xref: /freebsd/sys/ufs/ffs/ffs_snapshot.c (revision f856af0466c076beef4ea9b15d088e1119a945b8)
1 /*-
2  * Copyright 2000 Marshall Kirk McKusick. All Rights Reserved.
3  *
4  * Further information about snapshots can be obtained from:
5  *
6  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
7  *	1614 Oxford Street		mckusick@mckusick.com
8  *	Berkeley, CA 94709-1608		+1-510-843-9542
9  *	USA
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
22  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
23  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
24  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
25  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	@(#)ffs_snapshot.c	8.11 (McKusick) 7/23/00
34  */
35 
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 #include "opt_quota.h"
40 
41 #include <sys/param.h>
42 #include <sys/kernel.h>
43 #include <sys/systm.h>
44 #include <sys/conf.h>
45 #include <sys/bio.h>
46 #include <sys/buf.h>
47 #include <sys/proc.h>
48 #include <sys/namei.h>
49 #include <sys/sched.h>
50 #include <sys/stat.h>
51 #include <sys/malloc.h>
52 #include <sys/mount.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/vnode.h>
56 
57 #include <geom/geom.h>
58 
59 #include <ufs/ufs/extattr.h>
60 #include <ufs/ufs/quota.h>
61 #include <ufs/ufs/ufsmount.h>
62 #include <ufs/ufs/inode.h>
63 #include <ufs/ufs/ufs_extern.h>
64 
65 #include <ufs/ffs/fs.h>
66 #include <ufs/ffs/ffs_extern.h>
67 
68 #define KERNCRED thread0.td_ucred
69 #define DEBUG 1
70 
71 #include "opt_ffs.h"
72 
73 #ifdef NO_FFS_SNAPSHOT
74 int
75 ffs_snapshot(mp, snapfile)
76 	struct mount *mp;
77 	char *snapfile;
78 {
79 	return (EINVAL);
80 }
81 
82 int
83 ffs_snapblkfree(fs, devvp, bno, size, inum)
84 	struct fs *fs;
85 	struct vnode *devvp;
86 	ufs2_daddr_t bno;
87 	long size;
88 	ino_t inum;
89 {
90 	return (EINVAL);
91 }
92 
93 void
94 ffs_snapremove(vp)
95 	struct vnode *vp;
96 {
97 }
98 
99 void
100 ffs_snapshot_mount(mp)
101 	struct mount *mp;
102 {
103 }
104 
105 void
106 ffs_snapshot_unmount(mp)
107 	struct mount *mp;
108 {
109 }
110 
111 void
112 ffs_snapgone(ip)
113 	struct inode *ip;
114 {
115 }
116 
117 int
118 ffs_copyonwrite(devvp, bp)
119 	struct vnode *devvp;
120 	struct buf *bp;
121 {
122 	return (EINVAL);
123 }
124 
125 #else
126 
127 TAILQ_HEAD(snaphead, inode);
128 
129 struct snapdata {
130 	struct snaphead sn_head;
131 	daddr_t sn_listsize;
132 	daddr_t *sn_blklist;
133 	struct lock sn_lock;
134 };
135 
136 static int cgaccount(int, struct vnode *, struct buf *, int);
137 static int expunge_ufs1(struct vnode *, struct inode *, struct fs *,
138     int (*)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, struct fs *,
139     ufs_lbn_t, int), int);
140 static int indiracct_ufs1(struct vnode *, struct vnode *, int,
141     ufs1_daddr_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, struct fs *,
142     int (*)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, struct fs *,
143     ufs_lbn_t, int), int);
144 static int fullacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
145     struct fs *, ufs_lbn_t, int);
146 static int snapacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
147     struct fs *, ufs_lbn_t, int);
148 static int mapacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
149     struct fs *, ufs_lbn_t, int);
150 static int expunge_ufs2(struct vnode *, struct inode *, struct fs *,
151     int (*)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, struct fs *,
152     ufs_lbn_t, int), int);
153 static int indiracct_ufs2(struct vnode *, struct vnode *, int,
154     ufs2_daddr_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, struct fs *,
155     int (*)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, struct fs *,
156     ufs_lbn_t, int), int);
157 static int fullacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
158     struct fs *, ufs_lbn_t, int);
159 static int snapacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
160     struct fs *, ufs_lbn_t, int);
161 static int mapacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
162     struct fs *, ufs_lbn_t, int);
163 static int readblock(struct vnode *vp, struct buf *, ufs2_daddr_t);
164 static void process_deferred_inactive(struct mount *);
165 static void try_free_snapdata(struct vnode *devvp, struct thread *td);
166 
167 /*
168  * To ensure the consistency of snapshots across crashes, we must
169  * synchronously write out copied blocks before allowing the
170  * originals to be modified. Because of the rather severe speed
171  * penalty that this imposes, the following flag allows this
172  * crash persistence to be disabled.
173  */
174 int dopersistence = 0;
175 
176 #ifdef DEBUG
177 #include <sys/sysctl.h>
178 SYSCTL_INT(_debug, OID_AUTO, dopersistence, CTLFLAG_RW, &dopersistence, 0, "");
179 static int snapdebug = 0;
180 SYSCTL_INT(_debug, OID_AUTO, snapdebug, CTLFLAG_RW, &snapdebug, 0, "");
181 int collectsnapstats = 0;
182 SYSCTL_INT(_debug, OID_AUTO, collectsnapstats, CTLFLAG_RW, &collectsnapstats,
183 	0, "");
184 #endif /* DEBUG */
185 
186 /*
187  * Create a snapshot file and initialize it for the filesystem.
188  */
189 int
190 ffs_snapshot(mp, snapfile)
191 	struct mount *mp;
192 	char *snapfile;
193 {
194 	ufs2_daddr_t numblks, blkno, *blkp, *snapblklist;
195 	int error, cg, snaploc;
196 	int i, size, len, loc;
197 	int flag;
198 	struct timespec starttime = {0, 0}, endtime;
199 	char saved_nice = 0;
200 	long redo = 0, snaplistsize = 0;
201 	int32_t *lp;
202 	void *space;
203 	struct fs *copy_fs = NULL, *fs;
204 	struct thread *td = curthread;
205 	struct inode *ip, *xp;
206 	struct buf *bp, *nbp, *ibp, *sbp = NULL;
207 	struct nameidata nd;
208 	struct mount *wrtmp;
209 	struct vattr vat;
210 	struct vnode *vp, *xvp, *mvp, *devvp;
211 	struct uio auio;
212 	struct iovec aiov;
213 	struct snapdata *sn;
214 	struct ufsmount *ump;
215 
216 	ump = VFSTOUFS(mp);
217 	fs = ump->um_fs;
218 	sn = NULL;
219 	MNT_ILOCK(mp);
220 	flag = mp->mnt_flag;
221 	MNT_IUNLOCK(mp);
222 
223 	/*
224 	 * Need to serialize access to snapshot code per filesystem.
225 	 */
226 	/*
227 	 * Assign a snapshot slot in the superblock.
228 	 */
229 	UFS_LOCK(ump);
230 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++)
231 		if (fs->fs_snapinum[snaploc] == 0)
232 			break;
233 	UFS_UNLOCK(ump);
234 	if (snaploc == FSMAXSNAP)
235 		return (ENOSPC);
236 	/*
237 	 * Create the snapshot file.
238 	 */
239 restart:
240 	NDINIT(&nd, CREATE, LOCKPARENT | LOCKLEAF, UIO_SYSSPACE, snapfile, td);
241 	if ((error = namei(&nd)) != 0)
242 		return (error);
243 	if (nd.ni_vp != NULL) {
244 		vput(nd.ni_vp);
245 		error = EEXIST;
246 	}
247 	if (nd.ni_dvp->v_mount != mp)
248 		error = EXDEV;
249 	if (error) {
250 		NDFREE(&nd, NDF_ONLY_PNBUF);
251 		if (nd.ni_dvp == nd.ni_vp)
252 			vrele(nd.ni_dvp);
253 		else
254 			vput(nd.ni_dvp);
255 		return (error);
256 	}
257 	VATTR_NULL(&vat);
258 	vat.va_type = VREG;
259 	vat.va_mode = S_IRUSR;
260 	vat.va_vaflags |= VA_EXCLUSIVE;
261 	if (VOP_GETWRITEMOUNT(nd.ni_dvp, &wrtmp))
262 		wrtmp = NULL;
263 	if (wrtmp != mp)
264 		panic("ffs_snapshot: mount mismatch");
265 	vfs_rel(wrtmp);
266 	if (vn_start_write(NULL, &wrtmp, V_NOWAIT) != 0) {
267 		NDFREE(&nd, NDF_ONLY_PNBUF);
268 		vput(nd.ni_dvp);
269 		if ((error = vn_start_write(NULL, &wrtmp,
270 		    V_XSLEEP | PCATCH)) != 0)
271 			return (error);
272 		goto restart;
273 	}
274 	VOP_LEASE(nd.ni_dvp, td, KERNCRED, LEASE_WRITE);
275 	error = VOP_CREATE(nd.ni_dvp, &nd.ni_vp, &nd.ni_cnd, &vat);
276 	VOP_UNLOCK(nd.ni_dvp, 0, td);
277 	if (error) {
278 		NDFREE(&nd, NDF_ONLY_PNBUF);
279 		vn_finished_write(wrtmp);
280 		vrele(nd.ni_dvp);
281 		return (error);
282 	}
283 	vp = nd.ni_vp;
284 	ip = VTOI(vp);
285 	devvp = ip->i_devvp;
286 	/*
287 	 * Allocate and copy the last block contents so as to be able
288 	 * to set size to that of the filesystem.
289 	 */
290 	numblks = howmany(fs->fs_size, fs->fs_frag);
291 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)),
292 	    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
293 	if (error)
294 		goto out;
295 	ip->i_size = lblktosize(fs, (off_t)numblks);
296 	DIP_SET(ip, i_size, ip->i_size);
297 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
298 	error = readblock(vp, bp, numblks - 1);
299 	bawrite(bp);
300 	if (error != 0)
301 		goto out;
302 	/*
303 	 * Preallocate critical data structures so that we can copy
304 	 * them in without further allocation after we suspend all
305 	 * operations on the filesystem. We would like to just release
306 	 * the allocated buffers without writing them since they will
307 	 * be filled in below once we are ready to go, but this upsets
308 	 * the soft update code, so we go ahead and write the new buffers.
309 	 *
310 	 * Allocate all indirect blocks and mark all of them as not
311 	 * needing to be copied.
312 	 */
313 	for (blkno = NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
314 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
315 		    fs->fs_bsize, td->td_ucred, BA_METAONLY, &ibp);
316 		if (error)
317 			goto out;
318 		bawrite(ibp);
319 	}
320 	/*
321 	 * Allocate copies for the superblock and its summary information.
322 	 */
323 	error = UFS_BALLOC(vp, fs->fs_sblockloc, fs->fs_sbsize, KERNCRED,
324 	    0, &nbp);
325 	if (error)
326 		goto out;
327 	bawrite(nbp);
328 	blkno = fragstoblks(fs, fs->fs_csaddr);
329 	len = howmany(fs->fs_cssize, fs->fs_bsize);
330 	for (loc = 0; loc < len; loc++) {
331 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(blkno + loc)),
332 		    fs->fs_bsize, KERNCRED, 0, &nbp);
333 		if (error)
334 			goto out;
335 		bawrite(nbp);
336 	}
337 	/*
338 	 * Allocate all cylinder group blocks.
339 	 */
340 	for (cg = 0; cg < fs->fs_ncg; cg++) {
341 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
342 		    fs->fs_bsize, KERNCRED, 0, &nbp);
343 		if (error)
344 			goto out;
345 		bawrite(nbp);
346 	}
347 	/*
348 	 * Copy all the cylinder group maps. Although the
349 	 * filesystem is still active, we hope that only a few
350 	 * cylinder groups will change between now and when we
351 	 * suspend operations. Thus, we will be able to quickly
352 	 * touch up the few cylinder groups that changed during
353 	 * the suspension period.
