1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
24  * Copyright 2015, OmniTI Computer Consulting, Inc. All rights reserved.
25  */
26 
27 /*
28  * ZFS control directory (a.k.a. ".zfs")
29  *
30  * This directory provides a common location for all ZFS meta-objects.
31  * Currently, this is only the 'snapshot' directory, but this may expand in the
32  * future.  The elements are built using the GFS primitives, as the hierarchy
33  * does not actually exist on disk.
34  *
35  * For 'snapshot', we don't want to have all snapshots always mounted, because
36  * this would take up a huge amount of space in /etc/mnttab.  We have three
37  * types of objects:
38  *
39  * 	ctldir ------> snapshotdir -------> snapshot
40  *                                             |
41  *                                             |
42  *                                             V
43  *                                         mounted fs
44  *
45  * The 'snapshot' node contains just enough information to lookup '..' and act
46  * as a mountpoint for the snapshot.  Whenever we lookup a specific snapshot, we
47  * perform an automount of the underlying filesystem and return the
48  * corresponding vnode.
49  *
50  * All mounts are handled automatically by the kernel, but unmounts are
51  * (currently) handled from user land.  The main reason is that there is no
52  * reliable way to auto-unmount the filesystem when it's "no longer in use".
53  * When the user unmounts a filesystem, we call zfsctl_unmount(), which
54  * unmounts any snapshots within the snapshot directory.
55  *
56  * The '.zfs', '.zfs/snapshot', and all directories created under
57  * '.zfs/snapshot' (ie: '.zfs/snapshot/<snapname>') are all GFS nodes and
58  * share the same vfs_t as the head filesystem (what '.zfs' lives under).
59  *
60  * File systems mounted ontop of the GFS nodes '.zfs/snapshot/<snapname>'
61  * (ie: snapshots) are ZFS nodes and have their own unique vfs_t.
62  * However, vnodes within these mounted on file systems have their v_vfsp
63  * fields set to the head filesystem to make NFS happy (see
64  * zfsctl_snapdir_lookup()). We VFS_HOLD the head filesystem's vfs_t
65  * so that it cannot be freed until all snapshots have been unmounted.
66  */
67 
68 #include <sys/types.h>
69 #include <sys/param.h>
70 #include <sys/libkern.h>
71 #include <sys/dirent.h>
72 #include <sys/zfs_context.h>
73 #include <sys/zfs_ctldir.h>
74 #include <sys/zfs_ioctl.h>
75 #include <sys/zfs_vfsops.h>
76 #include <sys/namei.h>
77 #include <sys/stat.h>
78 #include <sys/dmu.h>
79 #include <sys/dsl_dataset.h>
80 #include <sys/dsl_destroy.h>
81 #include <sys/dsl_deleg.h>
82 #include <sys/mount.h>
83 #include <sys/zap.h>
84 #include <sys/sysproto.h>
85 
86 #include "zfs_namecheck.h"
87 
88 #include <sys/kernel.h>
89 #include <sys/ccompat.h>
90 
91 /* Common access mode for all virtual directories under the ctldir */
92 const uint16_t zfsctl_ctldir_mode = S_IRUSR | S_IXUSR | S_IRGRP | S_IXGRP |
93     S_IROTH | S_IXOTH;
94 
95 /*
96  * "Synthetic" filesystem implementation.
97  */
98 
99 /*
100  * Assert that A implies B.
101  */
102 #define	KASSERT_IMPLY(A, B, msg)	KASSERT(!(A) || (B), (msg));
103 
104 static MALLOC_DEFINE(M_SFSNODES, "sfs_nodes", "synthetic-fs nodes");
105 
106 typedef struct sfs_node {
107 	char		sn_name[ZFS_MAX_DATASET_NAME_LEN];
108 	uint64_t	sn_parent_id;
109 	uint64_t	sn_id;
110 } sfs_node_t;
111 
112 /*
113  * Check the parent's ID as well as the node's to account for a chance
114  * that IDs originating from different domains (snapshot IDs, artificial
115  * IDs, znode IDs) may clash.
116  */
117 static int
118 sfs_compare_ids(struct vnode *vp, void *arg)
119 {
120 	sfs_node_t *n1 = vp->v_data;
121 	sfs_node_t *n2 = arg;
122 	bool equal;
123 
124 	equal = n1->sn_id == n2->sn_id &&
125 	    n1->sn_parent_id == n2->sn_parent_id;
126 
127 	/* Zero means equality. */
128 	return (!equal);
129 }
130 
131 static int
132 sfs_vnode_get(const struct mount *mp, int flags, uint64_t parent_id,
133     uint64_t id, struct vnode **vpp)
134 {
135 	sfs_node_t search;
136 	int err;
137 
138 	search.sn_id = id;
139 	search.sn_parent_id = parent_id;
140 	err = vfs_hash_get(mp, (uint32_t)id, flags, curthread, vpp,
141 	    sfs_compare_ids, &search);
142 	return (err);
143 }
144 
145 static int
146 sfs_vnode_insert(struct vnode *vp, int flags, uint64_t parent_id,
147     uint64_t id, struct vnode **vpp)
148 {
149 	int err;
150 
151 	KASSERT(vp->v_data != NULL, ("sfs_vnode_insert with NULL v_data"));
152 	err = vfs_hash_insert(vp, (uint32_t)id, flags, curthread, vpp,
153 	    sfs_compare_ids, vp->v_data);
154 	return (err);
155 }
156 
157 static void
158 sfs_vnode_remove(struct vnode *vp)
159 {
160 	vfs_hash_remove(vp);
161 }
162 
163 typedef void sfs_vnode_setup_fn(vnode_t *vp, void *arg);
164 
165 static int
166 sfs_vgetx(struct mount *mp, int flags, uint64_t parent_id, uint64_t id,
167     const char *tag, struct vop_vector *vops,
168     sfs_vnode_setup_fn setup, void *arg,
169     struct vnode **vpp)
170 {
171 	struct vnode *vp;
172 	int error;
173 
174 	error = sfs_vnode_get(mp, flags, parent_id, id, vpp);
175 	if (error != 0 || *vpp != NULL) {
176 		KASSERT_IMPLY(error == 0, (*vpp)->v_data != NULL,
177 		    "sfs vnode with no data");
178 		return (error);
179 	}
180 
181 	/* Allocate a new vnode/inode. */
182 	error = getnewvnode(tag, mp, vops, &vp);
183 	if (error != 0) {
184 		*vpp = NULL;
185 		return (error);
186 	}
187 
188 	/*
189 	 * Exclusively lock the vnode vnode while it's being constructed.
