the advantage
157 * that the bypass routine already must do argument mapping.
158 * An example of this is null_getattrs in the null layer.
159 *
160 * A second approach is to directly invoke vnode operations on
161 * the lower layer with the VOP_OPERATIONNAME interface.
162 * The advantage of this method is that it is easy to invoke
163 * arbitrary operations on the lower layer. The disadvantage
164 * is that vnode arguments must be manualy mapped.
165 *
166 */
167
168 #include <sys/param.h>
169 #include <sys/systm.h>
170 #include <sys/conf.h>
171 #include <sys/kernel.h>
172 #include <sys/lock.h>
173 #include <sys/malloc.h>
174 #include <sys/mount.h>
175 #include <sys/mutex.h>
176 #include <sys/namei.h>
177 #include <sys/proc.h>
178 #include <sys/smr.h>
179 #include <sys/sysctl.h>
180 #include <sys/vnode.h>
181 #include <sys/stat.h>
182
183 #include <fs/nullfs/null.h>
184
185 #include <vm/vm.h>
186 #include <vm/vm_extern.h>
187 #include <vm/vm_object.h>
188 #include <vm/vnode_pager.h>
189
190 VFS_SMR_DECLARE;
191
192 static int null_bug_bypass = 0; /* for debugging: enables bypass printf'ing */
193 SYSCTL_INT(_debug, OID_AUTO, nullfs_bug_bypass, CTLFLAG_RW,
194 &null_bug_bypass, 0, "");
195
196 /*
197 * Synchronize inotify flags with the lower vnode:
198 * - If the upper vnode has the flag set and the lower does not, then the lower
199 * vnode is unwatched and the upper vnode does not need to go through
200 * VOP_INOTIFY.
201 * - If the lower vnode is watched, then the upper vnode should go through
202 * VOP_INOTIFY, so copy the flag up.
203 *
204 * The lockless check is only a fast path: the decision to change a flag
205 * is re-made under the upper vnode's interlock, since another thread may
206 * set or clear the flag concurrently.
207 */
208 static void
null_copy_inotify(struct vnode * vp,struct vnode * lvp,short flag)209 null_copy_inotify(struct vnode *vp, struct vnode *lvp, short flag)
210 {
211 if (__predict_true((vn_irflag_read(vp) & flag) ==
212 (vn_irflag_read(lvp) & flag)))
213 return;
214 VI_LOCK(vp);
215 if ((vn_irflag_read(vp) & flag) != 0) {
216 if ((vn_irflag_read(lvp) & flag) == 0)
217 vn_irflag_unset_locked(vp, flag);
218 } else {
219 if ((vn_irflag_read(lvp) & flag) != 0)
220 vn_irflag_set_locked(vp, flag);
221 }
222 VI_UNLOCK(vp);
223 }
224
225 /*
226 * This is the 10-Apr-92 bypass routine.
227 * This version has been optimized for speed, throwing away some
228 * safety checks. It should still always work, but it's not as
229 * robust to programmer errors.
230 *
231 * In general, we map all vnodes going down and unmap them on the way back.
232 * As an exception to this, vnodes can be marked "unmapped" by setting
233 * the Nth bit in operation's vdesc_flags.
234 *
235 * Also, some BSD vnode operations have the side effect of vrele'ing
236 * their arguments. With stacking, the reference counts are held
237 * by the upper node, not the lower one, so we must handle these
238 * side-effects here. This is not of concern in Sun-derived systems
239 * since there are no such side-effects.
240 *
241 * This makes the following assumptions:
242 * - only one returned vpp
243 * - no INOUT vpp's (Sun's vop_open has one of these)
244 * - the vnode operation vector of the first vnode should be used
245 * to determine what implementation of the op should be invoked
246 * - all mapped vnodes are of our vnode-type (NEEDSWORK:
247 * problems on rmdir'ing mount points and renaming?)
248 */
249 int
null_bypass(struct vop_generic_args * ap)250 null_bypass(struct vop_generic_args *ap)
251 {
252 struct vnode **this_vp_p;
253 struct vnode *old_vps[VDESC_MAX_VPS];
254 struct vnode **vps_p[VDESC_MAX_VPS];
255 struct vnode ***vppp;
256 struct vnode *lvp;
257 struct vnodeop_desc *descp = ap->a_desc;
258 int error, i, reles;
259
260 if (null_bug_bypass)
261 printf ("null_bypass: %s\n", descp->vdesc_name);
262
263 #ifdef DIAGNOSTIC
264 /*
265 * We require at least one vp.
266 */
267 if (descp->vdesc_vp_offsets == NULL ||
268 descp->vdesc_vp_offsets[0] == VDESC_NO_OFFSET)
269 panic ("null_bypass: no vp's in map");
270 #endif
271
272 /*
273 * Map the vnodes going in.
274 * Later, we'll invoke the operation based on
275 * the first mapped vnode's operation vector.
276 */
277 reles = descp->vdesc_flags;
278 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
279 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
280 break; /* bail out at end of list */
281 vps_p[i] = this_vp_p = VOPARG_OFFSETTO(struct vnode **,
282 descp->vdesc_vp_offsets[i], ap);
283
284 /*
285 * We're not guaranteed that any but the first vnode
286 * are of our type. Check for and don't map any
287 * that aren't. (We must always map first vp or vclean fails.)