354 	 */
355 	len = howmany(fs->fs_ncg, NBBY);
356 	MALLOC(space, void *, len, M_DEVBUF, M_WAITOK|M_ZERO);
357 	UFS_LOCK(ump);
358 	fs->fs_active = space;
359 	UFS_UNLOCK(ump);
360 	for (cg = 0; cg < fs->fs_ncg; cg++) {
361 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
362 		    fs->fs_bsize, KERNCRED, 0, &nbp);
363 		if (error)
364 			goto out;
365 		error = cgaccount(cg, vp, nbp, 1);
366 		bawrite(nbp);
367 		if (error)
368 			goto out;
369 	}
370 #ifdef QUOTA
371 	/*
372 	 * Turn off disk quotas for snapshot file.
373 	 */
374 	(void) chkdq(ip, -DIP(ip, i_blocks), KERNCRED, FORCE);
375 	for (i = 0; i < MAXQUOTAS; i++) {
376 		if (ip->i_dquot[i] != NODQUOT) {
377 			dqrele(vp, ip->i_dquot[i]);
378 			ip->i_dquot[i] = NODQUOT;
379 		}
380 	}
381 #endif
382 	/*
383 	 * Change inode to snapshot type file.
384 	 */
385 	ip->i_flags |= SF_SNAPSHOT;
386 	DIP_SET(ip, i_flags, ip->i_flags);
387 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
388 	/*
389 	 * Ensure that the snapshot is completely on disk.
390 	 * Since we have marked it as a snapshot it is safe to
391 	 * unlock it as no process will be allowed to write to it.
392 	 */
393 	if ((error = ffs_syncvnode(vp, MNT_WAIT)) != 0)
394 		goto out;
395 	VOP_UNLOCK(vp, 0, td);
396 	/*
397 	 * All allocations are done, so we can now snapshot the system.
398 	 *
399 	 * Recind nice scheduling while running with the filesystem suspended.
400 	 */
401 	if (td->td_proc->p_nice > 0) {
402 		PROC_LOCK(td->td_proc);
403 		mtx_lock_spin(&sched_lock);
404 		saved_nice = td->td_proc->p_nice;
405 		sched_nice(td->td_proc, 0);
406 		mtx_unlock_spin(&sched_lock);
407 		PROC_UNLOCK(td->td_proc);
408 	}
409 	/*
410 	 * Suspend operation on filesystem.
411 	 */
412 	for (;;) {
413 		vn_finished_write(wrtmp);
414 		if ((error = vfs_write_suspend(vp->v_mount)) != 0) {
415 			vn_start_write(NULL, &wrtmp, V_WAIT);
416 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
417 			goto out;
418 		}
419 		if (mp->mnt_kern_flag & MNTK_SUSPENDED)
420 			break;
421 		vn_start_write(NULL, &wrtmp, V_WAIT);
422 	}
423 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
424 	if (ip->i_effnlink == 0) {
425 		error = ENOENT;		/* Snapshot file unlinked */
426 		goto out1;
427 	}
428 	if (collectsnapstats)
429 		nanotime(&starttime);
430 
431 	/* The last block might have changed.  Copy it again to be sure. */
432 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)),
433 	    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
434 	if (error != 0)
435 		goto out1;
436 	error = readblock(vp, bp, numblks - 1);
437 	bp->b_flags |= B_VALIDSUSPWRT;
438 	bawrite(bp);
439 	if (error != 0)
440 		goto out1;
441 	/*
442 	 * First, copy all the cylinder group maps that have changed.
443 	 */
444 	for (cg = 0; cg < fs->fs_ncg; cg++) {
445 		if ((ACTIVECGNUM(fs, cg) & ACTIVECGOFF(cg)) != 0)
446 			continue;
447 		redo++;
448 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
449 		    fs->fs_bsize, KERNCRED, 0, &nbp);
450 		if (error)
451 			goto out1;
452 		error = cgaccount(cg, vp, nbp, 2);
453 		bawrite(nbp);
454 		if (error)
455 			goto out1;
456 	}
457 	/*
458 	 * Grab a copy of the superblock and its summary information.
459 	 * We delay writing it until the suspension is released below.
460 	 */
461 	error = bread(vp, lblkno(fs, fs->fs_sblockloc), fs->fs_bsize,
462 	    KERNCRED, &sbp);
463 	if (error) {
464 		brelse(sbp);
465 		sbp = NULL;
466 		goto out1;
467 	}
468 	loc = blkoff(fs, fs->fs_sblockloc);
469 	copy_fs = (struct fs *)(sbp->b_data + loc);
470 	bcopy(fs, copy_fs, fs->fs_sbsize);
471 	if ((fs->fs_flags & (FS_UNCLEAN | FS_NEEDSFSCK)) == 0)
472 		copy_fs->fs_clean = 1;
473 	size = fs->fs_bsize < SBLOCKSIZE ? fs->fs_bsize : SBLOCKSIZE;
474 	if (fs->fs_sbsize < size)
475 		bzero(&sbp->b_data[loc + fs->fs_sbsize], size - fs->fs_sbsize);
476 	size = blkroundup(fs, fs->fs_cssize);
477 	if (fs->fs_contigsumsize > 0)
478 		size += fs->fs_ncg * sizeof(int32_t);
479 	space = malloc((u_long)size, M_UFSMNT, M_WAITOK);
480 	copy_fs->fs_csp = space;
481 	bcopy(fs->fs_csp, copy_fs->fs_csp, fs->fs_cssize);
482 	space = (char *)space + fs->fs_cssize;
483 	loc = howmany(fs->fs_cssize, fs->fs_fsize);
484 	i = fs->fs_frag - loc % fs->fs_frag;
485 	len = (i == fs->fs_frag) ? 0 : i * fs->fs_fsize;
486 	if (len > 0) {
487 		if ((error = bread(devvp, fsbtodb(fs, fs->fs_csaddr + loc),
488 		    len, KERNCRED, &bp)) != 0) {
489 			brelse(bp);
490 			free(copy_fs->fs_csp, M_UFSMNT);
491 			bawrite(sbp);
492 			sbp = NULL;
493 			goto out1;
494 		}
495 		bcopy(bp->b_data, space, (u_int)len);
496 		space = (char *)space + len;
497 		bp->b_flags |= B_INVAL | B_NOCACHE;
498 		brelse(bp);
499 	}
500 	if (fs->fs_contigsumsize > 0) {
501 		copy_fs->fs_maxcluster = lp = space;
502 		for (i = 0; i < fs->fs_ncg; i++)
503 			*lp++ = fs->fs_contigsumsize;
504 	}
505 	/*
506 	 * We must check for active files that have been unlinked
507 	 * (e.g., with a zero link count). We have to expunge all
508 	 * trace of these files from the snapshot so that they are
509 	 * not reclaimed prematurely by fsck or unnecessarily dumped.
510 	 * We turn off the MNTK_SUSPENDED flag to avoid a panic from
511 	 * spec_strategy about writing on a suspended filesystem.
512 	 * Note that we skip unlinked snapshot files as they will
513 	 * be handled separately below.
514 	 *
515 	 * We also calculate the needed size for the snapshot list.
516 	 */
517 	snaplistsize = fs->fs_ncg + howmany(fs->fs_cssize, fs->fs_bsize) +
518 	    FSMAXSNAP + 1 /* superblock */ + 1 /* last block */ + 1 /* size */;
519 	MNT_ILOCK(mp);
520 	mp->mnt_kern_flag &= ~MNTK_SUSPENDED;
521 loop:
522 	MNT_VNODE_FOREACH(xvp, mp, mvp) {
523 		VI_LOCK(xvp);
524 		MNT_IUNLOCK(mp);
525 		if ((xvp->v_iflag & VI_DOOMED) ||
526 		    (xvp->v_usecount == 0 &&
527 		     (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) ||
528 		    xvp->v_type == VNON ||
529 		    (VTOI(xvp)->i_flags & SF_SNAPSHOT)) {
530 			VI_UNLOCK(xvp);
531 			MNT_ILOCK(mp);
532 			continue;
533 		}
534 		/*
535 		 * We can skip parent directory vnode because it must have
536 		 * this snapshot file in it.
537 		 */
538 		if (xvp == nd.ni_dvp) {
539 			VI_UNLOCK(xvp);
540 			MNT_ILOCK(mp);
541 			continue;
542 		}
543 		vholdl(xvp);
544 		if (vn_lock(xvp, LK_EXCLUSIVE | LK_INTERLOCK, td) != 0) {
545 			MNT_ILOCK(mp);
546 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
547 			vdrop(xvp);
548 			goto loop;
549 		}
550 		VI_LOCK(xvp);
551 		if (xvp->v_usecount == 0 &&
552 		    (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) {
553 			VI_UNLOCK(xvp);
554 			VOP_UNLOCK(xvp, 0, td);
555 			vdrop(xvp);
556 			MNT_ILOCK(mp);
557 			continue;
558 		}
559 		VI_UNLOCK(xvp);
560 		if (snapdebug)
561 			vprint("ffs_snapshot: busy vnode", xvp);
562 		if (VOP_GETATTR(xvp, &vat, td->td_ucred, td) == 0 &&
563 		    vat.va_nlink > 0) {
564 			VOP_UNLOCK(xvp, 0, td);
565 			vdrop(xvp);
566 			MNT_ILOCK(mp);
567 			continue;
568 		}
569 		xp = VTOI(xvp);
570 		if (ffs_checkfreefile(copy_fs, vp, xp->i_number)) {
571 			VOP_UNLOCK(xvp, 0, td);
572 			vdrop(xvp);
573 			MNT_ILOCK(mp);
574 			continue;
575 		}
576 		/*
577 		 * If there is a fragment, clear it here.
578 		 */
579 		blkno = 0;
580 		loc = howmany(xp->i_size, fs->fs_bsize) - 1;
581 		if (loc < NDADDR) {
582 			len = fragroundup(fs, blkoff(fs, xp->i_size));
583 			if (len != 0 && len < fs->fs_bsize) {
584 				ffs_blkfree(ump, copy_fs, vp,
585 				    DIP(xp, i_db[loc]), len, xp->i_number);
586 				blkno = DIP(xp, i_db[loc]);
587 				DIP_SET(xp, i_db[loc], 0);
588 			}
589 		}
590 		snaplistsize += 1;
591 		if (xp->i_ump->um_fstype == UFS1)
592 			error = expunge_ufs1(vp, xp, copy_fs, fullacct_ufs1,
593 			    BLK_NOCOPY);
594 		else
595 			error = expunge_ufs2(vp, xp, copy_fs, fullacct_ufs2,
596 			    BLK_NOCOPY);
597 		if (blkno)
598 			DIP_SET(xp, i_db[loc], blkno);
599 		if (!error)
600 			error = ffs_freefile(ump, copy_fs, vp, xp->i_number,
601 			    xp->i_mode);
602 		VOP_UNLOCK(xvp, 0, td);
603 		vdrop(xvp);
604 		if (error) {
605 			free(copy_fs->fs_csp, M_UFSMNT);
606 			bawrite(sbp);
607 			sbp = NULL;
608 			MNT_VNODE_FOREACH_ABORT(mp, mvp);
609 			goto out1;
610 		}
611 		MNT_ILOCK(mp);
612 	}
613 	MNT_IUNLOCK(mp);
614 	/*
615 	 * If there already exist snapshots on this filesystem, grab a
616 	 * reference to their shared lock. If this is the first snapshot
617 	 * on this filesystem, we need to allocate a lock for the snapshots
618 	 * to share. In either case, acquire the snapshot lock and give
619 	 * up our original private lock.
620 	 */
621 	VI_LOCK(devvp);
622 	sn = devvp->v_rdev->si_snapdata;
623 	if (sn != NULL) {
624 		xp = TAILQ_FIRST(&sn->sn_head);
625 		VI_UNLOCK(devvp);
626 		VI_LOCK(vp);
627 		vp->v_vnlock = &sn->sn_lock;
628 	} else {
629 		VI_UNLOCK(devvp);
630 		sn = malloc(sizeof *sn, M_UFSMNT, M_WAITOK | M_ZERO);
631 		TAILQ_INIT(&sn->sn_head);
632 		lockinit(&sn->sn_lock, PVFS, "snaplk", VLKTIMEOUT,
633 		    LK_CANRECURSE | LK_NOSHARE);
634 		VI_LOCK(vp);
635 		vp->v_vnlock = &sn->sn_lock;
636 		mp_fixme("si_snapdata setting is racey.");
637 		devvp->v_rdev->si_snapdata = sn;
638 		xp = NULL;
639 	}
640 	lockmgr(vp->v_vnlock, LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY,
641 	    VI_MTX(vp), td);
642 	lockmgr(&vp->v_lock, LK_RELEASE, NULL, td);
643 	/*
644 	 * If this is the first snapshot on this filesystem, then we need
645 	 * to allocate the space for the list of preallocated snapshot blocks.