190 	 */
191 	lockmgr(vp->v_vnlock, LK_EXCLUSIVE, NULL);
192 	error = insmntque(vp, mp);
193 	if (error != 0) {
194 		*vpp = NULL;
195 		return (error);
196 	}
197 
198 	setup(vp, arg);
199 
200 	error = sfs_vnode_insert(vp, flags, parent_id, id, vpp);
201 	if (error != 0 || *vpp != NULL) {
202 		KASSERT_IMPLY(error == 0, (*vpp)->v_data != NULL,
203 		    "sfs vnode with no data");
204 		return (error);
205 	}
206 
207 	*vpp = vp;
208 	return (0);
209 }
210 
211 static void
212 sfs_print_node(sfs_node_t *node)
213 {
214 	printf("\tname = %s\n", node->sn_name);
215 	printf("\tparent_id = %ju\n", (uintmax_t)node->sn_parent_id);
216 	printf("\tid = %ju\n", (uintmax_t)node->sn_id);
217 }
218 
219 static sfs_node_t *
220 sfs_alloc_node(size_t size, const char *name, uint64_t parent_id, uint64_t id)
221 {
222 	struct sfs_node *node;
223 
224 	KASSERT(strlen(name) < sizeof (node->sn_name),
225 	    ("sfs node name is too long"));
226 	KASSERT(size >= sizeof (*node), ("sfs node size is too small"));
227 	node = malloc(size, M_SFSNODES, M_WAITOK | M_ZERO);
228 	strlcpy(node->sn_name, name, sizeof (node->sn_name));
229 	node->sn_parent_id = parent_id;
230 	node->sn_id = id;
231 
232 	return (node);
233 }
234 
235 static void
236 sfs_destroy_node(sfs_node_t *node)
237 {
238 	free(node, M_SFSNODES);
239 }
240 
241 static void *
242 sfs_reclaim_vnode(vnode_t *vp)
243 {
244 	void *data;
245 
246 	sfs_vnode_remove(vp);
247 	data = vp->v_data;
248 	vp->v_data = NULL;
249 	return (data);
250 }
251 
252 static int
253 sfs_readdir_common(uint64_t parent_id, uint64_t id, struct vop_readdir_args *ap,
254     uio_t *uio, off_t *offp)
255 {
256 	struct dirent entry;
257 	int error;
258 
259 	/* Reset ncookies for subsequent use of vfs_read_dirent. */
260 	if (ap->a_ncookies != NULL)
261 		*ap->a_ncookies = 0;
262 
263 	if (uio->uio_resid < sizeof (entry))
264 		return (SET_ERROR(EINVAL));
265 
266 	if (uio->uio_offset < 0)
267 		return (SET_ERROR(EINVAL));
268 	if (uio->uio_offset == 0) {
269 		entry.d_fileno = id;
270 		entry.d_type = DT_DIR;
271 		entry.d_name[0] = '.';
272 		entry.d_name[1] = '\0';
273 		entry.d_namlen = 1;
274 		entry.d_reclen = sizeof (entry);
275 		error = vfs_read_dirent(ap, &entry, uio->uio_offset);
276 		if (error != 0)
277 			return (SET_ERROR(error));
278 	}
279 
280 	if (uio->uio_offset < sizeof (entry))
281 		return (SET_ERROR(EINVAL));
282 	if (uio->uio_offset == sizeof (entry)) {
283 		entry.d_fileno = parent_id;
284 		entry.d_type = DT_DIR;
285 		entry.d_name[0] = '.';
286 		entry.d_name[1] = '.';
287 		entry.d_name[2] = '\0';
288 		entry.d_namlen = 2;
289 		entry.d_reclen = sizeof (entry);
290 		error = vfs_read_dirent(ap, &entry, uio->uio_offset);
291 		if (error != 0)
292 			return (SET_ERROR(error));
293 	}
294 
295 	if (offp != NULL)
296 		*offp = 2 * sizeof (entry);
297 	return (0);
298 }
299 
300 
301 /*
302  * .zfs inode namespace
303  *
304  * We need to generate unique inode numbers for all files and directories
305  * within the .zfs pseudo-filesystem.  We use the following scheme:
306  *
307  * 	ENTRY			ZFSCTL_INODE
308  * 	.zfs			1
309  * 	.zfs/snapshot		2
310  * 	.zfs/snapshot/<snap>	objectid(snap)
311  */
312 #define	ZFSCTL_INO_SNAP(id)	(id)
313 
314 static struct vop_vector zfsctl_ops_root;
315 static struct vop_vector zfsctl_ops_snapdir;
316 static struct vop_vector zfsctl_ops_snapshot;
317 
318 void
319 zfsctl_init(void)
320 {
321 }
322 
323 void
324 zfsctl_fini(void)
325 {
326 }
327 
328 boolean_t
329 zfsctl_is_node(vnode_t *vp)
330 {
331 	return (vn_matchops(vp, zfsctl_ops_root) ||
332 	    vn_matchops(vp, zfsctl_ops_snapdir) ||
333 	    vn_matchops(vp, zfsctl_ops_snapshot));
334 
335 }
336 
337 typedef struct zfsctl_root {
338 	sfs_node_t	node;
339 	sfs_node_t	*snapdir;
340 	timestruc_t	cmtime;
341 } zfsctl_root_t;
342 
343 
344 /*
345  * Create the '.zfs' directory.
346  */
347 void
348 zfsctl_create(zfsvfs_t *zfsvfs)
349 {
350 	zfsctl_root_t *dot_zfs;
351 	sfs_node_t *snapdir;
352 	vnode_t *rvp;
353 	uint64_t crtime[2];
354 
355 	ASSERT(zfsvfs->z_ctldir == NULL);
356 
357 	snapdir = sfs_alloc_node(sizeof (*snapdir), "snapshot", ZFSCTL_INO_ROOT,
358 	    ZFSCTL_INO_SNAPDIR);
359 	dot_zfs = (zfsctl_root_t *)sfs_alloc_node(sizeof (*dot_zfs), ".zfs", 0,
360 	    ZFSCTL_INO_ROOT);
361 	dot_zfs->snapdir = snapdir;
362 
363 	VERIFY(VFS_ROOT(zfsvfs->z_vfs, LK_EXCLUSIVE, &rvp) == 0);
364 	VERIFY(0 == sa_lookup(VTOZ(rvp)->z_sa_hdl, SA_ZPL_CRTIME(zfsvfs),
365 	    &crtime, sizeof (crtime)));
366 	ZFS_TIME_DECODE(&dot_zfs->cmtime, crtime);
367 	vput(rvp);
368 
369 	zfsvfs->z_ctldir = dot_zfs;
370 }
371 
372 /*
373  * Destroy the '.zfs' directory.  Only called when the filesystem is unmounted.