288 */
289 if (i != 0 && (*this_vp_p == NULL ||
290 !null_is_nullfs_vnode(*this_vp_p))) {
291 old_vps[i] = NULL;
292 } else {
293 old_vps[i] = *this_vp_p;
294 *(vps_p[i]) = NULLVPTOLOWERVP(*this_vp_p);
295
296 /*
297 * The upper vnode reference to the lower
298 * vnode is the only reference that keeps our
299 * pointer to the lower vnode alive. If lower
300 * vnode is relocked during the VOP call,
301 * upper vnode might become unlocked and
302 * reclaimed, which invalidates our reference.
303 * Add a transient hold around VOP call.
304 */
305 vhold(*this_vp_p);
306
307 /*
308 * XXX - Several operations have the side effect
309 * of vrele'ing their vp's. We must account for
310 * that. (This should go away in the future.)
311 */
312 if (reles & VDESC_VP0_WILLRELE)
313 vref(*this_vp_p);
314 }
315 }
316
317 /*
318 * Call the operation on the lower layer
319 * with the modified argument structure.
320 */
321 if (vps_p[0] != NULL && *vps_p[0] != NULL) {
322 error = ap->a_desc->vdesc_call(ap);
323 } else {
324 printf("null_bypass: no map for %s\n", descp->vdesc_name);
325 error = EINVAL;
326 }
327
328 /*
329 * Maintain the illusion of call-by-value
330 * by restoring vnodes in the argument structure
331 * to their original value.
332 */
333 reles = descp->vdesc_flags;
334 for (i = 0; i < VDESC_MAX_VPS; reles >>= 1, i++) {
335 if (descp->vdesc_vp_offsets[i] == VDESC_NO_OFFSET)
336 break; /* bail out at end of list */
337 if (old_vps[i] != NULL) {
338 lvp = *(vps_p[i]);
339
340 /*
341 * Get rid of the transient hold on lvp. Copy inotify
342 * flags up in case something is watching the lower
343 * layer.
344 *
345 * If lowervp was unlocked during VOP
346 * operation, nullfs upper vnode could have
347 * been reclaimed, which changes its v_vnlock
348 * back to private v_lock. In this case we
349 * must move lock ownership from lower to
350 * upper (reclaimed) vnode.
351 */
352 if (lvp != NULL) {
353 null_copy_inotify(old_vps[i], lvp,
354 VIRF_INOTIFY);
355 null_copy_inotify(old_vps[i], lvp,
356 VIRF_INOTIFY_PARENT);
357 if (VOP_ISLOCKED(lvp) == LK_EXCLUSIVE &&
358 old_vps[i]->v_vnlock != lvp->v_vnlock) {
359 VOP_UNLOCK(lvp);
360 VOP_LOCK(old_vps[i], LK_EXCLUSIVE |
361 LK_RETRY);
362 }
363 vdrop(lvp);
364 }
365
366 *(vps_p[i]) = old_vps[i];
367 #if 0
368 if (reles & VDESC_VP0_WILLUNLOCK)
369 VOP_UNLOCK(*(vps_p[i]), 0);
370 #endif
371 if (reles & VDESC_VP0_WILLRELE)
372 vrele(*(vps_p[i]));
373 }
374 }
375
376 /*
377 * Map the possible out-going vpp
378 * (Assumes that the lower layer always returns
379 * a VREF'ed vpp unless it gets an error.)
380 */
381 if (descp->vdesc_vpp_offset != VDESC_NO_OFFSET && error == 0) {
382 /*
383 * XXX - even though some ops have vpp returned vp's,
384 * several ops actually vrele this before returning.
385 * We must avoid these ops.
386 * (This should go away when these ops are regularized.)
387 */
388 vppp = VOPARG_OFFSETTO(struct vnode ***,
389 descp->vdesc_vpp_offset, ap);
390 if (*vppp != NULL)
391 error = null_nodeget(old_vps[0]->v_mount, **vppp,
392 *vppp);
393 }
394
395 return (error);
396 }
397
398 static int
null_add_writecount(struct vop_add_writecount_args * ap)399 null_add_writecount(struct vop_add_writecount_args *ap)
400 {
401 struct vnode *lvp, *vp;
402 int error;
403
404 vp = ap->a_vp;
405 lvp = NULLVPTOLOWERVP(vp);
406 VI_LOCK(vp);
407 /* text refs are bypassed to lowervp */
408 VNASSERT(vp->v_writecount >= 0, vp, ("wrong null writecount"));
409 VNASSERT(vp->v_writecount + ap->a_inc >= 0, vp,
410 ("wrong writecount inc %d", ap->a_inc));
411 error = VOP_ADD_WRITECOUNT(lvp, ap->a_inc);
412 if (error == 0)
413 vp->v_writecount += ap->a_inc;
414 VI_UNLOCK(vp);
415 return (error);
416 }
417
418 /*
419 * We have to carry on the locking protocol on the null layer vnodes
420 * as we progress through the tree. We also have to enforce read-only
421 * if this layer is mounted read-only.