646 	 * This list will be refined below, but this preliminary one will
647 	 * keep us out of deadlock until the full one is ready.
648 	 */
649 	if (xp == NULL) {
650 		MALLOC(snapblklist, daddr_t *, snaplistsize * sizeof(daddr_t),
651 		    M_UFSMNT, M_WAITOK);
652 		blkp = &snapblklist[1];
653 		*blkp++ = lblkno(fs, fs->fs_sblockloc);
654 		blkno = fragstoblks(fs, fs->fs_csaddr);
655 		for (cg = 0; cg < fs->fs_ncg; cg++) {
656 			if (fragstoblks(fs, cgtod(fs, cg) > blkno))
657 				break;
658 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
659 		}
660 		len = howmany(fs->fs_cssize, fs->fs_bsize);
661 		for (loc = 0; loc < len; loc++)
662 			*blkp++ = blkno + loc;
663 		for (; cg < fs->fs_ncg; cg++)
664 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
665 		snapblklist[0] = blkp - snapblklist;
666 		VI_LOCK(devvp);
667 		if (sn->sn_blklist != NULL)
668 			panic("ffs_snapshot: non-empty list");
669 		sn->sn_blklist = snapblklist;
670 		sn->sn_listsize = blkp - snapblklist;
671 		VI_UNLOCK(devvp);
672 	}
673 	/*
674 	 * Record snapshot inode. Since this is the newest snapshot,
675 	 * it must be placed at the end of the list.
676 	 */
677 	VI_LOCK(devvp);
678 	fs->fs_snapinum[snaploc] = ip->i_number;
679 	if (ip->i_nextsnap.tqe_prev != 0)
680 		panic("ffs_snapshot: %d already on list", ip->i_number);
681 	TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
682 	devvp->v_vflag |= VV_COPYONWRITE;
683 	VI_UNLOCK(devvp);
684 	ASSERT_VOP_LOCKED(vp, "ffs_snapshot vp");
685 	vp->v_vflag |= VV_SYSTEM;
686 out1:
687 	KASSERT((sn != NULL && sbp != NULL && error == 0) ||
688 		(sn == NULL && sbp == NULL && error != 0),
689 		("email phk@ and mckusick@"));
690 	/*
691 	 * Resume operation on filesystem.
692 	 */
693 	vfs_write_resume(vp->v_mount);
694 	vn_start_write(NULL, &wrtmp, V_WAIT);
695 	if (collectsnapstats && starttime.tv_sec > 0) {
696 		nanotime(&endtime);
697 		timespecsub(&endtime, &starttime);
698 		printf("%s: suspended %ld.%03ld sec, redo %ld of %d\n",
699 		    vp->v_mount->mnt_stat.f_mntonname, (long)endtime.tv_sec,
700 		    endtime.tv_nsec / 1000000, redo, fs->fs_ncg);
701 	}
702 	if (sbp == NULL)
703 		goto out;
704 	/*
705 	 * Copy allocation information from all the snapshots in
706 	 * this snapshot and then expunge them from its view.
707 	 */
708 	TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap) {
709 		if (xp == ip)
710 			break;
711 		if (xp->i_ump->um_fstype == UFS1)
712 			error = expunge_ufs1(vp, xp, fs, snapacct_ufs1,
713 			    BLK_SNAP);
714 		else
715 			error = expunge_ufs2(vp, xp, fs, snapacct_ufs2,
716 			    BLK_SNAP);
717 		if (error == 0 && xp->i_effnlink == 0) {
718 			error = ffs_freefile(ump,
719 					     copy_fs,
720 					     vp,
721 					     xp->i_number,
722 					     xp->i_mode);
723 		}
724 		if (error) {
725 			fs->fs_snapinum[snaploc] = 0;
726 			goto done;
727 		}
728 	}
729 	/*
730 	 * Allocate space for the full list of preallocated snapshot blocks.
731 	 */
732 	MALLOC(snapblklist, daddr_t *, snaplistsize * sizeof(daddr_t),
733 	    M_UFSMNT, M_WAITOK);
734 	ip->i_snapblklist = &snapblklist[1];
735 	/*
736 	 * Expunge the blocks used by the snapshots from the set of
737 	 * blocks marked as used in the snapshot bitmaps. Also, collect
738 	 * the list of allocated blocks in i_snapblklist.
739 	 */
740 	if (ip->i_ump->um_fstype == UFS1)
741 		error = expunge_ufs1(vp, ip, copy_fs, mapacct_ufs1, BLK_SNAP);
742 	else
743 		error = expunge_ufs2(vp, ip, copy_fs, mapacct_ufs2, BLK_SNAP);
744 	if (error) {
745 		fs->fs_snapinum[snaploc] = 0;
746 		FREE(snapblklist, M_UFSMNT);
747 		goto done;
748 	}
749 	if (snaplistsize < ip->i_snapblklist - snapblklist)
750 		panic("ffs_snapshot: list too small");
751 	snaplistsize = ip->i_snapblklist - snapblklist;
752 	snapblklist[0] = snaplistsize;
753 	ip->i_snapblklist = 0;
754 	/*
755 	 * Write out the list of allocated blocks to the end of the snapshot.
756 	 */
757 	auio.uio_iov = &aiov;
758 	auio.uio_iovcnt = 1;
759 	aiov.iov_base = (void *)snapblklist;
760 	aiov.iov_len = snaplistsize * sizeof(daddr_t);
761 	auio.uio_resid = aiov.iov_len;;
762 	auio.uio_offset = ip->i_size;
763 	auio.uio_segflg = UIO_SYSSPACE;
764 	auio.uio_rw = UIO_WRITE;
765 	auio.uio_td = td;
766 	if ((error = VOP_WRITE(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
767 		fs->fs_snapinum[snaploc] = 0;
768 		FREE(snapblklist, M_UFSMNT);
769 		goto done;
770 	}
771 	/*
772 	 * Write the superblock and its summary information
773 	 * to the snapshot.
774 	 */
775 	blkno = fragstoblks(fs, fs->fs_csaddr);
776 	len = howmany(fs->fs_cssize, fs->fs_bsize);
777 	space = copy_fs->fs_csp;
778 	for (loc = 0; loc < len; loc++) {
779 		error = bread(vp, blkno + loc, fs->fs_bsize, KERNCRED, &nbp);
780 		if (error) {
781 			brelse(nbp);
782 			fs->fs_snapinum[snaploc] = 0;
783 			FREE(snapblklist, M_UFSMNT);
784 			goto done;
785 		}
786 		bcopy(space, nbp->b_data, fs->fs_bsize);
787 		space = (char *)space + fs->fs_bsize;
788 		bawrite(nbp);
789 	}
790 	/*
791 	 * As this is the newest list, it is the most inclusive, so
792 	 * should replace the previous list.
793 	 */
794 	VI_LOCK(devvp);
795 	space = sn->sn_blklist;
796 	sn->sn_blklist = snapblklist;
797 	sn->sn_listsize = snaplistsize;
798 	VI_UNLOCK(devvp);
799 	if (space != NULL)
800 		FREE(space, M_UFSMNT);
801 	/*
802 	 * If another process is currently writing the buffer containing
803 	 * the inode for this snapshot then a deadlock can occur. Drop
804 	 * the snapshot lock until the buffer has been written.
805 	 */
806 	VREF(vp);	/* Protect against ffs_snapgone() */
807 	VOP_UNLOCK(vp, 0, td);
808 	(void) bread(ip->i_devvp,
809 		     fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
810 		     (int) fs->fs_bsize, NOCRED, &nbp);
811 	brelse(nbp);
812 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
813 	if (ip->i_effnlink == 0)
814 		error = ENOENT;		/* Snapshot file unlinked */
815 	else
816 		vrele(vp);		/* Drop extra reference */
817 done:
818 	FREE(copy_fs->fs_csp, M_UFSMNT);
819 	bawrite(sbp);
820 out:
821 	if (saved_nice > 0) {
822 		PROC_LOCK(td->td_proc);
823 		mtx_lock_spin(&sched_lock);
824 		sched_nice(td->td_proc, saved_nice);
825 		mtx_unlock_spin(&sched_lock);
826 		PROC_UNLOCK(td->td_proc);
827 	}
828 	UFS_LOCK(ump);
829 	if (fs->fs_active != 0) {
830 		FREE(fs->fs_active, M_DEVBUF);
831 		fs->fs_active = 0;
832 	}
833 	UFS_UNLOCK(ump);
834 	MNT_ILOCK(mp);
835 	mp->mnt_flag = (mp->mnt_flag & MNT_QUOTA) | (flag & ~MNT_QUOTA);
836 	MNT_IUNLOCK(mp);
837 	if (error)
838 		(void) ffs_truncate(vp, (off_t)0, 0, NOCRED, td);
839 	(void) ffs_syncvnode(vp, MNT_WAIT);
840 	if (error)
841 		vput(vp);
842 	else
843 		VOP_UNLOCK(vp, 0, td);
844 	vrele(nd.ni_dvp);
845 	vn_finished_write(wrtmp);
846 	process_deferred_inactive(mp);
847 	return (error);
848 }
849 
850 /*
851  * Copy a cylinder group map. All the unallocated blocks are marked
852  * BLK_NOCOPY so that the snapshot knows that it need not copy them
853  * if they are later written. If passno is one, then this is a first
854  * pass, so only setting needs to be done. If passno is 2, then this
855  * is a revision to a previous pass which must be undone as the
856  * replacement pass is done.