374  * The nodes must not have any associated vnodes by now as they should be
375  * vflush-ed.
376  */
377 void
378 zfsctl_destroy(zfsvfs_t *zfsvfs)
379 {
380 	sfs_destroy_node(zfsvfs->z_ctldir->snapdir);
381 	sfs_destroy_node((sfs_node_t *)zfsvfs->z_ctldir);
382 	zfsvfs->z_ctldir = NULL;
383 }
384 
385 static int
386 zfsctl_fs_root_vnode(struct mount *mp, void *arg __unused, int flags,
387     struct vnode **vpp)
388 {
389 	return (VFS_ROOT(mp, flags, vpp));
390 }
391 
392 static void
393 zfsctl_common_vnode_setup(vnode_t *vp, void *arg)
394 {
395 	ASSERT_VOP_ELOCKED(vp, __func__);
396 
397 	/* We support shared locking. */
398 	VN_LOCK_ASHARE(vp);
399 	vp->v_type = VDIR;
400 	vp->v_data = arg;
401 }
402 
403 static int
404 zfsctl_root_vnode(struct mount *mp, void *arg __unused, int flags,
405     struct vnode **vpp)
406 {
407 	void *node;
408 	int err;
409 
410 	node = ((zfsvfs_t *)mp->mnt_data)->z_ctldir;
411 	err = sfs_vgetx(mp, flags, 0, ZFSCTL_INO_ROOT, "zfs", &zfsctl_ops_root,
412 	    zfsctl_common_vnode_setup, node, vpp);
413 	return (err);
414 }
415 
416 static int
417 zfsctl_snapdir_vnode(struct mount *mp, void *arg __unused, int flags,
418     struct vnode **vpp)
419 {
420 	void *node;
421 	int err;
422 
423 	node = ((zfsvfs_t *)mp->mnt_data)->z_ctldir->snapdir;
424 	err = sfs_vgetx(mp, flags, ZFSCTL_INO_ROOT, ZFSCTL_INO_SNAPDIR, "zfs",
425 	    &zfsctl_ops_snapdir, zfsctl_common_vnode_setup, node, vpp);
426 	return (err);
427 }
428 
429 /*
430  * Given a root znode, retrieve the associated .zfs directory.
431  * Add a hold to the vnode and return it.
432  */
433 int
434 zfsctl_root(zfsvfs_t *zfsvfs, int flags, vnode_t **vpp)
435 {
436 	int error;
437 
438 	error = zfsctl_root_vnode(zfsvfs->z_vfs, NULL, flags, vpp);
439 	return (error);
440 }
441 
442 /*
443  * Common open routine.  Disallow any write access.
444  */
445 static int
446 zfsctl_common_open(struct vop_open_args *ap)
447 {
448 	int flags = ap->a_mode;
449 
450 	if (flags & FWRITE)
451 		return (SET_ERROR(EACCES));
452 
453 	return (0);
454 }
455 
456 /*
457  * Common close routine.  Nothing to do here.
458  */
459 /* ARGSUSED */
460 static int
461 zfsctl_common_close(struct vop_close_args *ap)
462 {
463 	return (0);
464 }
465 
466 /*
467  * Common access routine.  Disallow writes.
468  */
469 static int
470 zfsctl_common_access(struct vop_access_args *ap)
471 {
472 	accmode_t accmode = ap->a_accmode;
473 
474 	if (accmode & VWRITE)
475 		return (SET_ERROR(EACCES));
476 	return (0);
477 }
478 
479 /*
480  * Common getattr function.  Fill in basic information.
481  */
482 static void
483 zfsctl_common_getattr(vnode_t *vp, vattr_t *vap)
484 {
485 	timestruc_t	now;
486 	sfs_node_t *node;
487 
488 	node = vp->v_data;
489 
490 	vap->va_uid = 0;
491 	vap->va_gid = 0;
492 	vap->va_rdev = 0;
493 	/*
494 	 * We are a purely virtual object, so we have no
495 	 * blocksize or allocated blocks.
496 	 */
497 	vap->va_blksize = 0;
498 	vap->va_nblocks = 0;
499 	vap->va_seq = 0;
500 	vn_fsid(vp, vap);
501 	vap->va_mode = zfsctl_ctldir_mode;
502 	vap->va_type = VDIR;
503 	/*
504 	 * We live in the now (for atime).
505 	 */
506 	gethrestime(&now);
507 	vap->va_atime = now;
508 	/* FreeBSD: Reset chflags(2) flags. */
509 	vap->va_flags = 0;
510 
511 	vap->va_nodeid = node->sn_id;
512 
513 	/* At least '.' and '..'. */
514 	vap->va_nlink = 2;
515 }
516 
517 #ifndef _OPENSOLARIS_SYS_VNODE_H_
518 struct vop_fid_args {
519 	struct vnode *a_vp;
520 	struct fid *a_fid;
521 };
522 #endif
523 
524 static int
525 zfsctl_common_fid(struct vop_fid_args *ap)
526 {
527 	vnode_t		*vp = ap->a_vp;
528 	fid_t		*fidp = (void *)ap->a_fid;
529 	sfs_node_t	*node = vp->v_data;
530 	uint64_t	object = node->sn_id;
531 	zfid_short_t	*zfid;
532 	int		i;
533 
534 	zfid = (zfid_short_t *)fidp;
535 	zfid->zf_len = SHORT_FID_LEN;
536 
537 	for (i = 0; i < sizeof (zfid->zf_object); i++)
538 		zfid->zf_object[i] = (uint8_t)(object >> (8 * i));
539 
540 	/* .zfs nodes always have a generation number of 0 */
541 	for (i = 0; i < sizeof (zfid->zf_gen); i++)
542 		zfid->zf_gen[i] = 0;
543 
544 	return (0);
545 }
546 
547 #ifndef _SYS_SYSPROTO_H_
548 struct vop_reclaim_args {
549 	struct vnode *a_vp;
550 	struct thread *a_td;
551 };
552 #endif
553 
554 static int
555 zfsctl_common_reclaim(struct vop_reclaim_args *ap)
556 {
557 	vnode_t *vp = ap->a_vp;
558 
559 	(void) sfs_reclaim_vnode(vp);
560 	return (0);
561 }
562 
563 #ifndef _SYS_SYSPROTO_H_
564 struct vop_print_args {
565 	struct vnode *a_vp;
566 };
567 #endif
568 
569 static int
570 zfsctl_common_print(struct vop_print_args *ap)
571 {
572 	sfs_print_node(ap->a_vp->v_data);
573 	return (0);
574 }
575 
576 #ifndef _SYS_SYSPROTO_H_
577 struct vop_getattr_args {
578 	struct vnode *a_vp;
579 	struct vattr *a_vap;
580 	struct ucred *a_cred;
581 };
582 #endif
583 
584 /*
585  * Get root directory attributes.
