xref: /freebsd/sys/ufs/ffs/ffs_snapshot.c (revision 94942af266ac119ede0ca836f9aa5a5ac0582938)
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 static int ffs_bp_snapblk(struct vnode *, struct buf *);
167 
168 /*
169  * To ensure the consistency of snapshots across crashes, we must
170  * synchronously write out copied blocks before allowing the
171  * originals to be modified. Because of the rather severe speed
172  * penalty that this imposes, the following flag allows this
173  * crash persistence to be disabled.
174  */
175 int dopersistence = 0;
176 
177 #ifdef DEBUG
178 #include <sys/sysctl.h>
179 SYSCTL_INT(_debug, OID_AUTO, dopersistence, CTLFLAG_RW, &dopersistence, 0, "");
180 static int snapdebug = 0;
181 SYSCTL_INT(_debug, OID_AUTO, snapdebug, CTLFLAG_RW, &snapdebug, 0, "");
182 int collectsnapstats = 0;
183 SYSCTL_INT(_debug, OID_AUTO, collectsnapstats, CTLFLAG_RW, &collectsnapstats,
184 	0, "");
185 #endif /* DEBUG */
186 
187 /*
188  * Create a snapshot file and initialize it for the filesystem.
189  */
190 int
191 ffs_snapshot(mp, snapfile)
192 	struct mount *mp;
193 	char *snapfile;
194 {
195 	ufs2_daddr_t numblks, blkno, *blkp, *snapblklist;
196 	int error, cg, snaploc;
197 	int i, size, len, loc;
198 	int flag;
199 	struct timespec starttime = {0, 0}, endtime;
200 	char saved_nice = 0;
201 	long redo = 0, snaplistsize = 0;
202 	int32_t *lp;
203 	void *space;
204 	struct fs *copy_fs = NULL, *fs;
205 	struct thread *td = curthread;
206 	struct inode *ip, *xp;
207 	struct buf *bp, *nbp, *ibp, *sbp = NULL;
208 	struct nameidata nd;
209 	struct mount *wrtmp;
210 	struct vattr vat;
211 	struct vnode *vp, *xvp, *mvp, *devvp;
212 	struct uio auio;
213 	struct iovec aiov;
214 	struct snapdata *sn;
215 	struct ufsmount *ump;
216 
217 	ump = VFSTOUFS(mp);
218 	fs = ump->um_fs;
219 	sn = NULL;
220 	MNT_ILOCK(mp);
221 	flag = mp->mnt_flag;
222 	MNT_IUNLOCK(mp);
223 
224 	/*
225 	 * Need to serialize access to snapshot code per filesystem.
226 	 */
227 	/*
228 	 * Assign a snapshot slot in the superblock.
229 	 */
230 	UFS_LOCK(ump);
231 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++)
232 		if (fs->fs_snapinum[snaploc] == 0)
233 			break;
234 	UFS_UNLOCK(ump);
235 	if (snaploc == FSMAXSNAP)
236 		return (ENOSPC);
237 	/*
238 	 * Create the snapshot file.