857  */
858 static int
859 cgaccount(cg, vp, nbp, passno)
860 	int cg;
861 	struct vnode *vp;
862 	struct buf *nbp;
863 	int passno;
864 {
865 	struct buf *bp, *ibp;
866 	struct inode *ip;
867 	struct cg *cgp;
868 	struct fs *fs;
869 	ufs2_daddr_t base, numblks;
870 	int error, len, loc, indiroff;
871 
872 	ip = VTOI(vp);
873 	fs = ip->i_fs;
874 	error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)),
875 		(int)fs->fs_cgsize, KERNCRED, &bp);
876 	if (error) {
877 		brelse(bp);
878 		return (error);
879 	}
880 	cgp = (struct cg *)bp->b_data;
881 	if (!cg_chkmagic(cgp)) {
882 		brelse(bp);
883 		return (EIO);
884 	}
885 	UFS_LOCK(ip->i_ump);
886 	ACTIVESET(fs, cg);
887 	UFS_UNLOCK(ip->i_ump);
888 	bcopy(bp->b_data, nbp->b_data, fs->fs_cgsize);
889 	if (fs->fs_cgsize < fs->fs_bsize)
890 		bzero(&nbp->b_data[fs->fs_cgsize],
891 		    fs->fs_bsize - fs->fs_cgsize);
892 	cgp = (struct cg *)nbp->b_data;
893 	bqrelse(bp);
894 	if (passno == 2)
895 		nbp->b_flags |= B_VALIDSUSPWRT;
896 	numblks = howmany(fs->fs_size, fs->fs_frag);
897 	len = howmany(fs->fs_fpg, fs->fs_frag);
898 	base = cgbase(fs, cg) / fs->fs_frag;
899 	if (base + len >= numblks)
900 		len = numblks - base - 1;
901 	loc = 0;
902 	if (base < NDADDR) {
903 		for ( ; loc < NDADDR; loc++) {
904 			if (ffs_isblock(fs, cg_blksfree(cgp), loc))
905 				DIP_SET(ip, i_db[loc], BLK_NOCOPY);
906 			else if (passno == 2 && DIP(ip, i_db[loc])== BLK_NOCOPY)
907 				DIP_SET(ip, i_db[loc], 0);
908 			else if (passno == 1 && DIP(ip, i_db[loc])== BLK_NOCOPY)
909 				panic("ffs_snapshot: lost direct block");
910 		}
911 	}
912 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(base + loc)),
913 	    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
914 	if (error) {
915 		return (error);
916 	}
917 	indiroff = (base + loc - NDADDR) % NINDIR(fs);
918 	for ( ; loc < len; loc++, indiroff++) {
919 		if (indiroff >= NINDIR(fs)) {
920 			if (passno == 2)
921 				ibp->b_flags |= B_VALIDSUSPWRT;
922 			bawrite(ibp);
923 			error = UFS_BALLOC(vp,
924 			    lblktosize(fs, (off_t)(base + loc)),
925 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
926 			if (error) {
927 				return (error);
928 			}
929 			indiroff = 0;
930 		}
931 		if (ip->i_ump->um_fstype == UFS1) {
932 			if (ffs_isblock(fs, cg_blksfree(cgp), loc))
933 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] =
934 				    BLK_NOCOPY;
935 			else if (passno == 2 && ((ufs1_daddr_t *)(ibp->b_data))
936 			    [indiroff] == BLK_NOCOPY)
937 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] = 0;
938 			else if (passno == 1 && ((ufs1_daddr_t *)(ibp->b_data))
939 			    [indiroff] == BLK_NOCOPY)
940 				panic("ffs_snapshot: lost indirect block");
941 			continue;
942 		}
943 		if (ffs_isblock(fs, cg_blksfree(cgp), loc))
944 			((ufs2_daddr_t *)(ibp->b_data))[indiroff] = BLK_NOCOPY;
945 		else if (passno == 2 &&
946 		    ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY)
947 			((ufs2_daddr_t *)(ibp->b_data))[indiroff] = 0;
948 		else if (passno == 1 &&
949 		    ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY)
950 			panic("ffs_snapshot: lost indirect block");
951 	}
952 	if (passno == 2)
953 		ibp->b_flags |= B_VALIDSUSPWRT;
954 	bdwrite(ibp);
955 	return (0);
956 }
957 
958 /*
959  * Before expunging a snapshot inode, note all the
960  * blocks that it claims with BLK_SNAP so that fsck will
961  * be able to account for those blocks properly and so
962  * that this snapshot knows that it need not copy them
963  * if the other snapshot holding them is freed. This code
964  * is reproduced once each for UFS1 and UFS2.
965  */
966 static int
967 expunge_ufs1(snapvp, cancelip, fs, acctfunc, expungetype)
968 	struct vnode *snapvp;
969 	struct inode *cancelip;
970 	struct fs *fs;
971 	int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
972 	    struct fs *, ufs_lbn_t, int);
973 	int expungetype;
974 {
975 	int i, error, indiroff;
976 	ufs_lbn_t lbn, rlbn;
977 	ufs2_daddr_t len, blkno, numblks, blksperindir;
978 	struct ufs1_dinode *dip;
979 	struct thread *td = curthread;
980 	struct buf *bp;
981 
982 	/*
983 	 * Prepare to expunge the inode. If its inode block has not
984 	 * yet been copied, then allocate and fill the copy.
985 	 */
986 	lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number));
987 	blkno = 0;
988 	if (lbn < NDADDR) {
989 		blkno = VTOI(snapvp)->i_din1->di_db[lbn];
990 	} else {
991 		td->td_pflags |= TDP_COWINPROGRESS;
992 		error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn),
993 		   fs->fs_bsize, KERNCRED, BA_METAONLY, &bp);
994 		td->td_pflags &= ~TDP_COWINPROGRESS;
995 		if (error)
996 			return (error);
997 		indiroff = (lbn - NDADDR) % NINDIR(fs);
998 		blkno = ((ufs1_daddr_t *)(bp->b_data))[indiroff];
999 		bqrelse(bp);
1000 	}
1001 	if (blkno != 0) {
1002 		if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp)))
1003 			return (error);
1004 	} else {
1005 		error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn),
1006 		    fs->fs_bsize, KERNCRED, 0, &bp);
1007 		if (error)
1008 			return (error);
1009 		if ((error = readblock(snapvp, bp, lbn)) != 0)
1010 			return (error);
1011 	}
1012 	/*
1013 	 * Set a snapshot inode to be a zero length file, regular files
1014 	 * or unlinked snapshots to be completely unallocated.
1015 	 */
1016 	dip = (struct ufs1_dinode *)bp->b_data +
1017 	    ino_to_fsbo(fs, cancelip->i_number);
1018 	if (expungetype == BLK_NOCOPY || cancelip->i_effnlink == 0)
1019 		dip->di_mode = 0;
1020 	dip->di_size = 0;
1021 	dip->di_blocks = 0;
1022 	dip->di_flags &= ~SF_SNAPSHOT;
1023 	bzero(&dip->di_db[0], (NDADDR + NIADDR) * sizeof(ufs1_daddr_t));
1024 	bdwrite(bp);
1025 	/*
1026 	 * Now go through and expunge all the blocks in the file
1027 	 * using the function requested.
1028 	 */
1029 	numblks = howmany(cancelip->i_size, fs->fs_bsize);
1030 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_db[0],
1031 	    &cancelip->i_din1->di_db[NDADDR], fs, 0, expungetype)))
1032 		return (error);
1033 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_ib[0],
1034 	    &cancelip->i_din1->di_ib[NIADDR], fs, -1, expungetype)))
1035 		return (error);
1036 	blksperindir = 1;
1037 	lbn = -NDADDR;
1038 	len = numblks - NDADDR;
1039 	rlbn = NDADDR;
1040 	for (i = 0; len > 0 && i < NIADDR; i++) {
1041 		error = indiracct_ufs1(snapvp, ITOV(cancelip), i,
1042 		    cancelip->i_din1->di_ib[i], lbn, rlbn, len,
1043 		    blksperindir, fs, acctfunc, expungetype);
1044 		if (error)
1045 			return (error);
1046 		blksperindir *= NINDIR(fs);
1047 		lbn -= blksperindir + 1;
1048 		len -= blksperindir;
1049 		rlbn += blksperindir;
1050 	}
1051 	return (0);
1052 }
1053 
1054 /*
1055  * Descend an indirect block chain for vnode cancelvp accounting for all
1056  * its indirect blocks in snapvp.
1057  */
1058 static int
1059 indiracct_ufs1(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks,
1060 	    blksperindir, fs, acctfunc, expungetype)
1061 	struct vnode *snapvp;
1062 	struct vnode *cancelvp;
1063 	int level;
1064 	ufs1_daddr_t blkno;
1065 	ufs_lbn_t lbn;
1066 	ufs_lbn_t rlbn;
1067 	ufs_lbn_t remblks;
1068 	ufs_lbn_t blksperindir;
1069 	struct fs *fs;
1070 	int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
1071 	    struct fs *, ufs_lbn_t, int);
1072 	int expungetype;
1073 {
1074 	int error, num, i;
1075 	ufs_lbn_t subblksperindir;
1076 	struct indir indirs[NIADDR + 2];
1077 	ufs1_daddr_t last, *bap;
1078 	struct buf *bp;
1079 
1080 	if (blkno == 0) {
1081 		if (expungetype == BLK_NOCOPY)
1082 			return (0);
1083 		panic("indiracct_ufs1: missing indir");
1084 	}
1085 	if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0)
1086 		return (error);
1087 	if (lbn != indirs[num - 1 - level].in_lbn || num < 2)
1088 		panic("indiracct_ufs1: botched params");
1089 	/*
1090 	 * We have to expand bread here since it will deadlock looking
1091 	 * up the block number for any blocks that are not in the cache.
1092 	 */
1093 	bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0);
1094 	bp->b_blkno = fsbtodb(fs, blkno);
1095 	if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0 &&
1096 	    (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) {
1097 		brelse(bp);
1098 		return (error);
1099 	}
1100 	/*
1101 	 * Account for the block pointers in this indirect block.
1102 	 */
1103 	last = howmany(remblks, blksperindir);
1104 	if (last > NINDIR(fs))
1105 		last = NINDIR(fs);
1106 	MALLOC(bap, ufs1_daddr_t *, fs->fs_bsize, M_DEVBUF, M_WAITOK);
1107 	bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize);
1108 	bqrelse(bp);
1109 	error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs,
1110 	    level == 0 ? rlbn : -1, expungetype);
1111 	if (error || level == 0)
1112 		goto out;
1113 	/*
1114 	 * Account for the block pointers in each of the indirect blocks
1115 	 * in the levels below us.
1116 	 */
1117 	subblksperindir = blksperindir / NINDIR(fs);
1118 	for (lbn++, level--, i = 0; i < last; i++) {
1119 		error = indiracct_ufs1(snapvp, cancelvp, level, bap[i], lbn,
1120 		    rlbn, remblks, subblksperindir, fs, acctfunc, expungetype);
1121 		if (error)
1122 			goto out;
1123 		rlbn += blksperindir;
1124 		lbn -= blksperindir;
1125 		remblks -= blksperindir;
1126 	}
1127 out:
1128 	FREE(bap, M_DEVBUF);
1129 	return (error);
1130 }
1131 
1132 /*
1133  * Do both snap accounting and map accounting.
1134  */
1135 static int
1136 fullacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype)
1137 	struct vnode *vp;
1138 	ufs1_daddr_t *oldblkp, *lastblkp;
1139 	struct fs *fs;
1140 	ufs_lbn_t lblkno;
1141 	int exptype;	/* BLK_SNAP or BLK_NOCOPY */
1142 {
1143 	int error;
1144 
1145 	if ((error = snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype)))
1146 		return (error);
1147 	return (mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype));
1148 }
1149 
1150 /*
1151  * Identify a set of blocks allocated in a snapshot inode.
1152  */
1153 static int
1154 snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1155 	struct vnode *vp;
1156 	ufs1_daddr_t *oldblkp, *lastblkp;
1157 	struct fs *fs;
1158 	ufs_lbn_t lblkno;
1159 	int expungetype;	/* BLK_SNAP or BLK_NOCOPY */
1160 {
1161 	struct inode *ip = VTOI(vp);
1162 	ufs1_daddr_t blkno, *blkp;
1163 	ufs_lbn_t lbn;
1164 	struct buf *ibp;
1165 	int error;
1166 
1167 	for ( ; oldblkp < lastblkp; oldblkp++) {
1168 		blkno = *oldblkp;
1169 		if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP)
1170 			continue;
1171 		lbn = fragstoblks(fs, blkno);
1172 		if (lbn < NDADDR) {
1173 			blkp = &ip->i_din1->di_db[lbn];
1174 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1175 		} else {
1176 			error = ffs_balloc_ufs1(vp, lblktosize(fs, (off_t)lbn),
1177 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1178 			if (error)
1179 				return (error);
1180 			blkp = &((ufs1_daddr_t *)(ibp->b_data))
1181 			    [(lbn - NDADDR) % NINDIR(fs)];
1182 		}
1183 		/*
1184 		 * If we are expunging a snapshot vnode and we
1185 		 * find a block marked BLK_NOCOPY, then it is
1186 		 * one that has been allocated to this snapshot after
1187 		 * we took our current snapshot and can be ignored.
1188 		 */
1189 		if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) {
1190 			if (lbn >= NDADDR)
1191 				brelse(ibp);
1192 		} else {
1193 			if (*blkp != 0)
1194 				panic("snapacct_ufs1: bad block");
1195 			*blkp = expungetype;
1196 			if (lbn >= NDADDR)
1197 				bdwrite(ibp);
1198 		}
1199 	}
1200 	return (0);
1201 }
1202 
1203 /*
1204  * Account for a set of blocks allocated in a snapshot inode.