586  */
587 static int
588 zfsctl_root_getattr(struct vop_getattr_args *ap)
589 {
590 	struct vnode *vp = ap->a_vp;
591 	struct vattr *vap = ap->a_vap;
592 	zfsctl_root_t *node = vp->v_data;
593 
594 	zfsctl_common_getattr(vp, vap);
595 	vap->va_ctime = node->cmtime;
596 	vap->va_mtime = vap->va_ctime;
597 	vap->va_birthtime = vap->va_ctime;
598 	vap->va_nlink += 1; /* snapdir */
599 	vap->va_size = vap->va_nlink;
600 	return (0);
601 }
602 
603 /*
604  * When we lookup "." we still can be asked to lock it
605  * differently, can't we?
606  */
607 static int
608 zfsctl_relock_dot(vnode_t *dvp, int ltype)
609 {
610 	vref(dvp);
611 	if (ltype != VOP_ISLOCKED(dvp)) {
612 		if (ltype == LK_EXCLUSIVE)
613 			vn_lock(dvp, LK_UPGRADE | LK_RETRY);
614 		else /* if (ltype == LK_SHARED) */
615 			vn_lock(dvp, LK_DOWNGRADE | LK_RETRY);
616 
617 		/* Relock for the "." case may left us with reclaimed vnode. */
618 		if (VN_IS_DOOMED(dvp)) {
619 			vrele(dvp);
620 			return (SET_ERROR(ENOENT));
621 		}
622 	}
623 	return (0);
624 }
625 
626 /*
627  * Special case the handling of "..".
628  */
629 static int
630 zfsctl_root_lookup(struct vop_lookup_args *ap)
631 {
632 	struct componentname *cnp = ap->a_cnp;
633 	vnode_t *dvp = ap->a_dvp;
634 	vnode_t **vpp = ap->a_vpp;
635 	int flags = ap->a_cnp->cn_flags;
636 	int lkflags = ap->a_cnp->cn_lkflags;
637 	int nameiop = ap->a_cnp->cn_nameiop;
638 	int err;
639 
640 	ASSERT(dvp->v_type == VDIR);
641 
642 	if ((flags & ISLASTCN) != 0 && nameiop != LOOKUP)
643 		return (SET_ERROR(ENOTSUP));
644 
645 	if (cnp->cn_namelen == 1 && *cnp->cn_nameptr == '.') {
646 		err = zfsctl_relock_dot(dvp, lkflags & LK_TYPE_MASK);
647 		if (err == 0)
648 			*vpp = dvp;
649 	} else if ((flags & ISDOTDOT) != 0) {
650 		err = vn_vget_ino_gen(dvp, zfsctl_fs_root_vnode, NULL,
651 		    lkflags, vpp);
652 	} else if (strncmp(cnp->cn_nameptr, "snapshot", cnp->cn_namelen) == 0) {
653 		err = zfsctl_snapdir_vnode(dvp->v_mount, NULL, lkflags, vpp);
654 	} else {
655 		err = SET_ERROR(ENOENT);
656 	}
657 	if (err != 0)
658 		*vpp = NULL;
659 	return (err);
660 }
661 
662 static int
663 zfsctl_root_readdir(struct vop_readdir_args *ap)
664 {
665 	struct dirent entry;
666 	vnode_t *vp = ap->a_vp;
667 	zfsvfs_t *zfsvfs = vp->v_vfsp->vfs_data;
668 	zfsctl_root_t *node = vp->v_data;
669 	uio_t *uio = ap->a_uio;
670 	int *eofp = ap->a_eofflag;
671 	off_t dots_offset;
672 	int error;
673 
674 	ASSERT(vp->v_type == VDIR);
675 
676 	error = sfs_readdir_common(zfsvfs->z_root, ZFSCTL_INO_ROOT, ap, uio,
677 	    &dots_offset);
678 	if (error != 0) {
679 		if (error == ENAMETOOLONG) /* ran out of destination space */
680 			error = 0;
681 		return (error);
682 	}
683 	if (uio->uio_offset != dots_offset)
684 		return (SET_ERROR(EINVAL));
685 
686 	CTASSERT(sizeof (node->snapdir->sn_name) <= sizeof (entry.d_name));
687 	entry.d_fileno = node->snapdir->sn_id;
688 	entry.d_type = DT_DIR;
689 	strcpy(entry.d_name, node->snapdir->sn_name);
690 	entry.d_namlen = strlen(entry.d_name);
691 	entry.d_reclen = sizeof (entry);
692 	error = vfs_read_dirent(ap, &entry, uio->uio_offset);
693 	if (error != 0) {
694 		if (error == ENAMETOOLONG)
695 			error = 0;
696 		return (SET_ERROR(error));
697 	}
698 	if (eofp != NULL)
699 		*eofp = 1;
700 	return (0);
701 }
702 
703 static int
704 zfsctl_root_vptocnp(struct vop_vptocnp_args *ap)
705 {
706 	static const char dotzfs_name[4] = ".zfs";
707 	vnode_t *dvp;
708 	int error;
709 
710 	if (*ap->a_buflen < sizeof (dotzfs_name))
711 		return (SET_ERROR(ENOMEM));
712 
713 	error = vn_vget_ino_gen(ap->a_vp, zfsctl_fs_root_vnode, NULL,
714 	    LK_SHARED, &dvp);
715 	if (error != 0)
716 		return (SET_ERROR(error));
717 
718 	VOP_UNLOCK1(dvp);
719 	*ap->a_vpp = dvp;
720 	*ap->a_buflen -= sizeof (dotzfs_name);
721 	bcopy(dotzfs_name, ap->a_buf + *ap->a_buflen, sizeof (dotzfs_name));
722 	return (0);
723 }
724 
725 static int
726 zfsctl_common_pathconf(struct vop_pathconf_args *ap)
727 {
728 	/*
729 	 * We care about ACL variables so that user land utilities like ls
730 	 * can display them correctly.  Since the ctldir's st_dev is set to be
731 	 * the same as the parent dataset, we must support all variables that
732 	 * it supports.