239 	 */
240 restart:
241 	NDINIT(&nd, CREATE, LOCKPARENT | LOCKLEAF, UIO_SYSSPACE, snapfile, td);
242 	if ((error = namei(&nd)) != 0)
243 		return (error);
244 	if (nd.ni_vp != NULL) {
245 		vput(nd.ni_vp);
246 		error = EEXIST;
247 	}
248 	if (nd.ni_dvp->v_mount != mp)
249 		error = EXDEV;
250 	if (error) {
251 		NDFREE(&nd, NDF_ONLY_PNBUF);
252 		if (nd.ni_dvp == nd.ni_vp)
253 			vrele(nd.ni_dvp);
254 		else
255 			vput(nd.ni_dvp);
256 		return (error);
257 	}
258 	VATTR_NULL(&vat);
259 	vat.va_type = VREG;
260 	vat.va_mode = S_IRUSR;
261 	vat.va_vaflags |= VA_EXCLUSIVE;
262 	if (VOP_GETWRITEMOUNT(nd.ni_dvp, &wrtmp))
263 		wrtmp = NULL;
264 	if (wrtmp != mp)
265 		panic("ffs_snapshot: mount mismatch");
266 	vfs_rel(wrtmp);
267 	if (vn_start_write(NULL, &wrtmp, V_NOWAIT) != 0) {
268 		NDFREE(&nd, NDF_ONLY_PNBUF);
269 		vput(nd.ni_dvp);
270 		if ((error = vn_start_write(NULL, &wrtmp,
271 		    V_XSLEEP | PCATCH)) != 0)
272 			return (error);
273 		goto restart;
274 	}
275 	VOP_LEASE(nd.ni_dvp, td, KERNCRED, LEASE_WRITE);
276 	error = VOP_CREATE(nd.ni_dvp, &nd.ni_vp, &nd.ni_cnd, &vat);
277 	VOP_UNLOCK(nd.ni_dvp, 0, td);
278 	if (error) {
279 		NDFREE(&nd, NDF_ONLY_PNBUF);
280 		vn_finished_write(wrtmp);
281 		vrele(nd.ni_dvp);
282 		return (error);
283 	}
284 	vp = nd.ni_vp;
285 	vp->v_vflag |= VV_SYSTEM;
286 	ip = VTOI(vp);
287 	devvp = ip->i_devvp;
288 	/*
289 	 * Allocate and copy the last block contents so as to be able
290 	 * to set size to that of the filesystem.
291 	 */
292 	numblks = howmany(fs->fs_size, fs->fs_frag);
293 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)),
294 	    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
295 	if (error)
296 		goto out;
297 	ip->i_size = lblktosize(fs, (off_t)numblks);
298 	DIP_SET(ip, i_size, ip->i_size);
299 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
300 	error = readblock(vp, bp, numblks - 1);
301 	bawrite(bp);
302 	if (error != 0)
303 		goto out;
304 	/*
305 	 * Preallocate critical data structures so that we can copy
306 	 * them in without further allocation after we suspend all
307 	 * operations on the filesystem. We would like to just release
308 	 * the allocated buffers without writing them since they will
309 	 * be filled in below once we are ready to go, but this upsets
310 	 * the soft update code, so we go ahead and write the new buffers.
311 	 *
312 	 * Allocate all indirect blocks and mark all of them as not
313 	 * needing to be copied.
314 	 */
315 	for (blkno = NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
316 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
317 		    fs->fs_bsize, td->td_ucred, BA_METAONLY, &ibp);
318 		if (error)
319 			goto out;
320 		bawrite(ibp);
321 	}
322 	/*
323 	 * Allocate copies for the superblock and its summary information.
324 	 */
325 	error = UFS_BALLOC(vp, fs->fs_sblockloc, fs->fs_sbsize, KERNCRED,
326 	    0, &nbp);
327 	if (error)
328 		goto out;
329 	bawrite(nbp);
330 	blkno = fragstoblks(fs, fs->fs_csaddr);
331 	len = howmany(fs->fs_cssize, fs->fs_bsize);
332 	for (loc = 0; loc < len; loc++) {
333 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(blkno + loc)),
334 		    fs->fs_bsize, KERNCRED, 0, &nbp);
335 		if (error)
336 			goto out;
337 		bawrite(nbp);
338 	}
339 	/*
340 	 * Allocate all cylinder group blocks.
341 	 */
342 	for (cg = 0; cg < fs->fs_ncg; cg++) {
343 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
344 		    fs->fs_bsize, KERNCRED, 0, &nbp);
345 		if (error)
346 			goto out;
347 		bawrite(nbp);
348 	}
349 	/*
350 	 * Copy all the cylinder group maps. Although the
351 	 * filesystem is still active, we hope that only a few
352 	 * cylinder groups will change between now and when we
353 	 * suspend operations. Thus, we will be able to quickly
354 	 * touch up the few cylinder groups that changed during
355 	 * the suspension period.