1205  */
1206 static int
1207 mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1208 	struct vnode *vp;
1209 	ufs1_daddr_t *oldblkp, *lastblkp;
1210 	struct fs *fs;
1211 	ufs_lbn_t lblkno;
1212 	int expungetype;
1213 {
1214 	ufs1_daddr_t blkno;
1215 	struct inode *ip;
1216 	ino_t inum;
1217 	int acctit;
1218 
1219 	ip = VTOI(vp);
1220 	inum = ip->i_number;
1221 	if (lblkno == -1)
1222 		acctit = 0;
1223 	else
1224 		acctit = 1;
1225 	for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) {
1226 		blkno = *oldblkp;
1227 		if (blkno == 0 || blkno == BLK_NOCOPY)
1228 			continue;
1229 		if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP)
1230 			*ip->i_snapblklist++ = lblkno;
1231 		if (blkno == BLK_SNAP)
1232 			blkno = blkstofrags(fs, lblkno);
1233 		ffs_blkfree(ip->i_ump, fs, vp, blkno, fs->fs_bsize, inum);
1234 	}
1235 	return (0);
1236 }
1237 
1238 /*
1239  * Before expunging a snapshot inode, note all the
1240  * blocks that it claims with BLK_SNAP so that fsck will
1241  * be able to account for those blocks properly and so
1242  * that this snapshot knows that it need not copy them
1243  * if the other snapshot holding them is freed. This code
1244  * is reproduced once each for UFS1 and UFS2.
1245  */
1246 static int
1247 expunge_ufs2(snapvp, cancelip, fs, acctfunc, expungetype)
1248 	struct vnode *snapvp;
1249 	struct inode *cancelip;
1250 	struct fs *fs;
1251 	int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
1252 	    struct fs *, ufs_lbn_t, int);
1253 	int expungetype;
1254 {
1255 	int i, error, indiroff;
1256 	ufs_lbn_t lbn, rlbn;
1257 	ufs2_daddr_t len, blkno, numblks, blksperindir;
1258 	struct ufs2_dinode *dip;
1259 	struct thread *td = curthread;
1260 	struct buf *bp;
1261 
1262 	/*
1263 	 * Prepare to expunge the inode. If its inode block has not
1264 	 * yet been copied, then allocate and fill the copy.
1265 	 */
1266 	lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number));
1267 	blkno = 0;
1268 	if (lbn < NDADDR) {
1269 		blkno = VTOI(snapvp)->i_din2->di_db[lbn];
1270 	} else {
1271 		td->td_pflags |= TDP_COWINPROGRESS;
1272 		error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn),
1273 		   fs->fs_bsize, KERNCRED, BA_METAONLY, &bp);
1274 		td->td_pflags &= ~TDP_COWINPROGRESS;
1275 		if (error)
1276 			return (error);
1277 		indiroff = (lbn - NDADDR) % NINDIR(fs);
1278 		blkno = ((ufs2_daddr_t *)(bp->b_data))[indiroff];
1279 		bqrelse(bp);
1280 	}
1281 	if (blkno != 0) {
1282 		if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp)))
1283 			return (error);
1284 	} else {
1285 		error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn),
1286 		    fs->fs_bsize, KERNCRED, 0, &bp);
1287 		if (error)
1288 			return (error);
1289 		if ((error = readblock(snapvp, bp, lbn)) != 0)
1290 			return (error);
1291 	}
1292 	/*
1293 	 * Set a snapshot inode to be a zero length file, regular files
1294 	 * to be completely unallocated.
1295 	 */
1296 	dip = (struct ufs2_dinode *)bp->b_data +
1297 	    ino_to_fsbo(fs, cancelip->i_number);
1298 	if (expungetype == BLK_NOCOPY)
1299 		dip->di_mode = 0;
1300 	dip->di_size = 0;
1301 	dip->di_blocks = 0;
1302 	dip->di_flags &= ~SF_SNAPSHOT;
1303 	bzero(&dip->di_db[0], (NDADDR + NIADDR) * sizeof(ufs2_daddr_t));
1304 	bdwrite(bp);
1305 	/*
1306 	 * Now go through and expunge all the blocks in the file
1307 	 * using the function requested.
1308 	 */
1309 	numblks = howmany(cancelip->i_size, fs->fs_bsize);
1310 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_db[0],
1311 	    &cancelip->i_din2->di_db[NDADDR], fs, 0, expungetype)))
1312 		return (error);
1313 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_ib[0],
1314 	    &cancelip->i_din2->di_ib[NIADDR], fs, -1, expungetype)))
1315 		return (error);
1316 	blksperindir = 1;
1317 	lbn = -NDADDR;
1318 	len = numblks - NDADDR;
1319 	rlbn = NDADDR;
1320 	for (i = 0; len > 0 && i < NIADDR; i++) {
1321 		error = indiracct_ufs2(snapvp, ITOV(cancelip), i,
1322 		    cancelip->i_din2->di_ib[i], lbn, rlbn, len,
1323 		    blksperindir, fs, acctfunc, expungetype);
1324 		if (error)
1325 			return (error);
1326 		blksperindir *= NINDIR(fs);
1327 		lbn -= blksperindir + 1;
1328 		len -= blksperindir;
1329 		rlbn += blksperindir;
1330 	}
1331 	return (0);
1332 }
1333 
1334 /*
1335  * Descend an indirect block chain for vnode cancelvp accounting for all
1336  * its indirect blocks in snapvp.
1337  */
1338 static int
1339 indiracct_ufs2(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks,
1340 	    blksperindir, fs, acctfunc, expungetype)
1341 	struct vnode *snapvp;
1342 	struct vnode *cancelvp;
1343 	int level;
1344 	ufs2_daddr_t blkno;
1345 	ufs_lbn_t lbn;
1346 	ufs_lbn_t rlbn;
1347 	ufs_lbn_t remblks;
1348 	ufs_lbn_t blksperindir;
1349 	struct fs *fs;
1350 	int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
1351 	    struct fs *, ufs_lbn_t, int);
1352 	int expungetype;
1353 {
1354 	int error, num, i;
1355 	ufs_lbn_t subblksperindir;
1356 	struct indir indirs[NIADDR + 2];
1357 	ufs2_daddr_t last, *bap;
1358 	struct buf *bp;
1359 
1360 	if (blkno == 0) {
1361 		if (expungetype == BLK_NOCOPY)
1362 			return (0);
1363 		panic("indiracct_ufs2: missing indir");
1364 	}
1365 	if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0)
1366 		return (error);
1367 	if (lbn != indirs[num - 1 - level].in_lbn || num < 2)
1368 		panic("indiracct_ufs2: botched params");
1369 	/*
1370 	 * We have to expand bread here since it will deadlock looking
1371 	 * up the block number for any blocks that are not in the cache.
1372 	 */
1373 	bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0);
1374 	bp->b_blkno = fsbtodb(fs, blkno);
1375 	if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0 &&
1376 	    (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) {
1377 		brelse(bp);
1378 		return (error);
1379 	}
1380 	/*
1381 	 * Account for the block pointers in this indirect block.
1382 	 */
1383 	last = howmany(remblks, blksperindir);
1384 	if (last > NINDIR(fs))
1385 		last = NINDIR(fs);
1386 	MALLOC(bap, ufs2_daddr_t *, fs->fs_bsize, M_DEVBUF, M_WAITOK);
1387 	bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize);
1388 	bqrelse(bp);
1389 	error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs,
1390 	    level == 0 ? rlbn : -1, expungetype);
1391 	if (error || level == 0)
1392 		goto out;
1393 	/*
1394 	 * Account for the block pointers in each of the indirect blocks
1395 	 * in the levels below us.
1396 	 */
1397 	subblksperindir = blksperindir / NINDIR(fs);
1398 	for (lbn++, level--, i = 0; i < last; i++) {
1399 		error = indiracct_ufs2(snapvp, cancelvp, level, bap[i], lbn,
1400 		    rlbn, remblks, subblksperindir, fs, acctfunc, expungetype);
1401 		if (error)
1402 			goto out;
1403 		rlbn += blksperindir;
1404 		lbn -= blksperindir;
1405 		remblks -= blksperindir;
1406 	}
1407 out:
1408 	FREE(bap, M_DEVBUF);
1409 	return (error);
1410 }
1411 
1412 /*
1413  * Do both snap accounting and map accounting.
1414  */
1415 static int
1416 fullacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype)
1417 	struct vnode *vp;
1418 	ufs2_daddr_t *oldblkp, *lastblkp;
1419 	struct fs *fs;
1420 	ufs_lbn_t lblkno;
1421 	int exptype;	/* BLK_SNAP or BLK_NOCOPY */
1422 {
1423 	int error;
1424 
1425 	if ((error = snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype)))
1426 		return (error);
1427 	return (mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype));
1428 }
1429 
1430 /*
1431  * Identify a set of blocks allocated in a snapshot inode.
1432  */
1433 static int
1434 snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1435 	struct vnode *vp;
1436 	ufs2_daddr_t *oldblkp, *lastblkp;
1437 	struct fs *fs;
1438 	ufs_lbn_t lblkno;
1439 	int expungetype;	/* BLK_SNAP or BLK_NOCOPY */
1440 {
1441 	struct inode *ip = VTOI(vp);
1442 	ufs2_daddr_t blkno, *blkp;
1443 	ufs_lbn_t lbn;
1444 	struct buf *ibp;
1445 	int error;
1446 
1447 	for ( ; oldblkp < lastblkp; oldblkp++) {
1448 		blkno = *oldblkp;
1449 		if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP)
1450 			continue;
1451 		lbn = fragstoblks(fs, blkno);
1452 		if (lbn < NDADDR) {
1453 			blkp = &ip->i_din2->di_db[lbn];
1454 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1455 		} else {
1456 			error = ffs_balloc_ufs2(vp, lblktosize(fs, (off_t)lbn),
1457 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1458 			if (error)
1459 				return (error);
1460 			blkp = &((ufs2_daddr_t *)(ibp->b_data))
1461 			    [(lbn - NDADDR) % NINDIR(fs)];
1462 		}
1463 		/*
1464 		 * If we are expunging a snapshot vnode and we
1465 		 * find a block marked BLK_NOCOPY, then it is
1466 		 * one that has been allocated to this snapshot after
1467 		 * we took our current snapshot and can be ignored.
1468 		 */
1469 		if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) {
1470 			if (lbn >= NDADDR)
1471 				brelse(ibp);
1472 		} else {
1473 			if (*blkp != 0)
1474 				panic("snapacct_ufs2: bad block");
1475 			*blkp = expungetype;
1476 			if (lbn >= NDADDR)
1477 				bdwrite(ibp);
1478 		}
1479 	}
1480 	return (0);
1481 }
1482 
1483 /*
1484  * Account for a set of blocks allocated in a snapshot inode.
1485  */
1486 static int
1487 mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1488 	struct vnode *vp;
1489 	ufs2_daddr_t *oldblkp, *lastblkp;
1490 	struct fs *fs;
1491 	ufs_lbn_t lblkno;
1492 	int expungetype;
1493 {
1494 	ufs2_daddr_t blkno;
1495 	struct inode *ip;
1496 	ino_t inum;
1497 	int acctit;
1498 
1499 	ip = VTOI(vp);
1500 	inum = ip->i_number;
1501 	if (lblkno == -1)
1502 		acctit = 0;
1503 	else
1504 		acctit = 1;
1505 	for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) {
1506 		blkno = *oldblkp;
1507 		if (blkno == 0 || blkno == BLK_NOCOPY)
1508 			continue;
1509 		if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP)
1510 			*ip->i_snapblklist++ = lblkno;
1511 		if (blkno == BLK_SNAP)
1512 			blkno = blkstofrags(fs, lblkno);
1513 		ffs_blkfree(ip->i_ump, fs, vp, blkno, fs->fs_bsize, inum);
1514 	}
1515 	return (0);
1516 }
1517 
1518 /*
1519  * Decrement extra reference on snapshot when last name is removed.
1520  * It will not be freed until the last open reference goes away.
1521  */
1522 void
1523 ffs_snapgone(ip)
1524 	struct inode *ip;
1525 {
1526 	struct inode *xp;
1527 	struct fs *fs;
1528 	int snaploc;
1529 	struct snapdata *sn;
1530 	struct ufsmount *ump;
1531 
1532 	/*
1533 	 * Find snapshot in incore list.
1534 	 */
1535 	xp = NULL;
1536 	sn = ip->i_devvp->v_rdev->si_snapdata;
1537 	if (sn != NULL)
1538 		TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap)
1539 			if (xp == ip)
1540 				break;
1541 	if (xp != NULL)
1542 		vrele(ITOV(ip));
1543 	else if (snapdebug)
1544 		printf("ffs_snapgone: lost snapshot vnode %d\n",
1545 		    ip->i_number);
1546 	/*
1547 	 * Delete snapshot inode from superblock. Keep list dense.