422 */
423 static int
null_lookup(struct vop_lookup_args * ap)424 null_lookup(struct vop_lookup_args *ap)
425 {
426 struct componentname *cnp = ap->a_cnp;
427 struct vnode *dvp = ap->a_dvp;
428 uint64_t flags = cnp->cn_flags;
429 struct vnode *vp, *ldvp, *lvp;
430 struct mount *mp;
431 int error;
432
433 mp = dvp->v_mount;
434 if ((flags & ISLASTCN) != 0 && (mp->mnt_flag & MNT_RDONLY) != 0 &&
435 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
436 return (EROFS);
437 /*
438 * Although it is possible to call null_bypass(), we'll do
439 * a direct call to reduce overhead
440 */
441 ldvp = NULLVPTOLOWERVP(dvp);
442 vp = lvp = NULL;
443
444 /*
445 * Renames in the lower mounts might create an inconsistent
446 * configuration where lower vnode is moved out of the directory tree
447 * remounted by our null mount.
448 *
449 * Do not try to handle it fancy, just avoid VOP_LOOKUP() with DOTDOT
450 * name which cannot be handled by the VOP.
451 */
452 if ((flags & ISDOTDOT) != 0) {
453 struct nameidata *ndp;
454
455 if ((ldvp->v_vflag & VV_ROOT) != 0) {
456 KASSERT((dvp->v_vflag & VV_ROOT) == 0,
457 ("ldvp %p fl %#x dvp %p fl %#x flags %#jx",
458 ldvp, ldvp->v_vflag, dvp, dvp->v_vflag,
459 (uintmax_t)flags));
460 return (ENOENT);
461 }
462 ndp = vfs_lookup_nameidata(cnp);
463 if (ndp != NULL && vfs_lookup_isroot(ndp, ldvp))
464 return (ENOENT);
465 }
466
467 /*
468 * Hold ldvp. The reference on it, owned by dvp, is lost in
469 * case of dvp reclamation, and we need ldvp to move our lock
470 * from ldvp to dvp.
471 */
472 vhold(ldvp);
473
474 error = VOP_LOOKUP(ldvp, &lvp, cnp);
475
476 /*
477 * VOP_LOOKUP() on lower vnode may unlock ldvp, which allows
478 * dvp to be reclaimed due to shared v_vnlock. Check for the
479 * doomed state and return error.
480 */
481 if (VN_IS_DOOMED(dvp)) {
482 if (error == 0 || error == EJUSTRETURN) {
483 if (lvp != NULL)
484 vput(lvp);
485 error = ENOENT;
486 }
487
488 /*
489 * If vgone() did reclaimed dvp before curthread
490 * relocked ldvp, the locks of dvp and ldpv are no
491 * longer shared. In this case, relock of ldvp in
492 * lower fs VOP_LOOKUP() does not restore the locking
493 * state of dvp. Compensate for this by unlocking
494 * ldvp and locking dvp, which is also correct if the
495 * locks are still shared.
496 */
497 VOP_UNLOCK(ldvp);
498 vn_lock(dvp, LK_EXCLUSIVE | LK_RETRY);
499 }
500 vdrop(ldvp);
501
502 if (error == EJUSTRETURN && (flags & ISLASTCN) != 0 &&
503 (mp->mnt_flag & MNT_RDONLY) != 0 &&
504 (cnp->cn_nameiop == CREATE || cnp->cn_nameiop == RENAME))
505 error = EROFS;
506
507 if ((error == 0 || error == EJUSTRETURN) && lvp != NULL) {
508 if (ldvp == lvp) {
509 *ap->a_vpp = dvp;
510 vref(dvp);
511 vrele(lvp);
512 } else {
513 error = null_nodeget(mp, lvp, &vp);
514 if (error == 0)
515 *ap->a_vpp = vp;
516 }
517 }
518 return (error);
519 }
520
521 static int
null_open(struct vop_open_args * ap)522 null_open(struct vop_open_args *ap)
523 {
524 int retval;
525 struct vnode *vp, *ldvp;
526
527 vp = ap->a_vp;
528 ldvp = NULLVPTOLOWERVP(vp);
529 retval = null_bypass(&ap->a_gen);
530 if (retval == 0) {
531 vp->v_object = ldvp->v_object;
532 if ((vn_irflag_read(ldvp) & VIRF_PGREAD) != 0) {
533 MPASS(vp->v_object != NULL);
534 if ((vn_irflag_read(vp) & VIRF_PGREAD) == 0) {
535 vn_irflag_set_cond(vp, VIRF_PGREAD);
536 }
537 }
538 }
539 return (retval);
540 }
541
542 /*
543 * Setattr call. Disallow write attempts if the layer is mounted read-only.
544 */
545 static int
null_setattr(struct vop_setattr_args * ap)546 null_setattr(struct vop_setattr_args *ap)
547 {
548 struct vnode *vp = ap->a_vp;
549 struct vattr *vap = ap->a_vap;
550
551 if ((vap->va_flags != VNOVAL || vap->va_uid != (uid_t)VNOVAL ||
552 vap->va_gid != (gid_t)VNOVAL || vap->va_atime.tv_sec != VNOVAL ||
553 vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) &&
554 (vp->v_mount->mnt_flag & MNT_RDONLY))
555 return (EROFS);
556 if (vap->va_size != VNOVAL) {
557 switch (vp->v_type) {
558 case VDIR:
559 return (EISDIR);
560 case VCHR:
561 case VBLK:
562 case VSOCK:
563 case VFIFO:
564 if (vap->va_flags != VNOVAL)
565 return (EOPNOTSUPP);
566 return (0);
567 case VREG:
568 case VLNK:
569 default:
570 /*
571 * Disallow write attempts if the filesystem is
572 * mounted read-only.