356 	 */
357 	len = howmany(fs->fs_ncg, NBBY);
358 	MALLOC(space, void *, len, M_DEVBUF, M_WAITOK|M_ZERO);
359 	UFS_LOCK(ump);
360 	fs->fs_active = space;
361 	UFS_UNLOCK(ump);
362 	for (cg = 0; cg < fs->fs_ncg; cg++) {
363 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
364 		    fs->fs_bsize, KERNCRED, 0, &nbp);
365 		if (error)
366 			goto out;
367 		error = cgaccount(cg, vp, nbp, 1);
368 		bawrite(nbp);
369 		if (error)
370 			goto out;
371 	}
372 	/*
373 	 * Change inode to snapshot type file.
374 	 */
375 	ip->i_flags |= SF_SNAPSHOT;
376 	DIP_SET(ip, i_flags, ip->i_flags);
377 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
378 	/*
379 	 * Ensure that the snapshot is completely on disk.
380 	 * Since we have marked it as a snapshot it is safe to
381 	 * unlock it as no process will be allowed to write to it.
382 	 */
383 	if ((error = ffs_syncvnode(vp, MNT_WAIT)) != 0)
384 		goto out;
385 	VOP_UNLOCK(vp, 0, td);
386 	/*
387 	 * All allocations are done, so we can now snapshot the system.
388 	 *
389 	 * Recind nice scheduling while running with the filesystem suspended.
390 	 */
391 	if (td->td_proc->p_nice > 0) {
392 		PROC_LOCK(td->td_proc);
393 		mtx_lock_spin(&sched_lock);
394 		saved_nice = td->td_proc->p_nice;
395 		sched_nice(td->td_proc, 0);
396 		mtx_unlock_spin(&sched_lock);
397 		PROC_UNLOCK(td->td_proc);
398 	}
399 	/*
400 	 * Suspend operation on filesystem.
401 	 */
402 	for (;;) {
403 		vn_finished_write(wrtmp);
404 		if ((error = vfs_write_suspend(vp->v_mount)) != 0) {
405 			vn_start_write(NULL, &wrtmp, V_WAIT);
406 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
407 			goto out;
408 		}
409 		if (mp->mnt_kern_flag & MNTK_SUSPENDED)
410 			break;
411 		vn_start_write(NULL, &wrtmp, V_WAIT);
412 	}
413 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
414 	if (ip->i_effnlink == 0) {
415 		error = ENOENT;		/* Snapshot file unlinked */
416 		goto out1;
417 	}
418 	if (collectsnapstats)
419 		nanotime(&starttime);
420 
421 	/* The last block might have changed.  Copy it again to be sure. */
422 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)),
423 	    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
424 	if (error != 0)
425 		goto out1;
426 	error = readblock(vp, bp, numblks - 1);
427 	bp->b_flags |= B_VALIDSUSPWRT;
428 	bawrite(bp);
429 	if (error != 0)
430 		goto out1;
431 	/*
432 	 * First, copy all the cylinder group maps that have changed.
433 	 */
434 	for (cg = 0; cg < fs->fs_ncg; cg++) {
435 		if ((ACTIVECGNUM(fs, cg) & ACTIVECGOFF(cg)) != 0)
436 			continue;
437 		redo++;
438 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
439 		    fs->fs_bsize, KERNCRED, 0, &nbp);
440 		if (error)
441 			goto out1;
442 		error = cgaccount(cg, vp, nbp, 2);
443 		bawrite(nbp);
444 		if (error)
445 			goto out1;
446 	}
447 	/*
448 	 * Grab a copy of the superblock and its summary information.
449 	 * We delay writing it until the suspension is released below.