1548 	 */
1549 	fs = ip->i_fs;
1550 	ump = ip->i_ump;
1551 	UFS_LOCK(ump);
1552 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++)
1553 		if (fs->fs_snapinum[snaploc] == ip->i_number)
1554 			break;
1555 	if (snaploc < FSMAXSNAP) {
1556 		for (snaploc++; snaploc < FSMAXSNAP; snaploc++) {
1557 			if (fs->fs_snapinum[snaploc] == 0)
1558 				break;
1559 			fs->fs_snapinum[snaploc - 1] = fs->fs_snapinum[snaploc];
1560 		}
1561 		fs->fs_snapinum[snaploc - 1] = 0;
1562 	}
1563 	UFS_UNLOCK(ump);
1564 }
1565 
1566 /*
1567  * Prepare a snapshot file for being removed.
1568  */
1569 void
1570 ffs_snapremove(vp)
1571 	struct vnode *vp;
1572 {
1573 	struct inode *ip;
1574 	struct vnode *devvp;
1575 	struct buf *ibp;
1576 	struct fs *fs;
1577 	struct thread *td = curthread;
1578 	ufs2_daddr_t numblks, blkno, dblk;
1579 	int error, loc, last;
1580 	struct snapdata *sn;
1581 
1582 	ip = VTOI(vp);
1583 	fs = ip->i_fs;
1584 	devvp = ip->i_devvp;
1585 	/*
1586 	 * If active, delete from incore list (this snapshot may
1587 	 * already have been in the process of being deleted, so
1588 	 * would not have been active).
1589 	 *
1590 	 * Clear copy-on-write flag if last snapshot.
1591 	 */
1592 	VI_LOCK(devvp);
1593 	if (ip->i_nextsnap.tqe_prev != 0) {
1594 		sn = devvp->v_rdev->si_snapdata;
1595 		TAILQ_REMOVE(&sn->sn_head, ip, i_nextsnap);
1596 		ip->i_nextsnap.tqe_prev = 0;
1597 		VI_UNLOCK(devvp);
1598 		lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL, td);
1599 		VI_LOCK(vp);
1600 		KASSERT(vp->v_vnlock == &sn->sn_lock,
1601 			("ffs_snapremove: lost lock mutation"));
1602 		vp->v_vnlock = &vp->v_lock;
1603 		VI_UNLOCK(vp);
1604 		VI_LOCK(devvp);
1605 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL, td);
1606 		try_free_snapdata(devvp, td);
1607 	} else
1608 		VI_UNLOCK(devvp);
1609 	/*
1610 	 * Clear all BLK_NOCOPY fields. Pass any block claims to other
1611 	 * snapshots that want them (see ffs_snapblkfree below).
1612 	 */
1613 	for (blkno = 1; blkno < NDADDR; blkno++) {
1614 		dblk = DIP(ip, i_db[blkno]);
1615 		if (dblk == 0)
1616 			continue;
1617 		if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1618 			DIP_SET(ip, i_db[blkno], 0);
1619 		else if ((dblk == blkstofrags(fs, blkno) &&
1620 		     ffs_snapblkfree(fs, ip->i_devvp, dblk, fs->fs_bsize,
1621 		     ip->i_number))) {
1622 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) -
1623 			    btodb(fs->fs_bsize));
1624 			DIP_SET(ip, i_db[blkno], 0);
1625 		}
1626 	}
1627 	numblks = howmany(ip->i_size, fs->fs_bsize);
1628 	for (blkno = NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
1629 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
1630 		    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1631 		if (error)
1632 			continue;
1633 		if (fs->fs_size - blkno > NINDIR(fs))
1634 			last = NINDIR(fs);
1635 		else
1636 			last = fs->fs_size - blkno;
1637 		for (loc = 0; loc < last; loc++) {
1638 			if (ip->i_ump->um_fstype == UFS1) {
1639 				dblk = ((ufs1_daddr_t *)(ibp->b_data))[loc];
1640 				if (dblk == 0)
1641 					continue;
1642 				if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1643 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1644 				else if ((dblk == blkstofrags(fs, blkno) &&
1645 				     ffs_snapblkfree(fs, ip->i_devvp, dblk,
1646 				     fs->fs_bsize, ip->i_number))) {
1647 					ip->i_din1->di_blocks -=
1648 					    btodb(fs->fs_bsize);
1649 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1650 				}
1651 				continue;
1652 			}
1653 			dblk = ((ufs2_daddr_t *)(ibp->b_data))[loc];
1654 			if (dblk == 0)
1655 				continue;
1656 			if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1657 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1658 			else if ((dblk == blkstofrags(fs, blkno) &&
1659 			     ffs_snapblkfree(fs, ip->i_devvp, dblk,
1660 			     fs->fs_bsize, ip->i_number))) {
1661 				ip->i_din2->di_blocks -= btodb(fs->fs_bsize);
1662 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1663 			}
1664 		}
1665 		bawrite(ibp);
1666 	}
1667 	/*
1668 	 * Clear snapshot flag and drop reference.
1669 	 */
1670 	ip->i_flags &= ~SF_SNAPSHOT;
1671 	DIP_SET(ip, i_flags, ip->i_flags);
1672 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
1673 #ifdef QUOTA
1674 	/*
1675 	 * Reenable disk quotas for ex-snapshot file.
1676 	 */
1677 	if (!getinoquota(ip))
1678 		(void) chkdq(ip, DIP(ip, i_blocks), KERNCRED, FORCE);
1679 #endif
1680 }
1681 
1682 /*
1683  * Notification that a block is being freed. Return zero if the free
1684  * should be allowed to proceed. Return non-zero if the snapshot file
1685  * wants to claim the block. The block will be claimed if it is an
1686  * uncopied part of one of the snapshots. It will be freed if it is
1687  * either a BLK_NOCOPY or has already been copied in all of the snapshots.
1688  * If a fragment is being freed, then all snapshots that care about
1689  * it must make a copy since a snapshot file can only claim full sized
1690  * blocks. Note that if more than one snapshot file maps the block,
1691  * we can pick one at random to claim it. Since none of the snapshots
1692  * can change, we are assurred that they will all see the same unmodified
1693  * image. When deleting a snapshot file (see ffs_snapremove above), we
1694  * must push any of these claimed blocks to one of the other snapshots
1695  * that maps it. These claimed blocks are easily identified as they will
1696  * have a block number equal to their logical block number within the
1697  * snapshot. A copied block can never have this property because they
1698  * must always have been allocated from a BLK_NOCOPY location.
1699  */
1700 int
1701 ffs_snapblkfree(fs, devvp, bno, size, inum)
1702 	struct fs *fs;
1703 	struct vnode *devvp;
1704 	ufs2_daddr_t bno;
1705 	long size;
1706 	ino_t inum;
1707 {
1708 	struct buf *ibp, *cbp, *savedcbp = 0;
1709 	struct thread *td = curthread;
1710 	struct inode *ip;
1711 	struct vnode *vp = NULL;
1712 	ufs_lbn_t lbn;
1713 	ufs2_daddr_t blkno;
1714 	int indiroff = 0, error = 0, claimedblk = 0;
1715 	struct snapdata *sn;
1716 
1717 	lbn = fragstoblks(fs, bno);
1718 retry:
1719 	VI_LOCK(devvp);
1720 	sn = devvp->v_rdev->si_snapdata;
1721 	if (sn == NULL) {
1722 		VI_UNLOCK(devvp);
1723 		return (0);
1724 	}
1725 	if (lockmgr(&sn->sn_lock,
1726 		    LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
1727 		    VI_MTX(devvp), td) != 0)
1728 		goto retry;
1729 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
1730 		vp = ITOV(ip);
1731 		/*
1732 		 * Lookup block being written.
1733 		 */
1734 		if (lbn < NDADDR) {
1735 			blkno = DIP(ip, i_db[lbn]);
1736 		} else {
1737 			td->td_pflags |= TDP_COWINPROGRESS;
1738 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1739 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1740 			td->td_pflags &= ~TDP_COWINPROGRESS;
1741 			if (error)
1742 				break;
1743 			indiroff = (lbn - NDADDR) % NINDIR(fs);
1744 			if (ip->i_ump->um_fstype == UFS1)
1745 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
1746 			else
1747 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
1748 		}
1749 		/*
1750 		 * Check to see if block needs to be copied.
1751 		 */
1752 		if (blkno == 0) {
1753 			/*
1754 			 * A block that we map is being freed. If it has not
1755 			 * been claimed yet, we will claim or copy it (below).
1756 			 */
1757 			claimedblk = 1;
1758 		} else if (blkno == BLK_SNAP) {
1759 			/*
1760 			 * No previous snapshot claimed the block,
1761 			 * so it will be freed and become a BLK_NOCOPY
1762 			 * (don't care) for us.
1763 			 */
1764 			if (claimedblk)
1765 				panic("snapblkfree: inconsistent block type");
1766 			if (lbn < NDADDR) {
1767 				DIP_SET(ip, i_db[lbn], BLK_NOCOPY);
1768 				ip->i_flag |= IN_CHANGE | IN_UPDATE;
1769 			} else if (ip->i_ump->um_fstype == UFS1) {
1770 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] =
1771 				    BLK_NOCOPY;
1772 				bdwrite(ibp);
1773 			} else {
1774 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] =
1775 				    BLK_NOCOPY;
1776 				bdwrite(ibp);
1777 			}
1778 			continue;
1779 		} else /* BLK_NOCOPY or default */ {
1780 			/*
1781 			 * If the snapshot has already copied the block
1782 			 * (default), or does not care about the block,
1783 			 * it is not needed.
1784 			 */
1785 			if (lbn >= NDADDR)
1786 				bqrelse(ibp);
1787 			continue;
1788 		}
1789 		/*
1790 		 * If this is a full size block, we will just grab it
1791 		 * and assign it to the snapshot inode. Otherwise we
1792 		 * will proceed to copy it. See explanation for this
1793 		 * routine as to why only a single snapshot needs to
1794 		 * claim this block.
1795 		 */
1796 		if (size == fs->fs_bsize) {
1797 #ifdef DEBUG
1798 			if (snapdebug)
1799 				printf("%s %d lbn %jd from inum %d\n",
1800 				    "Grabonremove: snapino", ip->i_number,
1801 				    (intmax_t)lbn, inum);
1802 #endif
1803 			if (lbn < NDADDR) {
1804 				DIP_SET(ip, i_db[lbn], bno);
1805 			} else if (ip->i_ump->um_fstype == UFS1) {
1806 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] = bno;
1807 				bdwrite(ibp);
1808 			} else {
1809 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] = bno;
1810 				bdwrite(ibp);
1811 			}
1812 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + btodb(size));
1813 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1814 			lockmgr(vp->v_vnlock, LK_RELEASE, NULL, td);
1815 			return (1);
1816 		}
1817 		if (lbn >= NDADDR)
1818 			bqrelse(ibp);
1819 		/*
1820 		 * Allocate the block into which to do the copy. Note that this
1821 		 * allocation will never require any additional allocations for
1822 		 * the snapshot inode.
1823 		 */
1824 		td->td_pflags |= TDP_COWINPROGRESS;
1825 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1826 		    fs->fs_bsize, KERNCRED, 0, &cbp);
1827 		td->td_pflags &= ~TDP_COWINPROGRESS;
1828 		if (error)
1829 			break;
1830 #ifdef DEBUG
1831 		if (snapdebug)
1832 			printf("%s%d lbn %jd %s %d size %ld to blkno %jd\n",
1833 			    "Copyonremove: snapino ", ip->i_number,
1834 			    (intmax_t)lbn, "for inum", inum, size,
1835 			    (intmax_t)cbp->b_blkno);
1836 #endif
1837 		/*
1838 		 * If we have already read the old block contents, then
1839 		 * simply copy them to the new block. Note that we need
1840 		 * to synchronously write snapshots that have not been
1841 		 * unlinked, and hence will be visible after a crash,
1842 		 * to ensure their integrity.