733 	 */
734 	switch (ap->a_name) {
735 	case _PC_LINK_MAX:
736 		*ap->a_retval = MIN(LONG_MAX, ZFS_LINK_MAX);
737 		return (0);
738 
739 	case _PC_FILESIZEBITS:
740 		*ap->a_retval = 64;
741 		return (0);
742 
743 	case _PC_MIN_HOLE_SIZE:
744 		*ap->a_retval = (int)SPA_MINBLOCKSIZE;
745 		return (0);
746 
747 	case _PC_ACL_EXTENDED:
748 		*ap->a_retval = 0;
749 		return (0);
750 
751 	case _PC_ACL_NFS4:
752 		*ap->a_retval = 1;
753 		return (0);
754 
755 	case _PC_ACL_PATH_MAX:
756 		*ap->a_retval = ACL_MAX_ENTRIES;
757 		return (0);
758 
759 	case _PC_NAME_MAX:
760 		*ap->a_retval = NAME_MAX;
761 		return (0);
762 
763 	default:
764 		return (vop_stdpathconf(ap));
765 	}
766 }
767 
768 /*
769  * Returns a trivial ACL
770  */
771 static int
772 zfsctl_common_getacl(struct vop_getacl_args *ap)
773 {
774 	int i;
775 
776 	if (ap->a_type != ACL_TYPE_NFS4)
777 		return (EINVAL);
778 
779 	acl_nfs4_sync_acl_from_mode(ap->a_aclp, zfsctl_ctldir_mode, 0);
780 	/*
781 	 * acl_nfs4_sync_acl_from_mode assumes that the owner can always modify
782 	 * attributes.  That is not the case for the ctldir, so we must clear
783 	 * those bits.  We also must clear ACL_READ_NAMED_ATTRS, because xattrs
784 	 * aren't supported by the ctldir.
785 	 */
786 	for (i = 0; i < ap->a_aclp->acl_cnt; i++) {
787 		struct acl_entry *entry;
788 		entry = &(ap->a_aclp->acl_entry[i]);
789 		entry->ae_perm &= ~(ACL_WRITE_ACL | ACL_WRITE_OWNER |
790 		    ACL_WRITE_ATTRIBUTES | ACL_WRITE_NAMED_ATTRS |
791 		    ACL_READ_NAMED_ATTRS);
792 	}
793 
794 	return (0);
795 }
796 
797 static struct vop_vector zfsctl_ops_root = {
798 	.vop_default =	&default_vnodeops,
799 #if __FreeBSD_version >= 1300121
800 	.vop_fplookup_vexec = VOP_EAGAIN,
801 #endif
802 	.vop_open =	zfsctl_common_open,
803 	.vop_close =	zfsctl_common_close,
804 	.vop_ioctl =	VOP_EINVAL,
805 	.vop_getattr =	zfsctl_root_getattr,
806 	.vop_access =	zfsctl_common_access,
807 	.vop_readdir =	zfsctl_root_readdir,
808 	.vop_lookup =	zfsctl_root_lookup,
809 	.vop_inactive =	VOP_NULL,
810 	.vop_reclaim =	zfsctl_common_reclaim,
811 	.vop_fid =	zfsctl_common_fid,
812 	.vop_print =	zfsctl_common_print,
813 	.vop_vptocnp =	zfsctl_root_vptocnp,
814 	.vop_pathconf =	zfsctl_common_pathconf,
815 	.vop_getacl =	zfsctl_common_getacl,
816 };
817 VFS_VOP_VECTOR_REGISTER(zfsctl_ops_root);
818 
819 static int
820 zfsctl_snapshot_zname(vnode_t *vp, const char *name, int len, char *zname)
821 {
822 	objset_t *os = ((zfsvfs_t *)((vp)->v_vfsp->vfs_data))->z_os;
823 
824 	dmu_objset_name(os, zname);
825 	if (strlen(zname) + 1 + strlen(name) >= len)
826 		return (SET_ERROR(ENAMETOOLONG));
827 	(void) strcat(zname, "@");
828 	(void) strcat(zname, name);
829 	return (0);
830 }
831 
832 static int
833 zfsctl_snapshot_lookup(vnode_t *vp, const char *name, uint64_t *id)
834 {
835 	objset_t *os = ((zfsvfs_t *)((vp)->v_vfsp->vfs_data))->z_os;
836 	int err;
837 
838 	err = dsl_dataset_snap_lookup(dmu_objset_ds(os), name, id);
839 	return (err);
840 }
841 
842 /*
843  * Given a vnode get a root vnode of a filesystem mounted on top of
844  * the vnode, if any.  The root vnode is referenced and locked.
845  * If no filesystem is mounted then the orinal vnode remains referenced
846  * and locked.  If any error happens the orinal vnode is unlocked and
847  * released.
848  */
849 static int
850 zfsctl_mounted_here(vnode_t **vpp, int flags)
851 {
852 	struct mount *mp;
853 	int err;
854 
855 	ASSERT_VOP_LOCKED(*vpp, __func__);
856 	ASSERT3S((*vpp)->v_type, ==, VDIR);
857 
858 	if ((mp = (*vpp)->v_mountedhere) != NULL) {
859 		err = vfs_busy(mp, 0);
860 		KASSERT(err == 0, ("vfs_busy(mp, 0) failed with %d", err));
861 		KASSERT(vrefcnt(*vpp) > 1, ("unreferenced mountpoint"));
862 		vput(*vpp);
863 		err = VFS_ROOT(mp, flags, vpp);
864 		vfs_unbusy(mp);
865 		return (err);
866 	}
867 	return (EJUSTRETURN);
868 }
869 
870 typedef struct {
871 	const char *snap_name;
872 	uint64_t    snap_id;
873 } snapshot_setup_arg_t;
874 
875 static void
876 zfsctl_snapshot_vnode_setup(vnode_t *vp, void *arg)
877 {
878 	snapshot_setup_arg_t *ssa = arg;
879 	sfs_node_t *node;
880 
881 	ASSERT_VOP_ELOCKED(vp, __func__);
882 
883 	node = sfs_alloc_node(sizeof (sfs_node_t),
884 	    ssa->snap_name, ZFSCTL_INO_SNAPDIR, ssa->snap_id);
885 	zfsctl_common_vnode_setup(vp, node);
886 
887 	/* We have to support recursive locking. */
888 	VN_LOCK_AREC(vp);
889 }
890 
891 /*
892  * Lookup entry point for the 'snapshot' directory.  Try to open the
893  * snapshot if it exist, creating the pseudo filesystem vnode as necessary.
894  * Perform a mount of the associated dataset on top of the vnode.
895  * There are four possibilities:
896  * - the snapshot node and vnode do not exist
897  * - the snapshot vnode is covered by the mounted snapshot
898  * - the snapshot vnode is not covered yet, the mount operation is in progress
899  * - the snapshot vnode is not covered, because the snapshot has been unmounted
900  * The last two states are transient and should be relatively short-lived.