450 	 */
451 	error = bread(vp, lblkno(fs, fs->fs_sblockloc), fs->fs_bsize,
452 	    KERNCRED, &sbp);
453 	if (error) {
454 		brelse(sbp);
455 		sbp = NULL;
456 		goto out1;
457 	}
458 	loc = blkoff(fs, fs->fs_sblockloc);
459 	copy_fs = (struct fs *)(sbp->b_data + loc);
460 	bcopy(fs, copy_fs, fs->fs_sbsize);
461 	if ((fs->fs_flags & (FS_UNCLEAN | FS_NEEDSFSCK)) == 0)
462 		copy_fs->fs_clean = 1;
463 	size = fs->fs_bsize < SBLOCKSIZE ? fs->fs_bsize : SBLOCKSIZE;
464 	if (fs->fs_sbsize < size)
465 		bzero(&sbp->b_data[loc + fs->fs_sbsize], size - fs->fs_sbsize);
466 	size = blkroundup(fs, fs->fs_cssize);
467 	if (fs->fs_contigsumsize > 0)
468 		size += fs->fs_ncg * sizeof(int32_t);
469 	space = malloc((u_long)size, M_UFSMNT, M_WAITOK);
470 	copy_fs->fs_csp = space;
471 	bcopy(fs->fs_csp, copy_fs->fs_csp, fs->fs_cssize);
472 	space = (char *)space + fs->fs_cssize;
473 	loc = howmany(fs->fs_cssize, fs->fs_fsize);
474 	i = fs->fs_frag - loc % fs->fs_frag;
475 	len = (i == fs->fs_frag) ? 0 : i * fs->fs_fsize;
476 	if (len > 0) {
477 		if ((error = bread(devvp, fsbtodb(fs, fs->fs_csaddr + loc),
478 		    len, KERNCRED, &bp)) != 0) {
479 			brelse(bp);
480 			free(copy_fs->fs_csp, M_UFSMNT);
481 			bawrite(sbp);
482 			sbp = NULL;
483 			goto out1;
484 		}
485 		bcopy(bp->b_data, space, (u_int)len);
486 		space = (char *)space + len;
487 		bp->b_flags |= B_INVAL | B_NOCACHE;
488 		brelse(bp);
489 	}
490 	if (fs->fs_contigsumsize > 0) {
491 		copy_fs->fs_maxcluster = lp = space;
492 		for (i = 0; i < fs->fs_ncg; i++)
493 			*lp++ = fs->fs_contigsumsize;
494 	}
495 	/*
496 	 * We must check for active files that have been unlinked
497 	 * (e.g., with a zero link count). We have to expunge all
498 	 * trace of these files from the snapshot so that they are
499 	 * not reclaimed prematurely by fsck or unnecessarily dumped.
500 	 * We turn off the MNTK_SUSPENDED flag to avoid a panic from
501 	 * spec_strategy about writing on a suspended filesystem.
502 	 * Note that we skip unlinked snapshot files as they will
503 	 * be handled separately below.
504 	 *
505 	 * We also calculate the needed size for the snapshot list.
506 	 */
507 	snaplistsize = fs->fs_ncg + howmany(fs->fs_cssize, fs->fs_bsize) +
508 	    FSMAXSNAP + 1 /* superblock */ + 1 /* last block */ + 1 /* size */;
509 	MNT_ILOCK(mp);
510 	mp->mnt_kern_flag &= ~MNTK_SUSPENDED;
511 loop:
512 	MNT_VNODE_FOREACH(xvp, mp, mvp) {
513 		VI_LOCK(xvp);
514 		MNT_IUNLOCK(mp);
515 		if ((xvp->v_iflag & VI_DOOMED) ||
516 		    (xvp->v_usecount == 0 &&
517 		     (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) ||
518 		    xvp->v_type == VNON ||
519 		    (VTOI(xvp)->i_flags & SF_SNAPSHOT)) {
520 			VI_UNLOCK(xvp);
521 			MNT_ILOCK(mp);
522 			continue;
523 		}
524 		/*
525 		 * We can skip parent directory vnode because it must have
526 		 * this snapshot file in it.