573 */
574 if (vp->v_mount->mnt_flag & MNT_RDONLY)
575 return (EROFS);
576 }
577 }
578
579 return (null_bypass(&ap->a_gen));
580 }
581
582 /*
583 * We handle stat and getattr only to change the fsid.
584 */
585 static int
null_stat(struct vop_stat_args * ap)586 null_stat(struct vop_stat_args *ap)
587 {
588 int error;
589
590 if ((error = null_bypass(&ap->a_gen)) != 0)
591 return (error);
592
593 ap->a_sb->st_dev = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
594 return (0);
595 }
596
597 static int
null_getattr(struct vop_getattr_args * ap)598 null_getattr(struct vop_getattr_args *ap)
599 {
600 int error;
601
602 if ((error = null_bypass(&ap->a_gen)) != 0)
603 return (error);
604
605 ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
606 return (0);
607 }
608
609 /*
610 * Handle to disallow write access if mounted read-only.
611 */
612 static int
null_access(struct vop_access_args * ap)613 null_access(struct vop_access_args *ap)
614 {
615 struct vnode *vp = ap->a_vp;
616 accmode_t accmode = ap->a_accmode;
617
618 /*
619 * Disallow write attempts on read-only layers;
620 * unless the file is a socket, fifo, or a block or
621 * character device resident on the filesystem.
622 */
623 if (accmode & VWRITE) {
624 switch (vp->v_type) {
625 case VDIR:
626 case VLNK:
627 case VREG:
628 if (vp->v_mount->mnt_flag & MNT_RDONLY)
629 return (EROFS);
630 break;
631 default:
632 break;
633 }
634 }
635 return (null_bypass(&ap->a_gen));
636 }
637
638 static int
null_accessx(struct vop_accessx_args * ap)639 null_accessx(struct vop_accessx_args *ap)
640 {
641 struct vnode *vp = ap->a_vp;
642 accmode_t accmode = ap->a_accmode;
643
644 /*
645 * Disallow write attempts on read-only layers;
646 * unless the file is a socket, fifo, or a block or
647 * character device resident on the filesystem.
648 */
649 if (accmode & VWRITE) {
650 switch (vp->v_type) {
651 case VDIR:
652 case VLNK:
653 case VREG:
654 if (vp->v_mount->mnt_flag & MNT_RDONLY)
655 return (EROFS);
656 break;
657 default:
658 break;
659 }
660 }
661 return (null_bypass(&ap->a_gen));
662 }
663
664 /*
665 * Increasing refcount of lower vnode is needed at least for the case
666 * when lower FS is NFS to do sillyrename if the file is in use.
667 * Unfortunately v_usecount is incremented in many places in
668 * the kernel and, as such, there may be races that result in
669 * the NFS client doing an extraneous silly rename, but that seems
670 * preferable to not doing a silly rename when it is needed.
671 */
672 static int
null_remove(struct vop_remove_args * ap)673 null_remove(struct vop_remove_args *ap)
674 {
675 int retval, vreleit;
676 struct vnode *lvp, *vp;
677
678 vp = ap->a_vp;
679 if (vrefcnt(vp) > 1) {
680 lvp = NULLVPTOLOWERVP(vp);
681 vref(lvp);
682 vreleit = 1;
683 } else
684 vreleit = 0;
685 VTONULL(vp)->null_flags |= NULLV_DROP;
686 retval = null_bypass(&ap->a_gen);
687 if (vreleit != 0)
688 vrele(lvp);
689 return (retval);
690 }
691
692 /*
693 * We handle this to eliminate null FS to lower FS
694 * file moving. Don't know why we don't allow this,
695 * possibly we should.