1843 		 */
1844 		if (savedcbp != 0) {
1845 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
1846 			bawrite(cbp);
1847 			if (dopersistence && ip->i_effnlink > 0)
1848 				(void) ffs_syncvnode(vp, MNT_WAIT);
1849 			continue;
1850 		}
1851 		/*
1852 		 * Otherwise, read the old block contents into the buffer.
1853 		 */
1854 		if ((error = readblock(vp, cbp, lbn)) != 0) {
1855 			bzero(cbp->b_data, fs->fs_bsize);
1856 			bawrite(cbp);
1857 			if (dopersistence && ip->i_effnlink > 0)
1858 				(void) ffs_syncvnode(vp, MNT_WAIT);
1859 			break;
1860 		}
1861 		savedcbp = cbp;
1862 	}
1863 	/*
1864 	 * Note that we need to synchronously write snapshots that
1865 	 * have not been unlinked, and hence will be visible after
1866 	 * a crash, to ensure their integrity.
1867 	 */
1868 	if (savedcbp) {
1869 		vp = savedcbp->b_vp;
1870 		bawrite(savedcbp);
1871 		if (dopersistence && VTOI(vp)->i_effnlink > 0)
1872 			(void) ffs_syncvnode(vp, MNT_WAIT);
1873 	}
1874 	/*
1875 	 * If we have been unable to allocate a block in which to do
1876 	 * the copy, then return non-zero so that the fragment will
1877 	 * not be freed. Although space will be lost, the snapshot
1878 	 * will stay consistent.
1879 	 */
1880 	lockmgr(vp->v_vnlock, LK_RELEASE, NULL, td);
1881 	return (error);
1882 }
1883 
1884 /*
1885  * Associate snapshot files when mounting.
1886  */
1887 void
1888 ffs_snapshot_mount(mp)
1889 	struct mount *mp;
1890 {
1891 	struct ufsmount *ump = VFSTOUFS(mp);
1892 	struct vnode *devvp = ump->um_devvp;
1893 	struct fs *fs = ump->um_fs;
1894 	struct thread *td = curthread;
1895 	struct snapdata *sn;
1896 	struct vnode *vp;
1897 	struct vnode *lastvp;
1898 	struct inode *ip;
1899 	struct uio auio;
1900 	struct iovec aiov;
1901 	void *snapblklist;
1902 	char *reason;
1903 	daddr_t snaplistsize;
1904 	int error, snaploc, loc;
1905 
1906 	/*
1907 	 * XXX The following needs to be set before ffs_truncate or
1908 	 * VOP_READ can be called.
1909 	 */
1910 	mp->mnt_stat.f_iosize = fs->fs_bsize;
1911 	/*
1912 	 * Process each snapshot listed in the superblock.
1913 	 */
1914 	vp = NULL;
1915 	lastvp = NULL;
1916 	sn = devvp->v_rdev->si_snapdata;
1917 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++) {
1918 		if (fs->fs_snapinum[snaploc] == 0)
1919 			break;
1920 		if ((error = ffs_vget(mp, fs->fs_snapinum[snaploc],
1921 		    LK_EXCLUSIVE, &vp)) != 0){
1922 			printf("ffs_snapshot_mount: vget failed %d\n", error);
1923 			continue;
1924 		}
1925 		ip = VTOI(vp);
1926 		if ((ip->i_flags & SF_SNAPSHOT) == 0 || ip->i_size ==
1927 		    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag))) {
1928 			if ((ip->i_flags & SF_SNAPSHOT) == 0) {
1929 				reason = "non-snapshot";
1930 			} else {
1931 				reason = "old format snapshot";
1932 				(void)ffs_truncate(vp, (off_t)0, 0, NOCRED, td);
1933 				(void)ffs_syncvnode(vp, MNT_WAIT);
1934 			}
1935 			printf("ffs_snapshot_mount: %s inode %d\n",
1936 			    reason, fs->fs_snapinum[snaploc]);
1937 			vput(vp);
1938 			vp = NULL;
1939 			for (loc = snaploc + 1; loc < FSMAXSNAP; loc++) {
1940 				if (fs->fs_snapinum[loc] == 0)
1941 					break;
1942 				fs->fs_snapinum[loc - 1] = fs->fs_snapinum[loc];
1943 			}
1944 			fs->fs_snapinum[loc - 1] = 0;
1945 			snaploc--;
1946 			continue;
1947 		}
1948 		/*
1949 		 * If there already exist snapshots on this filesystem, grab a
1950 		 * reference to their shared lock. If this is the first snapshot
1951 		 * on this filesystem, we need to allocate a lock for the
1952 		 * snapshots to share. In either case, acquire the snapshot
1953 		 * lock and give up our original private lock.
1954 		 */
1955 		VI_LOCK(devvp);
1956 		if (sn != NULL) {
1957 
1958 			VI_UNLOCK(devvp);
1959 			VI_LOCK(vp);
1960 			vp->v_vnlock = &sn->sn_lock;
1961 		} else {
1962 			VI_UNLOCK(devvp);
1963 			sn = malloc(sizeof *sn, M_UFSMNT, M_WAITOK | M_ZERO);
1964 			TAILQ_INIT(&sn->sn_head);
1965 			lockinit(&sn->sn_lock, PVFS, "snaplk", VLKTIMEOUT,
1966 			    LK_CANRECURSE | LK_NOSHARE);
1967 			VI_LOCK(vp);
1968 			vp->v_vnlock = &sn->sn_lock;
1969 			devvp->v_rdev->si_snapdata = sn;
1970 		}
1971 		lockmgr(vp->v_vnlock, LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY,
1972 		    VI_MTX(vp), td);
1973 		lockmgr(&vp->v_lock, LK_RELEASE, NULL, td);
1974 		/*
1975 		 * Link it onto the active snapshot list.
1976 		 */
1977 		VI_LOCK(devvp);
1978 		if (ip->i_nextsnap.tqe_prev != 0)
1979 			panic("ffs_snapshot_mount: %d already on list",
1980 			    ip->i_number);
1981 		else
1982 			TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
1983 		vp->v_vflag |= VV_SYSTEM;
1984 		VI_UNLOCK(devvp);
1985 		VOP_UNLOCK(vp, 0, td);
1986 		lastvp = vp;
1987 	}
1988 	vp = lastvp;
1989 	/*
1990 	 * No usable snapshots found.
1991 	 */
1992 	if (vp == NULL)
1993 		return;
1994 	/*
1995 	 * Allocate the space for the block hints list. We always want to
1996 	 * use the list from the newest snapshot.
1997 	 */
1998 	auio.uio_iov = &aiov;
1999 	auio.uio_iovcnt = 1;
2000 	aiov.iov_base = (void *)&snaplistsize;
2001 	aiov.iov_len = sizeof(snaplistsize);
2002 	auio.uio_resid = aiov.iov_len;
2003 	auio.uio_offset =
2004 	    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag));
2005 	auio.uio_segflg = UIO_SYSSPACE;
2006 	auio.uio_rw = UIO_READ;
2007 	auio.uio_td = td;
2008 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2009 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
2010 		printf("ffs_snapshot_mount: read_1 failed %d\n", error);
2011 		VOP_UNLOCK(vp, 0, td);
2012 		return;
2013 	}
2014 	MALLOC(snapblklist, void *, snaplistsize * sizeof(daddr_t),
2015 	    M_UFSMNT, M_WAITOK);
2016 	auio.uio_iovcnt = 1;
2017 	aiov.iov_base = snapblklist;
2018 	aiov.iov_len = snaplistsize * sizeof (daddr_t);
2019 	auio.uio_resid = aiov.iov_len;
2020 	auio.uio_offset -= sizeof(snaplistsize);
2021 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
2022 		printf("ffs_snapshot_mount: read_2 failed %d\n", error);
2023 		VOP_UNLOCK(vp, 0, td);
2024 		FREE(snapblklist, M_UFSMNT);
2025 		return;
2026 	}
2027 	VOP_UNLOCK(vp, 0, td);
2028 	VI_LOCK(devvp);
2029 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_mount");
2030 	sn->sn_listsize = snaplistsize;
2031 	sn->sn_blklist = (daddr_t *)snapblklist;
2032 	devvp->v_vflag |= VV_COPYONWRITE;
2033 	VI_UNLOCK(devvp);
2034 }
2035 
2036 /*
2037  * Disassociate snapshot files when unmounting.
2038  */
2039 void
2040 ffs_snapshot_unmount(mp)
2041 	struct mount *mp;
2042 {
2043 	struct vnode *devvp = VFSTOUFS(mp)->um_devvp;
2044 	struct snapdata *sn;
2045 	struct inode *xp;
2046 	struct vnode *vp;
2047 	struct thread *td = curthread;
2048 
2049 	VI_LOCK(devvp);
2050 	sn = devvp->v_rdev->si_snapdata;
2051 	while (sn != NULL && (xp = TAILQ_FIRST(&sn->sn_head)) != NULL) {
2052 		vp = ITOV(xp);
2053 		TAILQ_REMOVE(&sn->sn_head, xp, i_nextsnap);
2054 		xp->i_nextsnap.tqe_prev = 0;
2055 		lockmgr(&sn->sn_lock,
2056 			LK_INTERLOCK | LK_EXCLUSIVE,
2057 			VI_MTX(devvp),
2058 			td);
2059 		VI_LOCK(vp);
2060 		lockmgr(&vp->v_lock,
2061 			LK_INTERLOCK | LK_EXCLUSIVE,
2062 			VI_MTX(vp), td);
2063 		VI_LOCK(vp);
2064 		KASSERT(vp->v_vnlock == &sn->sn_lock,
2065 		("ffs_snapshot_unmount: lost lock mutation"));
2066 		vp->v_vnlock = &vp->v_lock;
2067 		VI_UNLOCK(vp);
2068 		lockmgr(&vp->v_lock, LK_RELEASE, NULL, td);
2069 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL, td);
2070 		if (xp->i_effnlink > 0)
2071 			vrele(vp);
2072 		VI_LOCK(devvp);
2073 		sn = devvp->v_rdev->si_snapdata;
2074 	}
2075 	try_free_snapdata(devvp, td);
2076 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_unmount");
2077 }
2078 
2079 /*
2080  * Check for need to copy block that is about to be written,
2081  * copying the block if necessary.
2082  */
2083 int
2084 ffs_copyonwrite(devvp, bp)
2085 	struct vnode *devvp;
2086 	struct buf *bp;
2087 {
2088 	struct snapdata *sn;
2089 	struct buf *ibp, *cbp, *savedcbp = 0;
2090 	struct thread *td = curthread;
2091 	struct fs *fs;
2092 	struct inode *ip;
2093 	struct vnode *vp = 0;
2094 	ufs2_daddr_t lbn, blkno, *snapblklist;
2095 	int lower, upper, mid, indiroff, error = 0;
2096 	int launched_async_io, prev_norunningbuf;
2097 	long saved_runningbufspace;
2098 
2099 	if (devvp != bp->b_vp && (VTOI(bp->b_vp)->i_flags & SF_SNAPSHOT) != 0)
2100 		return (0);		/* Update on a snapshot file */
2101 	if (td->td_pflags & TDP_COWINPROGRESS)
2102 		panic("ffs_copyonwrite: recursive call");
2103 	/*
2104 	 * First check to see if it is in the preallocated list.
2105 	 * By doing this check we avoid several potential deadlocks.
2106 	 */
2107 	VI_LOCK(devvp);
2108 	sn = devvp->v_rdev->si_snapdata;
2109 	if (sn == NULL ||
2110 	    TAILQ_FIRST(&sn->sn_head) == NULL) {
2111 		VI_UNLOCK(devvp);
2112 		return (0);		/* No snapshot */
2113 	}
2114 	ip = TAILQ_FIRST(&sn->sn_head);
2115 	fs = ip->i_fs;
2116 	lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno));
2117 	snapblklist = sn->sn_blklist;
2118 	upper = sn->sn_listsize - 1;
2119 	lower = 1;
2120 	while (lower <= upper) {
2121 		mid = (lower + upper) / 2;
2122 		if (snapblklist[mid] == lbn)
2123 			break;
2124 		if (snapblklist[mid] < lbn)
2125 			lower = mid + 1;
2126 		else
2127 			upper = mid - 1;
2128 	}
2129 	if (lower <= upper) {
2130 		VI_UNLOCK(devvp);
2131 		return (0);
2132 	}
2133 	launched_async_io = 0;
2134 	prev_norunningbuf = td->td_pflags & TDP_NORUNNINGBUF;
2135 	/*
2136 	 * Since I/O on bp isn't yet in progress and it may be blocked
2137 	 * for a long time waiting on snaplk, back it out of
2138 	 * runningbufspace, possibly waking other threads waiting for space.