527 		 */
528 		if (xvp == nd.ni_dvp) {
529 			VI_UNLOCK(xvp);
530 			MNT_ILOCK(mp);
531 			continue;
532 		}
533 		vholdl(xvp);
534 		if (vn_lock(xvp, LK_EXCLUSIVE | LK_INTERLOCK, td) != 0) {
535 			MNT_ILOCK(mp);
536 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
537 			vdrop(xvp);
538 			goto loop;
539 		}
540 		VI_LOCK(xvp);
541 		if (xvp->v_usecount == 0 &&
542 		    (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) {
543 			VI_UNLOCK(xvp);
544 			VOP_UNLOCK(xvp, 0, td);
545 			vdrop(xvp);
546 			MNT_ILOCK(mp);
547 			continue;
548 		}
549 		VI_UNLOCK(xvp);
550 		if (snapdebug)
551 			vprint("ffs_snapshot: busy vnode", xvp);
552 		if (VOP_GETATTR(xvp, &vat, td->td_ucred, td) == 0 &&
553 		    vat.va_nlink > 0) {
554 			VOP_UNLOCK(xvp, 0, td);
555 			vdrop(xvp);
556 			MNT_ILOCK(mp);
557 			continue;
558 		}
559 		xp = VTOI(xvp);
560 		if (ffs_checkfreefile(copy_fs, vp, xp->i_number)) {
561 			VOP_UNLOCK(xvp, 0, td);
562 			vdrop(xvp);
563 			MNT_ILOCK(mp);
564 			continue;
565 		}
566 		/*
567 		 * If there is a fragment, clear it here.
568 		 */
569 		blkno = 0;
570 		loc = howmany(xp->i_size, fs->fs_bsize) - 1;
571 		if (loc < NDADDR) {
572 			len = fragroundup(fs, blkoff(fs, xp->i_size));
573 			if (len != 0 && len < fs->fs_bsize) {
574 				ffs_blkfree(ump, copy_fs, vp,
575 				    DIP(xp, i_db[loc]), len, xp->i_number);
576 				blkno = DIP(xp, i_db[loc]);
577 				DIP_SET(xp, i_db[loc], 0);
578 			}
579 		}
580 		snaplistsize += 1;
581 		if (xp->i_ump->um_fstype == UFS1)
582 			error = expunge_ufs1(vp, xp, copy_fs, fullacct_ufs1,
583 			    BLK_NOCOPY);
584 		else
585 			error = expunge_ufs2(vp, xp, copy_fs, fullacct_ufs2,
586 			    BLK_NOCOPY);
587 		if (blkno)
588 			DIP_SET(xp, i_db[loc], blkno);
589 		if (!error)
590 			error = ffs_freefile(ump, copy_fs, vp, xp->i_number,
591 			    xp->i_mode);
592 		VOP_UNLOCK(xvp, 0, td);
593 		vdrop(xvp);
594 		if (error) {
595 			free(copy_fs->fs_csp, M_UFSMNT);
596 			bawrite(sbp);
597 			sbp = NULL;
598 			MNT_VNODE_FOREACH_ABORT(mp, mvp);
599 			goto out1;
600 		}
601 		MNT_ILOCK(mp);
602 	}
603 	MNT_IUNLOCK(mp);
604 	/*
605 	 * If there already exist snapshots on this filesystem, grab a
606 	 * reference to their shared lock. If this is the first snapshot
607 	 * on this filesystem, we need to allocate a lock for the snapshots
608 	 * to share. In either case, acquire the snapshot lock and give
609 	 * up our original private lock.
610 	 */
611 	VI_LOCK(devvp);
612 	sn = devvp->v_rdev->si_snapdata;
613 	if (sn != NULL) {
614 		xp = TAILQ_FIRST(&sn->sn_head);
615 		VI_UNLOCK(devvp);
616 		VI_LOCK(vp);
617 		vp->v_vnlock = &sn->sn_lock;
618 	} else {
619 		VI_UNLOCK(devvp);
620 		sn = malloc(sizeof *sn, M_UFSMNT, M_WAITOK | M_ZERO);
621 		TAILQ_INIT(&sn->sn_head);
622 		lockinit(&sn->sn_lock, PVFS, "snaplk", VLKTIMEOUT,
623 		    LK_CANRECURSE | LK_NOSHARE);
624 		VI_LOCK(vp);
625 		vp->v_vnlock = &sn->sn_lock;
626 		mp_fixme("si_snapdata setting is racey.");
627 		devvp->v_rdev->si_snapdata = sn;
628 		xp = NULL;
629 	}
630 	lockmgr(vp->v_vnlock, LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY,
631 	    VI_MTX(vp), td);
632 	lockmgr(&vp->v_lock, LK_RELEASE, NULL, td);
633 	/*
634 	 * If this is the first snapshot on this filesystem, then we need
635 	 * to allocate the space for the list of preallocated snapshot blocks.