901  */
902 static int
903 zfsctl_snapdir_lookup(struct vop_lookup_args *ap)
904 {
905 	vnode_t *dvp = ap->a_dvp;
906 	vnode_t **vpp = ap->a_vpp;
907 	struct componentname *cnp = ap->a_cnp;
908 	char name[NAME_MAX + 1];
909 	char fullname[ZFS_MAX_DATASET_NAME_LEN];
910 	char *mountpoint;
911 	size_t mountpoint_len;
912 	zfsvfs_t *zfsvfs = dvp->v_vfsp->vfs_data;
913 	uint64_t snap_id;
914 	int nameiop = cnp->cn_nameiop;
915 	int lkflags = cnp->cn_lkflags;
916 	int flags = cnp->cn_flags;
917 	int err;
918 
919 	ASSERT(dvp->v_type == VDIR);
920 
921 	if ((flags & ISLASTCN) != 0 && nameiop != LOOKUP)
922 		return (SET_ERROR(ENOTSUP));
923 
924 	if (cnp->cn_namelen == 1 && *cnp->cn_nameptr == '.') {
925 		err = zfsctl_relock_dot(dvp, lkflags & LK_TYPE_MASK);
926 		if (err == 0)
927 			*vpp = dvp;
928 		return (err);
929 	}
930 	if (flags & ISDOTDOT) {
931 		err = vn_vget_ino_gen(dvp, zfsctl_root_vnode, NULL, lkflags,
932 		    vpp);
933 		return (err);
934 	}
935 
936 	if (cnp->cn_namelen >= sizeof (name))
937 		return (SET_ERROR(ENAMETOOLONG));
938 
939 	strlcpy(name, ap->a_cnp->cn_nameptr, ap->a_cnp->cn_namelen + 1);
940 	err = zfsctl_snapshot_lookup(dvp, name, &snap_id);
941 	if (err != 0)
942 		return (SET_ERROR(ENOENT));
943 
944 	for (;;) {
945 		snapshot_setup_arg_t ssa;
946 
947 		ssa.snap_name = name;
948 		ssa.snap_id = snap_id;
949 		err = sfs_vgetx(dvp->v_mount, LK_SHARED, ZFSCTL_INO_SNAPDIR,
950 		    snap_id, "zfs", &zfsctl_ops_snapshot,
951 		    zfsctl_snapshot_vnode_setup, &ssa, vpp);
952 		if (err != 0)
953 			return (err);
954 
955 		/* Check if a new vnode has just been created. */
956 		if (VOP_ISLOCKED(*vpp) == LK_EXCLUSIVE)
957 			break;
958 
959 		/*
960 		 * Check if a snapshot is already mounted on top of the vnode.
961 		 */
962 		err = zfsctl_mounted_here(vpp, lkflags);
963 		if (err != EJUSTRETURN)
964 			return (err);
965 
966 		/*
967 		 * If the vnode is not covered, then either the mount operation
968 		 * is in progress or the snapshot has already been unmounted
969 		 * but the vnode hasn't been inactivated and reclaimed yet.
970 		 * We can try to re-use the vnode in the latter case.
971 		 */
972 		VI_LOCK(*vpp);
973 		if (((*vpp)->v_iflag & VI_MOUNT) == 0) {
974 			/*
975 			 * Upgrade to exclusive lock in order to:
976 			 * - avoid race conditions
977 			 * - satisfy the contract of mount_snapshot()
978 			 */
979 			err = VOP_LOCK(*vpp, LK_TRYUPGRADE | LK_INTERLOCK);
980 			if (err == 0)
981 				break;
982 		} else {
983 			VI_UNLOCK(*vpp);
984 		}
985 
986 		/*
987 		 * In this state we can loop on uncontested locks and starve
988 		 * the thread doing the lengthy, non-trivial mount operation.
989 		 * So, yield to prevent that from happening.
990 		 */
991 		vput(*vpp);
992 		kern_yield(PRI_USER);
993 	}
994 
995 	VERIFY0(zfsctl_snapshot_zname(dvp, name, sizeof (fullname), fullname));
996 
997 	mountpoint_len = strlen(dvp->v_vfsp->mnt_stat.f_mntonname) +
998 	    strlen("/" ZFS_CTLDIR_NAME "/snapshot/") + strlen(name) + 1;
999 	mountpoint = kmem_alloc(mountpoint_len, KM_SLEEP);
1000 	(void) snprintf(mountpoint, mountpoint_len,
1001 	    "%s/" ZFS_CTLDIR_NAME "/snapshot/%s",
1002 	    dvp->v_vfsp->mnt_stat.f_mntonname, name);
1003 
1004 	err = mount_snapshot(curthread, vpp, "zfs", mountpoint, fullname, 0);
1005 	kmem_free(mountpoint, mountpoint_len);
1006 	if (err == 0) {
1007 		/*
1008 		 * Fix up the root vnode mounted on .zfs/snapshot/<snapname>.
1009 		 *
1010 		 * This is where we lie about our v_vfsp in order to
1011 		 * make .zfs/snapshot/<snapname> accessible over NFS
1012 		 * without requiring manual mounts of <snapname>.