696 */
697 static int
null_rename(struct vop_rename_args * ap)698 null_rename(struct vop_rename_args *ap)
699 {
700 struct vnode *fdvp, *fvp, *tdvp, *tvp;
701 struct vnode *lfdvp, *lfvp, *ltdvp, *ltvp;
702 struct null_node *fdnn, *fnn, *tdnn, *tnn;
703 int error;
704
705 tdvp = ap->a_tdvp;
706 fvp = ap->a_fvp;
707 fdvp = ap->a_fdvp;
708 tvp = ap->a_tvp;
709 lfdvp = NULL;
710
711 /* Check for cross-device rename. */
712 if ((fvp->v_mount != tdvp->v_mount) ||
713 (tvp != NULL && fvp->v_mount != tvp->v_mount)) {
714 error = EXDEV;
715 goto upper_err;
716 }
717
718 VI_LOCK(fdvp);
719 fdnn = VTONULL(fdvp);
720 if (fdnn == NULL) { /* fdvp is not locked, can be doomed */
721 VI_UNLOCK(fdvp);
722 error = ENOENT;
723 goto upper_err;
724 }
725 lfdvp = fdnn->null_lowervp;
726 vref(lfdvp);
727 VI_UNLOCK(fdvp);
728
729 VI_LOCK(fvp);
730 fnn = VTONULL(fvp);
731 if (fnn == NULL) {
732 VI_UNLOCK(fvp);
733 error = ENOENT;
734 goto upper_err;
735 }
736 lfvp = fnn->null_lowervp;
737 vref(lfvp);
738 VI_UNLOCK(fvp);
739
740 tdnn = VTONULL(tdvp);
741 ltdvp = tdnn->null_lowervp;
742 vref(ltdvp);
743
744 if (tvp != NULL) {
745 tnn = VTONULL(tvp);
746 ltvp = tnn->null_lowervp;
747 vref(ltvp);
748 tnn->null_flags |= NULLV_DROP;
749 } else {
750 ltvp = NULL;
751 }
752
753 error = VOP_RENAME(lfdvp, lfvp, ap->a_fcnp, ltdvp, ltvp, ap->a_tcnp,
754 ap->a_flags);
755 vrele(fdvp);
756 vrele(fvp);
757 vrele(tdvp);
758 if (tvp != NULL)
759 vrele(tvp);
760 return (error);
761
762 upper_err:
763 if (tdvp == tvp)
764 vrele(tdvp);
765 else
766 vput(tdvp);
767 if (tvp)
768 vput(tvp);
769 if (lfdvp != NULL)
770 vrele(lfdvp);
771 vrele(fdvp);
772 vrele(fvp);
773 return (error);
774 }
775
776 static int
null_rmdir(struct vop_rmdir_args * ap)777 null_rmdir(struct vop_rmdir_args *ap)
778 {
779
780 VTONULL(ap->a_vp)->null_flags |= NULLV_DROP;
781 return (null_bypass(&ap->a_gen));
782 }
783
784 /*
785 * We need to process our own vnode lock and then clear the interlock flag as
786 * it applies only to our vnode, not the vnodes below us on the stack.
787 *
788 * We have to hold the vnode here to solve a potential reclaim race. If we're
789 * forcibly vgone'd while we still have refs, a thread could be sleeping inside
790 * the lowervp's vop_lock routine. When we vgone we will drop our last ref to
791 * the lowervp, which would allow it to be reclaimed. The lowervp could then
792 * be recycled, in which case it is not legal to be sleeping in its VOP. We
793 * prevent it from being recycled by holding the vnode here.
794 */
795 static struct vnode *
null_lock_prep_with_smr(struct vop_lock1_args * ap)796 null_lock_prep_with_smr(struct vop_lock1_args *ap)
797 {
798 struct null_node *nn;
799 struct vnode *lvp;
800
801 lvp = NULL;
802
803 vfs_smr_enter();
804
805 nn = VTONULL_SMR(ap->a_vp);
806 if (__predict_true(nn != NULL)) {
807 lvp = nn->null_lowervp;
808 if (lvp != NULL && !vhold_smr(lvp))
809 lvp = NULL;
810 }
811
812 vfs_smr_exit();
813 return (lvp);
814 }
815
816 static struct vnode *
null_lock_prep_with_interlock(struct vop_lock1_args * ap)817 null_lock_prep_with_interlock(struct vop_lock1_args *ap)
818 {
819 struct null_node *nn;
820 struct vnode *lvp;
821
822 ASSERT_VI_LOCKED(ap->a_vp, __func__);
823
824 ap->a_flags &= ~LK_INTERLOCK;
825
826 lvp = NULL;
827
828 nn = VTONULL(ap->a_vp);
829 if (__predict_true(nn != NULL)) {
830 lvp = nn->null_lowervp;
831 if (lvp != NULL)
832 vholdnz(lvp);
833 }
834 VI_UNLOCK(ap->a_vp);
835 return (lvp);
836 }
837
838 static int
null_lock(struct vop_lock1_args * ap)839 null_lock(struct vop_lock1_args *ap)
840 {
841 struct vnode *lvp;
842 int error, flags;
843
844 if (__predict_true((ap->a_flags & LK_INTERLOCK) == 0)) {
845 lvp = null_lock_prep_with_smr(ap);
846 if (__predict_false(lvp == NULL)) {
847 VI_LOCK(ap->a_vp);
848 lvp = null_lock_prep_with_interlock(ap);
849 }
850 } else {
851 lvp = null_lock_prep_with_interlock(ap);
852 }
853
854 ASSERT_VI_UNLOCKED(ap->a_vp, __func__);
855
856 if (__predict_false(lvp == NULL))
857 return (vop_stdlock(ap));
858
859 VNPASS(lvp->v_holdcnt > 0, lvp);
860 error = VOP_LOCK(lvp, ap->a_flags);
861 /*
862 * We might have slept to get the lock and someone might have
863 * clean our vnode already, switching vnode lock from one in
864 * lowervp to v_lock in our own vnode structure. Handle this
865 * case by reacquiring correct lock in requested mode.