636 	 * This list will be refined below, but this preliminary one will
637 	 * keep us out of deadlock until the full one is ready.
638 	 */
639 	if (xp == NULL) {
640 		MALLOC(snapblklist, daddr_t *, snaplistsize * sizeof(daddr_t),
641 		    M_UFSMNT, M_WAITOK);
642 		blkp = &snapblklist[1];
643 		*blkp++ = lblkno(fs, fs->fs_sblockloc);
644 		blkno = fragstoblks(fs, fs->fs_csaddr);
645 		for (cg = 0; cg < fs->fs_ncg; cg++) {
646 			if (fragstoblks(fs, cgtod(fs, cg) > blkno))
647 				break;
648 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
649 		}
650 		len = howmany(fs->fs_cssize, fs->fs_bsize);
651 		for (loc = 0; loc < len; loc++)
652 			*blkp++ = blkno + loc;
653 		for (; cg < fs->fs_ncg; cg++)
654 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
655 		snapblklist[0] = blkp - snapblklist;
656 		VI_LOCK(devvp);
657 		if (sn->sn_blklist != NULL)
658 			panic("ffs_snapshot: non-empty list");
659 		sn->sn_blklist = snapblklist;
660 		sn->sn_listsize = blkp - snapblklist;
661 		VI_UNLOCK(devvp);
662 	}
663 	/*
664 	 * Record snapshot inode. Since this is the newest snapshot,
665 	 * it must be placed at the end of the list.
666 	 */
667 	VI_LOCK(devvp);
668 	fs->fs_snapinum[snaploc] = ip->i_number;
669 	if (ip->i_nextsnap.tqe_prev != 0)
670 		panic("ffs_snapshot: %d already on list", ip->i_number);
671 	TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
672 	devvp->v_vflag |= VV_COPYONWRITE;
673 	VI_UNLOCK(devvp);
674 	ASSERT_VOP_LOCKED(vp, "ffs_snapshot vp");
675 out1:
676 	KASSERT((sn != NULL && sbp != NULL && error == 0) ||
677 		(sn == NULL && sbp == NULL && error != 0),
678 		("email phk@ and mckusick@"));
679 	/*
680 	 * Resume operation on filesystem.
681 	 */
682 	vfs_write_resume(vp->v_mount);
683 	vn_start_write(NULL, &wrtmp, V_WAIT);
684 	if (collectsnapstats && starttime.tv_sec > 0) {
685 		nanotime(&endtime);
686 		timespecsub(&endtime, &starttime);
687 		printf("%s: suspended %ld.%03ld sec, redo %ld of %d\n",
688 		    vp->v_mount->mnt_stat.f_mntonname, (long)endtime.tv_sec,
689 		    endtime.tv_nsec / 1000000, redo, fs->fs_ncg);
690 	}
691 	if (sbp == NULL)
692 		goto out;
693 	/*
694 	 * Copy allocation information from all the snapshots in
695 	 * this snapshot and then expunge them from its view.
696 	 */
697 	TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap) {
698 		if (xp == ip)
699 			break;
700 		if (xp->i_ump->um_fstype == UFS1)
701 			error = expunge_ufs1(vp, xp, fs, snapacct_ufs1,
702 			    BLK_SNAP);
703 		else
704 			error = expunge_ufs2(vp, xp, fs, snapacct_ufs2,
705 			    BLK_SNAP);
706 		if (error == 0 && xp->i_effnlink == 0) {
707 			error = ffs_freefile(ump,
708 					     copy_fs,
709 					     vp,
710 					     xp->i_number,
711 					     xp->i_mode);
712 		}
713 		if (error) {
714 			fs->fs_snapinum[snaploc] = 0;
715 			goto done;
716 		}
717 	}
718 	/*
719 	 * Allocate space for the full list of preallocated snapshot blocks.