2139 	 */
2140 	saved_runningbufspace = bp->b_runningbufspace;
2141 	if (saved_runningbufspace != 0)
2142 		runningbufwakeup(bp);
2143 	/*
2144 	 * Not in the precomputed list, so check the snapshots.
2145 	 */
2146 	while (lockmgr(&sn->sn_lock,
2147 		       LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
2148 		       VI_MTX(devvp), td) != 0) {
2149 		VI_LOCK(devvp);
2150 		sn = devvp->v_rdev->si_snapdata;
2151 		if (sn == NULL ||
2152 		    TAILQ_FIRST(&sn->sn_head) == NULL) {
2153 			VI_UNLOCK(devvp);
2154 			if (saved_runningbufspace != 0) {
2155 				bp->b_runningbufspace = saved_runningbufspace;
2156 				atomic_add_int(&runningbufspace,
2157 					       bp->b_runningbufspace);
2158 			}
2159 			return (0);		/* Snapshot gone */
2160 		}
2161 	}
2162 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
2163 		vp = ITOV(ip);
2164 		/*
2165 		 * We ensure that everything of our own that needs to be
2166 		 * copied will be done at the time that ffs_snapshot is
2167 		 * called. Thus we can skip the check here which can
2168 		 * deadlock in doing the lookup in UFS_BALLOC.
2169 		 */
2170 		if (bp->b_vp == vp)
2171 			continue;
2172 		/*
2173 		 * Check to see if block needs to be copied. We do not have
2174 		 * to hold the snapshot lock while doing this lookup as it
2175 		 * will never require any additional allocations for the
2176 		 * snapshot inode.
2177 		 */
2178 		if (lbn < NDADDR) {
2179 			blkno = DIP(ip, i_db[lbn]);
2180 		} else {
2181 			td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2182 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2183 			   fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
2184 			td->td_pflags &= ~TDP_COWINPROGRESS;
2185 			if (error)
2186 				break;
2187 			indiroff = (lbn - NDADDR) % NINDIR(fs);
2188 			if (ip->i_ump->um_fstype == UFS1)
2189 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
2190 			else
2191 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
2192 			bqrelse(ibp);
2193 		}
2194 #ifdef DIAGNOSTIC
2195 		if (blkno == BLK_SNAP && bp->b_lblkno >= 0)
2196 			panic("ffs_copyonwrite: bad copy block");
2197 #endif
2198 		if (blkno != 0)
2199 			continue;
2200 		/*
2201 		 * Allocate the block into which to do the copy. Since
2202 		 * multiple processes may all try to copy the same block,
2203 		 * we have to recheck our need to do a copy if we sleep
2204 		 * waiting for the lock.
2205 		 *
2206 		 * Because all snapshots on a filesystem share a single
2207 		 * lock, we ensure that we will never be in competition
2208 		 * with another process to allocate a block.
2209 		 */
2210 		td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2211 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2212 		    fs->fs_bsize, KERNCRED, 0, &cbp);
2213 		td->td_pflags &= ~TDP_COWINPROGRESS;
2214 		if (error)
2215 			break;
2216 #ifdef DEBUG
2217 		if (snapdebug) {
2218 			printf("Copyonwrite: snapino %d lbn %jd for ",
2219 			    ip->i_number, (intmax_t)lbn);
2220 			if (bp->b_vp == devvp)
2221 				printf("fs metadata");
2222 			else
2223 				printf("inum %d", VTOI(bp->b_vp)->i_number);
2224 			printf(" lblkno %jd to blkno %jd\n",
2225 			    (intmax_t)bp->b_lblkno, (intmax_t)cbp->b_blkno);
2226 		}
2227 #endif
2228 		/*
2229 		 * If we have already read the old block contents, then
2230 		 * simply copy them to the new block. Note that we need
2231 		 * to synchronously write snapshots that have not been
2232 		 * unlinked, and hence will be visible after a crash,
2233 		 * to ensure their integrity.
2234 		 */
2235 		if (savedcbp != 0) {
2236 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
2237 			bawrite(cbp);
2238 			if (dopersistence && ip->i_effnlink > 0)
2239 				(void) ffs_syncvnode(vp, MNT_WAIT);
2240 			else
2241 				launched_async_io = 1;
2242 			continue;
2243 		}
2244 		/*
2245 		 * Otherwise, read the old block contents into the buffer.
2246 		 */
2247 		if ((error = readblock(vp, cbp, lbn)) != 0) {
2248 			bzero(cbp->b_data, fs->fs_bsize);
2249 			bawrite(cbp);
2250 			if (dopersistence && ip->i_effnlink > 0)
2251 				(void) ffs_syncvnode(vp, MNT_WAIT);
2252 			else
2253 				launched_async_io = 1;
2254 			break;
2255 		}
2256 		savedcbp = cbp;
2257 	}
2258 	/*
2259 	 * Note that we need to synchronously write snapshots that
2260 	 * have not been unlinked, and hence will be visible after
2261 	 * a crash, to ensure their integrity.
2262 	 */
2263 	if (savedcbp) {
2264 		vp = savedcbp->b_vp;
2265 		bawrite(savedcbp);
2266 		if (dopersistence && VTOI(vp)->i_effnlink > 0)
2267 			(void) ffs_syncvnode(vp, MNT_WAIT);
2268 		else
2269 			launched_async_io = 1;
2270 	}
2271 	lockmgr(vp->v_vnlock, LK_RELEASE, NULL, td);
2272 	td->td_pflags = (td->td_pflags & ~TDP_NORUNNINGBUF) |
2273 		prev_norunningbuf;
2274 	if (launched_async_io && (td->td_pflags & TDP_NORUNNINGBUF) == 0)
2275 		waitrunningbufspace();
2276 	/*
2277 	 * I/O on bp will now be started, so count it in runningbufspace.
2278 	 */
2279 	if (saved_runningbufspace != 0) {
2280 		bp->b_runningbufspace = saved_runningbufspace;
2281 		atomic_add_int(&runningbufspace, bp->b_runningbufspace);
2282 	}
2283 	return (error);
2284 }
2285 
2286 /*
2287  * Read the specified block into the given buffer.
2288  * Much of this boiler-plate comes from bwrite().
2289  */
2290 static int
2291 readblock(vp, bp, lbn)
2292 	struct vnode *vp;
2293 	struct buf *bp;
2294 	ufs2_daddr_t lbn;
2295 {
2296 	struct inode *ip = VTOI(vp);
2297 	struct bio *bip;
2298 
2299 	bip = g_alloc_bio();
2300 	bip->bio_cmd = BIO_READ;
2301 	bip->bio_offset = dbtob(fsbtodb(ip->i_fs, blkstofrags(ip->i_fs, lbn)));
2302 	bip->bio_data = bp->b_data;
2303 	bip->bio_length = bp->b_bcount;
2304 	bip->bio_done = NULL;
2305 
2306 	g_io_request(bip, ip->i_devvp->v_bufobj.bo_private);
2307 	bp->b_error = biowait(bip, "snaprdb");
2308 	g_destroy_bio(bip);
2309 	return (bp->b_error);
2310 }
2311 
2312 /*
2313  * Process file deletes that were deferred by ufs_inactive() due to
2314  * the file system being suspended. Transfer IN_LAZYACCESS into
2315  * IN_MODIFIED for vnodes that were accessed during suspension.
2316  */
2317 static void
2318 process_deferred_inactive(struct mount *mp)
2319 {
2320 	struct vnode *vp, *mvp;
2321 	struct inode *ip;
2322 	struct thread *td;
2323 	int error;
2324 
2325 	td = curthread;
2326 	(void) vn_start_secondary_write(NULL, &mp, V_WAIT);
2327 	MNT_ILOCK(mp);
2328  loop:
2329 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2330 		VI_LOCK(vp);
2331 		/*
2332 		 * IN_LAZYACCESS is checked here without holding any
2333 		 * vnode lock, but this flag is set only while holding
2334 		 * vnode interlock.
2335 		 */
2336 		if (vp->v_type == VNON || (vp->v_iflag & VI_DOOMED) != 0 ||
2337 		    ((VTOI(vp)->i_flag & IN_LAZYACCESS) == 0 &&
2338 			((vp->v_iflag & VI_OWEINACT) == 0 ||
2339 			vp->v_usecount > 0))) {
2340 			VI_UNLOCK(vp);
2341 			continue;
2342 		}
2343 		MNT_IUNLOCK(mp);
2344 		vholdl(vp);
2345 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td);
2346 		if (error != 0) {
2347 			vdrop(vp);
2348 			MNT_ILOCK(mp);
2349 			if (error == ENOENT)
2350 				continue;	/* vnode recycled */
2351 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2352 			goto loop;
2353 		}
2354 		ip = VTOI(vp);
2355 		if ((ip->i_flag & IN_LAZYACCESS) != 0) {
2356 			ip->i_flag &= ~IN_LAZYACCESS;
2357 			ip->i_flag |= IN_MODIFIED;
2358 		}
2359 		VI_LOCK(vp);
2360 		if ((vp->v_iflag & VI_OWEINACT) == 0 || vp->v_usecount > 0) {
2361 			VI_UNLOCK(vp);
2362 			VOP_UNLOCK(vp, 0, td);
2363 			vdrop(vp);
2364 			MNT_ILOCK(mp);
2365 			continue;
2366 		}
2367 
2368 		VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2369 			 ("process_deferred_inactive: "
2370 			  "recursed on VI_DOINGINACT"));
2371 		vp->v_iflag |= VI_DOINGINACT;
2372 		vp->v_iflag &= ~VI_OWEINACT;
2373 		VI_UNLOCK(vp);
2374 		(void) VOP_INACTIVE(vp, td);
2375 		VI_LOCK(vp);
2376 		VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2377 			 ("process_deferred_inactive: lost VI_DOINGINACT"));
2378 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2379 			 ("process_deferred_inactive: got VI_OWEINACT"));
2380 		vp->v_iflag &= ~VI_DOINGINACT;
2381 		VI_UNLOCK(vp);
2382 		VOP_UNLOCK(vp, 0, td);
2383 		vdrop(vp);
2384 		MNT_ILOCK(mp);
2385 	}
2386 	MNT_IUNLOCK(mp);
2387 	vn_finished_secondary_write(mp);
2388 }
2389 
2390 /* Try to free snapdata associated with devvp */
2391 static void
2392 try_free_snapdata(struct vnode *devvp,
2393 		  struct thread *td)
2394 {
2395 	struct snapdata *sn;
2396 	ufs2_daddr_t *snapblklist;
2397 
2398 	sn = devvp->v_rdev->si_snapdata;
2399 
2400 	if (sn == NULL || TAILQ_FIRST(&sn->sn_head) != NULL ||
2401 	    (devvp->v_vflag & VV_COPYONWRITE) == 0) {
2402 		VI_UNLOCK(devvp);
2403 		return;
2404 	}
2405 
2406 	devvp->v_rdev->si_snapdata = NULL;
2407 	devvp->v_vflag &= ~VV_COPYONWRITE;
2408 	snapblklist = sn->sn_blklist;
2409 	sn->sn_blklist = NULL;
2410 	sn->sn_listsize = 0;
2411 	lockmgr(&sn->sn_lock, LK_DRAIN|LK_INTERLOCK, VI_MTX(devvp), td);
2412 	lockmgr(&sn->sn_lock, LK_RELEASE, NULL, td);
2413 	lockdestroy(&sn->sn_lock);
2414 	free(sn, M_UFSMNT);
2415 	if (snapblklist != NULL)
2416 		FREE(snapblklist, M_UFSMNT);
2417 }
2418 #endif
2419