1013 		 */
1014 		ASSERT(VTOZ(*vpp)->z_zfsvfs != zfsvfs);
1015 		VTOZ(*vpp)->z_zfsvfs->z_parent = zfsvfs;
1016 
1017 		/* Clear the root flag (set via VFS_ROOT) as well. */
1018 		(*vpp)->v_vflag &= ~VV_ROOT;
1019 	}
1020 
1021 	if (err != 0)
1022 		*vpp = NULL;
1023 	return (err);
1024 }
1025 
1026 static int
1027 zfsctl_snapdir_readdir(struct vop_readdir_args *ap)
1028 {
1029 	char snapname[ZFS_MAX_DATASET_NAME_LEN];
1030 	struct dirent entry;
1031 	vnode_t *vp = ap->a_vp;
1032 	zfsvfs_t *zfsvfs = vp->v_vfsp->vfs_data;
1033 	uio_t *uio = ap->a_uio;
1034 	int *eofp = ap->a_eofflag;
1035 	off_t dots_offset;
1036 	int error;
1037 
1038 	ASSERT(vp->v_type == VDIR);
1039 
1040 	error = sfs_readdir_common(ZFSCTL_INO_ROOT, ZFSCTL_INO_SNAPDIR, ap, uio,
1041 	    &dots_offset);
1042 	if (error != 0) {
1043 		if (error == ENAMETOOLONG) /* ran out of destination space */
1044 			error = 0;
1045 		return (error);
1046 	}
1047 
1048 	ZFS_ENTER(zfsvfs);
1049 	for (;;) {
1050 		uint64_t cookie;
1051 		uint64_t id;
1052 
1053 		cookie = uio->uio_offset - dots_offset;
1054 
1055 		dsl_pool_config_enter(dmu_objset_pool(zfsvfs->z_os), FTAG);
1056 		error = dmu_snapshot_list_next(zfsvfs->z_os, sizeof (snapname),
1057 		    snapname, &id, &cookie, NULL);
1058 		dsl_pool_config_exit(dmu_objset_pool(zfsvfs->z_os), FTAG);
1059 		if (error != 0) {
1060 			if (error == ENOENT) {
1061 				if (eofp != NULL)
1062 					*eofp = 1;
1063 				error = 0;
1064 			}
1065 			ZFS_EXIT(zfsvfs);
1066 			return (error);
1067 		}
1068 
1069 		entry.d_fileno = id;
1070 		entry.d_type = DT_DIR;
1071 		strcpy(entry.d_name, snapname);
1072 		entry.d_namlen = strlen(entry.d_name);
1073 		entry.d_reclen = sizeof (entry);
1074 		error = vfs_read_dirent(ap, &entry, uio->uio_offset);
1075 		if (error != 0) {
1076 			if (error == ENAMETOOLONG)
1077 				error = 0;
1078 			ZFS_EXIT(zfsvfs);
1079 			return (SET_ERROR(error));
1080 		}
1081 		uio->uio_offset = cookie + dots_offset;
1082 	}
1083 	/* NOTREACHED */
1084 }
1085 
1086 static int
1087 zfsctl_snapdir_getattr(struct vop_getattr_args *ap)
1088 {
1089 	vnode_t *vp = ap->a_vp;
1090 	vattr_t *vap = ap->a_vap;
1091 	zfsvfs_t *zfsvfs = vp->v_vfsp->vfs_data;
1092 	dsl_dataset_t *ds;
1093 	uint64_t snap_count;
1094 	int err;
1095 
1096 	ZFS_ENTER(zfsvfs);
1097 	ds = dmu_objset_ds(zfsvfs->z_os);
1098 	zfsctl_common_getattr(vp, vap);
1099 	vap->va_ctime = dmu_objset_snap_cmtime(zfsvfs->z_os);
1100 	vap->va_mtime = vap->va_ctime;
1101 	vap->va_birthtime = vap->va_ctime;
1102 	if (dsl_dataset_phys(ds)->ds_snapnames_zapobj != 0) {
1103 		err = zap_count(dmu_objset_pool(ds->ds_objset)->dp_meta_objset,
1104 		    dsl_dataset_phys(ds)->ds_snapnames_zapobj, &snap_count);
1105 		if (err != 0) {
1106 			ZFS_EXIT(zfsvfs);
1107 			return (err);
1108 		}
1109 		vap->va_nlink += snap_count;
1110 	}
1111 	vap->va_size = vap->va_nlink;
1112 
1113 	ZFS_EXIT(zfsvfs);
1114 	return (0);
1115 }
1116 
1117 static struct vop_vector zfsctl_ops_snapdir = {
1118 	.vop_default =	&default_vnodeops,
1119 #if __FreeBSD_version >= 1300121
1120 	.vop_fplookup_vexec = VOP_EAGAIN,
1121 #endif
1122 	.vop_open =	zfsctl_common_open,
1123 	.vop_close =	zfsctl_common_close,
1124 	.vop_getattr =	zfsctl_snapdir_getattr,
1125 	.vop_access =	zfsctl_common_access,
1126 	.vop_readdir =	zfsctl_snapdir_readdir,
1127 	.vop_lookup =	zfsctl_snapdir_lookup,
1128 	.vop_reclaim =	zfsctl_common_reclaim,
1129 	.vop_fid =	zfsctl_common_fid,
1130 	.vop_print =	zfsctl_common_print,
1131 	.vop_pathconf =	zfsctl_common_pathconf,
1132 	.vop_getacl =	zfsctl_common_getacl,
1133 };
1134 VFS_VOP_VECTOR_REGISTER(zfsctl_ops_snapdir);
1135 
1136 
1137 static int
1138 zfsctl_snapshot_inactive(struct vop_inactive_args *ap)
1139 {
1140 	vnode_t *vp = ap->a_vp;
1141 
1142 	VERIFY(vrecycle(vp) == 1);
1143 	return (0);
1144 }
1145 
1146 static int
1147 zfsctl_snapshot_reclaim(struct vop_reclaim_args *ap)
1148 {
1149 	vnode_t *vp = ap->a_vp;
1150 	void *data = vp->v_data;
1151 
1152 	sfs_reclaim_vnode(vp);
1153 	sfs_destroy_node(data);
1154 	return (0);
1155 }
1156 
1157 static int
1158 zfsctl_snapshot_vptocnp(struct vop_vptocnp_args *ap)
1159 {
1160 	struct mount *mp;
1161 	vnode_t *dvp;
1162 	vnode_t *vp;
1163 	sfs_node_t *node;
1164 	size_t len;
1165 	int locked;
1166 	int error;
1167 
1168 	vp = ap->a_vp;
1169 	node = vp->v_data;
1170 	len = strlen(node->sn_name);
1171 	if (*ap->a_buflen < len)
1172 		return (SET_ERROR(ENOMEM));
1173 
1174 	/*
1175 	 * Prevent unmounting of the snapshot while the vnode lock
1176 	 * is not held.  That is not strictly required, but allows
1177 	 * us to assert that an uncovered snapshot vnode is never
1178 	 * "leaked".
1179 	 */
1180 	mp = vp->v_mountedhere;
1181 	if (mp == NULL)
1182 		return (SET_ERROR(ENOENT));
1183 	error = vfs_busy(mp, 0);
1184 	KASSERT(error == 0, ("vfs_busy(mp, 0) failed with %d", error));
1185 
1186 	/*
1187 	 * We can vput the vnode as we can now depend on the reference owned
1188 	 * by the busied mp.  But we also need to hold the vnode, because
1189 	 * the reference may go after vfs_unbusy() which has to be called
1190 	 * before we can lock the vnode again.