866 */
867 if (VTONULL(ap->a_vp) == NULL && error == 0) {
868 VOP_UNLOCK(lvp);
869
870 flags = ap->a_flags;
871 ap->a_flags &= ~LK_TYPE_MASK;
872 switch (flags & LK_TYPE_MASK) {
873 case LK_SHARED:
874 ap->a_flags |= LK_SHARED;
875 break;
876 case LK_UPGRADE:
877 case LK_EXCLUSIVE:
878 ap->a_flags |= LK_EXCLUSIVE;
879 break;
880 default:
881 panic("Unsupported lock request %d\n",
882 flags);
883 }
884 error = vop_stdlock(ap);
885 }
886 vdrop(lvp);
887 return (error);
888 }
889
890 static int
null_unlock(struct vop_unlock_args * ap)891 null_unlock(struct vop_unlock_args *ap)
892 {
893 struct vnode *vp = ap->a_vp;
894 struct null_node *nn;
895 struct vnode *lvp;
896 int error;
897
898 /*
899 * Contrary to null_lock, we don't need to hold the vnode around
900 * unlock.
901 *
902 * We hold the lock, which means we can't be racing against vgone.
903 *
904 * At the same time VOP_UNLOCK promises to not touch anything after
905 * it finishes unlock, just like we don't.
906 *
907 * vop_stdunlock for a doomed vnode matches doomed locking in null_lock.
908 */
909 nn = VTONULL(vp);
910 if (nn != NULL && (lvp = NULLVPTOLOWERVP(vp)) != NULL) {
911 error = VOP_UNLOCK(lvp);
912 } else {
913 error = vop_stdunlock(ap);
914 }
915
916 return (error);
917 }
918
919 /*
920 * Do not allow the VOP_INACTIVE to be passed to the lower layer,
921 * since the reference count on the lower vnode is not related to
922 * ours.
923 */
924 static int
null_want_recycle(struct vnode * vp)925 null_want_recycle(struct vnode *vp)
926 {
927 struct vnode *lvp;
928 struct null_node *xp;
929 struct mount *mp;
930 struct null_mount *xmp;
931
932 xp = VTONULL(vp);
933 lvp = NULLVPTOLOWERVP(vp);
934 mp = vp->v_mount;
935 xmp = MOUNTTONULLMOUNT(mp);
936 if ((xmp->nullm_flags & NULLM_CACHE) == 0 ||
937 (xp->null_flags & NULLV_DROP) != 0 ||
938 (lvp->v_vflag & VV_NOSYNC) != 0) {
939 /*
940 * If this is the last reference and caching of the
941 * nullfs vnodes is not enabled, or the lower vnode is
942 * deleted, then free up the vnode so as not to tie up
943 * the lower vnodes.
944 */
945 return (1);
946 }
947 return (0);
948 }
949
950 static int
null_inactive(struct vop_inactive_args * ap)951 null_inactive(struct vop_inactive_args *ap)
952 {
953 struct vnode *vp;
954
955 vp = ap->a_vp;
956 if (null_want_recycle(vp)) {
957 vp->v_object = NULL;
958 vrecycle(vp);
959 }
960 return (0);
961 }
962
963 static int
null_need_inactive(struct vop_need_inactive_args * ap)964 null_need_inactive(struct vop_need_inactive_args *ap)
965 {
966
967 return (null_want_recycle(ap->a_vp) || vn_need_pageq_flush(ap->a_vp));
968 }
969
970 /*
971 * Now, the nullfs vnode and, due to the sharing lock, the lower
972 * vnode, are exclusively locked, and we shall destroy the null vnode.
973 */
974 static int
null_reclaim(struct vop_reclaim_args * ap)975 null_reclaim(struct vop_reclaim_args *ap)
976 {
977 struct vnode *vp;
978 struct null_node *xp;
979 struct vnode *lowervp;
980 short flags;
981
982 vp = ap->a_vp;
983 xp = VTONULL(vp);
984 lowervp = xp->null_lowervp;
985
986 KASSERT(lowervp != NULL && vp->v_vnlock != &vp->v_lock,
987 ("Reclaiming incomplete null vnode %p", vp));
988
989 null_hashrem(xp);
990 /*
991 * Use the interlock to protect the clearing of v_data to
992 * prevent faults in null_lock().
993 */
994 lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
995 VI_LOCK(vp);
996 vp->v_data = NULL;
997 vp->v_object = NULL;
998 vp->v_vnlock = &vp->v_lock;
999
1000 /*
1001 * If we were opened for write, we leased the write reference
1002 * to the lower vnode. If this is a reclamation due to the
1003 * forced unmount, undo the reference now.
1004 */
1005 if (vp->v_writecount > 0)
1006 VOP_ADD_WRITECOUNT(lowervp, -vp->v_writecount);
1007 else if (vp->v_writecount < 0)
1008 vp->v_writecount = 0;
1009
1010 /*
1011 * Undo the effects of null_copy_inotify(): setting VIRF_INOTIFY* causes
1012 * the VFS to invoke VOP_INOTIFY on the marked vnode, and for nullfs
1013 * vnodes this is bypassed to the lower vnode. The inotify watch holds
1014 * a ref on the lower vnode, but not the upper vnode, so VOP_INOTIFY
1015 * must not be called on the upper vnode after this point.