720 	 */
721 	MALLOC(snapblklist, daddr_t *, snaplistsize * sizeof(daddr_t),
722 	    M_UFSMNT, M_WAITOK);
723 	ip->i_snapblklist = &snapblklist[1];
724 	/*
725 	 * Expunge the blocks used by the snapshots from the set of
726 	 * blocks marked as used in the snapshot bitmaps. Also, collect
727 	 * the list of allocated blocks in i_snapblklist.
728 	 */
729 	if (ip->i_ump->um_fstype == UFS1)
730 		error = expunge_ufs1(vp, ip, copy_fs, mapacct_ufs1, BLK_SNAP);
731 	else
732 		error = expunge_ufs2(vp, ip, copy_fs, mapacct_ufs2, BLK_SNAP);
733 	if (error) {
734 		fs->fs_snapinum[snaploc] = 0;
735 		FREE(snapblklist, M_UFSMNT);
736 		goto done;
737 	}
738 	if (snaplistsize < ip->i_snapblklist - snapblklist)
739 		panic("ffs_snapshot: list too small");
740 	snaplistsize = ip->i_snapblklist - snapblklist;
741 	snapblklist[0] = snaplistsize;
742 	ip->i_snapblklist = 0;
743 	/*
744 	 * Write out the list of allocated blocks to the end of the snapshot.
745 	 */
746 	auio.uio_iov = &aiov;
747 	auio.uio_iovcnt = 1;
748 	aiov.iov_base = (void *)snapblklist;
749 	aiov.iov_len = snaplistsize * sizeof(daddr_t);
750 	auio.uio_resid = aiov.iov_len;;
751 	auio.uio_offset = ip->i_size;
752 	auio.uio_segflg = UIO_SYSSPACE;
753 	auio.uio_rw = UIO_WRITE;
754 	auio.uio_td = td;
755 	if ((error = VOP_WRITE(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
756 		fs->fs_snapinum[snaploc] = 0;
757 		FREE(snapblklist, M_UFSMNT);
758 		goto done;
759 	}
760 	/*
761 	 * Write the superblock and its summary information
762 	 * to the snapshot.
763 	 */
764 	blkno = fragstoblks(fs, fs->fs_csaddr);
765 	len = howmany(fs->fs_cssize, fs->fs_bsize);
766 	space = copy_fs->fs_csp;
767 	for (loc = 0; loc < len; loc++) {
768 		error = bread(vp, blkno + loc, fs->fs_bsize, KERNCRED, &nbp);
769 		if (error) {
770 			brelse(nbp);
771 			fs->fs_snapinum[snaploc] = 0;
772 			FREE(snapblklist, M_UFSMNT);
773 			goto done;
774 		}
775 		bcopy(space, nbp->b_data, fs->fs_bsize);
776 		space = (char *)space + fs->fs_bsize;
777 		bawrite(nbp);
778 	}
779 	/*
780 	 * As this is the newest list, it is the most inclusive, so
781 	 * should replace the previous list.
782 	 */
783 	VI_LOCK(devvp);
784 	space = sn->sn_blklist;
785 	sn->sn_blklist = snapblklist;
786 	sn->sn_listsize = snaplistsize;
787 	VI_UNLOCK(devvp);
788 	if (space != NULL)
789 		FREE(space, M_UFSMNT);
790 	/*
791 	 * If another process is currently writing the buffer containing
792 	 * the inode for this snapshot then a deadlock can occur. Drop
793 	 * the snapshot lock until the buffer has been written.