1191 	 */
1192 	locked = VOP_ISLOCKED(vp);
1193 #if __FreeBSD_version >= 1300045
1194 	enum vgetstate vs = vget_prep(vp);
1195 #else
1196 	vhold(vp);
1197 #endif
1198 	vput(vp);
1199 
1200 	/* Look up .zfs/snapshot, our parent. */
1201 	error = zfsctl_snapdir_vnode(vp->v_mount, NULL, LK_SHARED, &dvp);
1202 	if (error == 0) {
1203 		VOP_UNLOCK1(dvp);
1204 		*ap->a_vpp = dvp;
1205 		*ap->a_buflen -= len;
1206 		bcopy(node->sn_name, ap->a_buf + *ap->a_buflen, len);
1207 	}
1208 	vfs_unbusy(mp);
1209 #if __FreeBSD_version >= 1300045
1210 	vget_finish(vp, locked | LK_RETRY, vs);
1211 #else
1212 	vget(vp, locked | LK_VNHELD | LK_RETRY, curthread);
1213 #endif
1214 	return (error);
1215 }
1216 
1217 /*
1218  * These VP's should never see the light of day.  They should always
1219  * be covered.
1220  */
1221 static struct vop_vector zfsctl_ops_snapshot = {
1222 #if __FreeBSD_version >= 1300121
1223 	.vop_fplookup_vexec =	VOP_EAGAIN,
1224 #endif
1225 	.vop_inactive =		zfsctl_snapshot_inactive,
1226 #if __FreeBSD_version >= 1300045
1227 	.vop_need_inactive = vop_stdneed_inactive,
1228 #endif
1229 	.vop_reclaim =		zfsctl_snapshot_reclaim,
1230 	.vop_vptocnp =		zfsctl_snapshot_vptocnp,
1231 	.vop_lock1 =		vop_stdlock,
1232 	.vop_unlock =		vop_stdunlock,
1233 	.vop_islocked =		vop_stdislocked,
1234 	.vop_advlockpurge =	vop_stdadvlockpurge, /* called by vgone */
1235 	.vop_print =		zfsctl_common_print,
1236 };
1237 VFS_VOP_VECTOR_REGISTER(zfsctl_ops_snapshot);
1238 
1239 int
1240 zfsctl_lookup_objset(vfs_t *vfsp, uint64_t objsetid, zfsvfs_t **zfsvfsp)
1241 {
1242 	zfsvfs_t *zfsvfs __unused = vfsp->vfs_data;
1243 	vnode_t *vp;
1244 	int error;
1245 
1246 	ASSERT(zfsvfs->z_ctldir != NULL);
1247 	*zfsvfsp = NULL;
1248 	error = sfs_vnode_get(vfsp, LK_EXCLUSIVE,
1249 	    ZFSCTL_INO_SNAPDIR, objsetid, &vp);
1250 	if (error == 0 && vp != NULL) {
1251 		/*
1252 		 * XXX Probably need to at least reference, if not busy, the mp.
1253 		 */
1254 		if (vp->v_mountedhere != NULL)
1255 			*zfsvfsp = vp->v_mountedhere->mnt_data;
1256 		vput(vp);
1257 	}
1258 	if (*zfsvfsp == NULL)
1259 		return (SET_ERROR(EINVAL));
1260 	return (0);
1261 }
1262 
1263 /*
1264  * Unmount any snapshots for the given filesystem.  This is called from
1265  * zfs_umount() - if we have a ctldir, then go through and unmount all the
1266  * snapshots.
1267  */
1268 int
1269 zfsctl_umount_snapshots(vfs_t *vfsp, int fflags, cred_t *cr)
1270 {
1271 	char snapname[ZFS_MAX_DATASET_NAME_LEN];
1272 	zfsvfs_t *zfsvfs = vfsp->vfs_data;
1273 	struct mount *mp;
1274 	vnode_t *vp;
1275 	uint64_t cookie;
1276 	int error;
1277 
1278 	ASSERT(zfsvfs->z_ctldir != NULL);
1279 
1280 	cookie = 0;
1281 	for (;;) {
1282 		uint64_t id;
1283 
1284 		dsl_pool_config_enter(dmu_objset_pool(zfsvfs->z_os), FTAG);
1285 		error = dmu_snapshot_list_next(zfsvfs->z_os, sizeof (snapname),
1286 		    snapname, &id, &cookie, NULL);
1287 		dsl_pool_config_exit(dmu_objset_pool(zfsvfs->z_os), FTAG);
1288 		if (error != 0) {
1289 			if (error == ENOENT)
1290 				error = 0;
1291 			break;
1292 		}
1293 
1294 		for (;;) {
1295 			error = sfs_vnode_get(vfsp, LK_EXCLUSIVE,
1296 			    ZFSCTL_INO_SNAPDIR, id, &vp);
1297 			if (error != 0 || vp == NULL)
1298 				break;
1299 
1300 			mp = vp->v_mountedhere;
1301 
1302 			/*
1303 			 * v_mountedhere being NULL means that the
1304 			 * (uncovered) vnode is in a transient state
1305 			 * (mounting or unmounting), so loop until it
1306 			 * settles down.
1307 			 */
1308 			if (mp != NULL)
1309 				break;
1310 			vput(vp);
1311 		}
1312 		if (error != 0)
1313 			break;
1314 		if (vp == NULL)
1315 			continue;	/* no mountpoint, nothing to do */
1316 
1317 		/*
1318 		 * The mount-point vnode is kept locked to avoid spurious EBUSY
1319 		 * from a concurrent umount.
1320 		 * The vnode lock must have recursive locking enabled.
1321 		 */
1322 		vfs_ref(mp);
1323 		error = dounmount(mp, fflags, curthread);
1324 		KASSERT_IMPLY(error == 0, vrefcnt(vp) == 1,
1325 		    ("extra references after unmount"));
1326 		vput(vp);
1327 		if (error != 0)
1328 			break;
1329 	}
1330 	KASSERT_IMPLY((fflags & MS_FORCE) != 0, error == 0,
1331 	    ("force unmounting failed"));
1332 	return (error);
1333 }
1334 
1335 int
1336 zfsctl_snapshot_unmount(const char *snapname, int flags __unused)
1337 {
1338 	vfs_t *vfsp = NULL;
1339 	zfsvfs_t *zfsvfs = NULL;
1340 
1341 	if (strchr(snapname, '@') == NULL)
1342 		return (0);
1343 
1344 	int err = getzfsvfs(snapname, &zfsvfs);
1345 	if (err != 0) {
1346 		ASSERT3P(zfsvfs, ==, NULL);
1347 		return (0);
1348 	}
1349 	vfsp = zfsvfs->z_vfs;
1350 
1351 	ASSERT(!dsl_pool_config_held(dmu_objset_pool(zfsvfs->z_os)));
1352 
1353 	vfs_ref(vfsp);
1354 	vfs_unbusy(vfsp);
1355 	return (dounmount(vfsp, MS_FORCE, curthread));
1356 }
1357