1016 */
1017 flags = vn_irflag_read(vp) & (VIRF_INOTIFY | VIRF_INOTIFY_PARENT);
1018 if (flags != 0)
1019 vn_irflag_unset_locked(vp, flags);
1020
1021 VI_UNLOCK(vp);
1022
1023 if ((xp->null_flags & NULLV_NOUNLOCK) != 0)
1024 vunref(lowervp);
1025 else
1026 vput(lowervp);
1027 uma_zfree_smr(null_node_zone, xp);
1028
1029 return (0);
1030 }
1031
1032 static int
null_print(struct vop_print_args * ap)1033 null_print(struct vop_print_args *ap)
1034 {
1035 struct vnode *vp = ap->a_vp;
1036
1037 printf("\tvp=%p, lowervp=%p\n", vp, VTONULL(vp)->null_lowervp);
1038 return (0);
1039 }
1040
1041 /* ARGSUSED */
1042 static int
null_getwritemount(struct vop_getwritemount_args * ap)1043 null_getwritemount(struct vop_getwritemount_args *ap)
1044 {
1045 struct null_node *xp;
1046 struct vnode *lowervp;
1047 struct vnode *vp;
1048
1049 vp = ap->a_vp;
1050 VI_LOCK(vp);
1051 xp = VTONULL(vp);
1052 if (xp && (lowervp = xp->null_lowervp)) {
1053 vholdnz(lowervp);
1054 VI_UNLOCK(vp);
1055 VOP_GETWRITEMOUNT(lowervp, ap->a_mpp);
1056 vdrop(lowervp);
1057 } else {
1058 VI_UNLOCK(vp);
1059 *(ap->a_mpp) = NULL;
1060 }
1061 return (0);
1062 }
1063
1064 static int
null_vptofh(struct vop_vptofh_args * ap)1065 null_vptofh(struct vop_vptofh_args *ap)
1066 {
1067 struct vnode *lvp;
1068
1069 lvp = NULLVPTOLOWERVP(ap->a_vp);
1070 return VOP_VPTOFH(lvp, ap->a_fhp);
1071 }
1072
1073 static int
null_vptocnp(struct vop_vptocnp_args * ap)1074 null_vptocnp(struct vop_vptocnp_args *ap)
1075 {
1076 struct vnode *vp = ap->a_vp;
1077 struct vnode **dvp = ap->a_vpp;
1078 struct vnode *lvp, *ldvp;
1079 struct mount *mp;
1080 int error, locked;
1081
1082 locked = VOP_ISLOCKED(vp);
1083 lvp = NULLVPTOLOWERVP(vp);
1084 mp = vp->v_mount;
1085 error = vfs_busy(mp, MBF_NOWAIT);
1086 if (error != 0)
1087 return (error);
1088 vhold(lvp);
1089 VOP_UNLOCK(vp); /* vp is held by vn_vptocnp_locked that called us */
1090 ldvp = lvp;
1091 vref(lvp);
1092 error = vn_vptocnp(&ldvp, ap->a_buf, ap->a_buflen);
1093 vdrop(lvp);
1094 if (error != 0) {
1095 vn_lock(vp, locked | LK_RETRY);
1096 vfs_unbusy(mp);
1097 return (ENOENT);
1098 }
1099
1100 error = vn_lock(ldvp, LK_SHARED);
1101 if (error != 0) {
1102 vrele(ldvp);
1103 vn_lock(vp, locked | LK_RETRY);
1104 vfs_unbusy(mp);
1105 return (ENOENT);
1106 }
1107 error = null_nodeget(mp, ldvp, dvp);
1108 if (error == 0) {
1109 #ifdef DIAGNOSTIC
1110 NULLVPTOLOWERVP(*dvp);
1111 #endif
1112 VOP_UNLOCK(*dvp); /* keep reference on *dvp */
1113 }
1114 vn_lock(vp, locked | LK_RETRY);
1115 vfs_unbusy(mp);
1116 return (error);
1117 }
1118
1119 static int
null_read_pgcache(struct vop_read_pgcache_args * ap)1120 null_read_pgcache(struct vop_read_pgcache_args *ap)
1121 {
1122 struct vnode *lvp, *vp;
1123 struct null_node *xp;
1124 int error;
1125
1126 vp = ap->a_vp;
1127 VI_LOCK(vp);
1128 xp = VTONULL(vp);
1129 if (xp == NULL) {
1130 VI_UNLOCK(vp);
1131 return (EJUSTRETURN);
1132 }
1133 lvp = xp->null_lowervp;
1134 vref(lvp);
1135 VI_UNLOCK(vp);
1136 error = VOP_READ_PGCACHE(lvp, ap->a_uio, ap->a_ioflag, ap->a_cred);
1137 vrele(lvp);
1138 return (error);
1139 }
1140
1141 static int
null_advlock(struct vop_advlock_args * ap)1142 null_advlock(struct vop_advlock_args *ap)
1143 {
1144 struct vnode *lvp, *vp;
1145 struct null_node *xp;
1146 int error;
1147
1148 vp = ap->a_vp;
1149 VI_LOCK(vp);
1150 xp = VTONULL(vp);
1151 if (xp == NULL) {
1152 VI_UNLOCK(vp);
1153 return (EBADF);
1154 }
1155 lvp = xp->null_lowervp;
1156 vref(lvp);
1157 VI_UNLOCK(vp);
1158 error = VOP_ADVLOCK(lvp, ap->a_id, ap->a_op, ap->a_fl, ap->a_flags);
1159 vrele(lvp);
1160 return (error);
1161 }
1162
1163 /*
1164 * Avoid standard bypass, since lower dvp and vp could be no longer
1165 * valid after vput().
1166 */
1167 static int
null_vput_pair(struct vop_vput_pair_args * ap)1168 null_vput_pair(struct vop_vput_pair_args *ap)
1169 {
1170 struct mount *mp;
1171 struct vnode *dvp, *ldvp, *lvp, *vp, *vp1, **vpp;
1172 int error, res;
1173
1174 dvp = ap->a_dvp;
1175 ldvp = NULLVPTOLOWERVP(dvp);
1176 vref(ldvp);
1177
1178 vpp = ap->a_vpp;
1179 vp = NULL;
1180 lvp = NULL;
1181 mp = NULL;
1182 if (vpp != NULL)
1183 vp = *vpp;
1184 if (vp != NULL) {
1185 lvp = NULLVPTOLOWERVP(vp);
1186 vref(lvp);
1187 if (!ap->a_unlock_vp) {
1188 vhold(vp);
1189 vhold(lvp);
1190 mp = vp->v_mount;
1191 vfs_ref(mp);
1192 }
1193 }
1194
1195 res = VOP_VPUT_PAIR(ldvp, lvp != NULL ? &lvp : NULL, true);
1196 if (vp != NULL && ap->a_unlock_vp)
1197 vrele(vp);
1198 vrele(dvp);
1199
1200 if (vp == NULL || ap->a_unlock_vp)
1201 return (res);
1202
1203 /* lvp has been unlocked and vp might be reclaimed */
1204 VOP_LOCK(vp, LK_EXCLUSIVE | LK_RETRY);
1205 if (vp->v_data == NULL && vfs_busy(mp, MBF_NOWAIT) == 0) {
1206 vput(vp);
1207 vget(lvp, LK_EXCLUSIVE | LK_RETRY);
1208 if (VN_IS_DOOMED(lvp)) {
1209 vput(lvp);
1210 vget(vp, LK_EXCLUSIVE | LK_RETRY);
1211 } else {
1212 error = null_nodeget(mp, lvp, &vp1);
1213 if (error == 0) {
1214 *vpp = vp1;
1215 } else {
1216 vget(vp, LK_EXCLUSIVE | LK_RETRY);
1217 }
1218 }
1219 vfs_unbusy(mp);
1220 }
1221 vdrop(lvp);
1222 vdrop(vp);
1223 vfs_rel(mp);
1224
1225 return (res);
1226 }
1227
1228 static int
null_getlowvnode(struct vop_getlowvnode_args * ap)1229 null_getlowvnode(struct vop_getlowvnode_args *ap)
1230 {
1231 struct vnode *vp, *vpl;
1232
1233 vp = ap->a_vp;
1234 if (vn_lock(vp, LK_SHARED) != 0)
1235 return (EBADF);
1236
1237 vpl = NULLVPTOLOWERVP(vp);
1238 vhold(vpl);
1239 VOP_UNLOCK(vp);
1240 VOP_GETLOWVNODE(vpl, ap->a_vplp, ap->a_flags);
1241 vdrop(vpl);
1242 return (0);
1243 }
1244
1245 /*
1246 * Global vfs data structures
1247 */
1248 struct vop_vector null_vnodeops = {
1249 .vop_bypass = null_bypass,
1250 .vop_access = null_access,
1251 .vop_accessx = null_accessx,
1252 .vop_advlock = null_advlock,
1253 .vop_advlockpurge = vop_stdadvlockpurge,
1254 .vop_bmap = VOP_EOPNOTSUPP,
1255 .vop_stat = null_stat,
1256 .vop_getattr = null_getattr,
1257 .vop_getlowvnode = null_getlowvnode,
1258 .vop_getwritemount = null_getwritemount,
1259 .vop_inactive = null_inactive,
1260 .vop_need_inactive = null_need_inactive,
1261 .vop_islocked = vop_stdislocked,
1262 .vop_lock1 = null_lock,
1263 .vop_lookup = null_lookup,
1264 .vop_open = null_open,
1265 .vop_print = null_print,
1266 .vop_read_pgcache = null_read_pgcache,
1267 .vop_reclaim = null_reclaim,
1268 .vop_remove = null_remove,
1269 .vop_rename = null_rename,
1270 .vop_rmdir = null_rmdir,
1271 .vop_setattr = null_setattr,
1272 .vop_strategy = VOP_EOPNOTSUPP,
1273 .vop_unlock = null_unlock,
1274 .vop_vptocnp = null_vptocnp,
1275 .vop_vptofh = null_vptofh,
1276 .vop_add_writecount = null_add_writecount,
1277 .vop_vput_pair = null_vput_pair,
1278 .vop_copy_file_range = VOP_PANIC,
1279 };
1280 VFS_VOP_VECTOR_REGISTER(null_vnodeops);
1281
1282 struct vop_vector null_vnodeops_no_unp_bypass = {
1283 .vop_default = &null_vnodeops,
1284 .vop_unp_bind = vop_stdunp_bind,
1285 .vop_unp_connect = vop_stdunp_connect,
1286 .vop_unp_detach = vop_stdunp_detach,
1287 };
1288 VFS_VOP_VECTOR_REGISTER(null_vnodeops_no_unp_bypass);
1289