794 	 */
795 	VREF(vp);	/* Protect against ffs_snapgone() */
796 	VOP_UNLOCK(vp, 0, td);
797 	(void) bread(ip->i_devvp,
798 		     fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
799 		     (int) fs->fs_bsize, NOCRED, &nbp);
800 	brelse(nbp);
801 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
802 	if (ip->i_effnlink == 0)
803 		error = ENOENT;		/* Snapshot file unlinked */
804 	else
805 		vrele(vp);		/* Drop extra reference */
806 done:
807 	FREE(copy_fs->fs_csp, M_UFSMNT);
808 	bawrite(sbp);
809 out:
810 	NDFREE(&nd, NDF_ONLY_PNBUF);
811 	if (saved_nice > 0) {
812 		PROC_LOCK(td->td_proc);
813 		mtx_lock_spin(&sched_lock);
814 		sched_nice(td->td_proc, saved_nice);
815 		mtx_unlock_spin(&sched_lock);
816 		PROC_UNLOCK(td->td_proc);
817 	}
818 	UFS_LOCK(ump);
819 	if (fs->fs_active != 0) {
820 		FREE(fs->fs_active, M_DEVBUF);
821 		fs->fs_active = 0;
822 	}
823 	UFS_UNLOCK(ump);
824 	MNT_ILOCK(mp);
825 	mp->mnt_flag = (mp->mnt_flag & MNT_QUOTA) | (flag & ~MNT_QUOTA);
826 	MNT_IUNLOCK(mp);
827 	if (error)
828 		(void) ffs_truncate(vp, (off_t)0, 0, NOCRED, td);
829 	(void) ffs_syncvnode(vp, MNT_WAIT);
830 	if (error)
831 		vput(vp);
832 	else
833 		VOP_UNLOCK(vp, 0, td);
834 	vrele(nd.ni_dvp);
835 	vn_finished_write(wrtmp);
836 	process_deferred_inactive(mp);
837 	return (error);
838 }
839 
840 /*
841  * Copy a cylinder group map. All the unallocated blocks are marked
842  * BLK_NOCOPY so that the snapshot knows that it need not copy them
843  * if they are later written. If passno is one, then this is a first
844  * pass, so only setting needs to be done. If passno is 2, then this
845  * is a revision to a previous pass which must be undone as the
846  * replacement pass is done.
847  */
848 static int
849 cgaccount(cg, vp, nbp, passno)
850 	int cg;
851 	struct vnode *vp;
852 	struct buf *nbp;
853 	int passno;
854 {
855 	struct buf *bp, *ibp;
856 	struct inode *ip;
857 	struct cg *cgp;
858 	struct fs *fs;
859 	ufs2_daddr_t base, numblks;
860 	int error, len, loc, indiroff;
861 
862 	ip = VTOI(vp);
863 	fs = ip->i_fs;
864 	error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)),
865 		(int)fs->fs_cgsize, KERNCRED, &bp);
866 	if (error) {
867 		brelse(bp);
868 		return (error);
869 	}
870 	cgp = (struct cg *)bp->b_data;
871 	if (!cg_chkmagic(cgp)) {
872 		brelse(bp);
873 		return (EIO);
874 	}
875 	UFS_LOCK(ip->i_ump);
876 	ACTIVESET(fs, cg);
877 	UFS_UNLOCK(ip->i_ump);
878 	bcopy(bp->b_data, nbp->b_data, fs->fs_cgsize);
879 	if (fs->fs_cgsize < fs->fs_bsize)
880 		bzero(&nbp->b_data[fs->fs_cgsize],
881 		    fs->fs_bsize - fs->fs_cgsize);
882 	cgp = (struct cg *)nbp->b_data;
883 	bqrelse(bp);
884 	if (passno == 2)
885 		nbp->b_flags |= B_VALIDSUSPWRT;
886 	numblks = howmany(fs->fs_size, fs->fs_frag);
887 	len = howmany(fs->fs_fpg, fs->fs_frag);
888 	base = cgbase(fs, cg) / fs->fs_frag;
889 	if (base + len >= numblks)
890 		len = numblks - base - 1;
891 	loc = 0;
892 	if (base < NDADDR) {
893 		for ( ; loc < NDADDR; loc++) {
894 			if (ffs_isblock(fs, cg_blksfree(cgp), loc))
895 				DIP_SET(ip, i_db[loc], BLK_NOCOPY);
896 			else if (passno == 2 && DIP(ip, i_db[loc])== BLK_NOCOPY)
897 				DIP_SET(ip, i_db[loc], 0);
898 			else if (passno == 1 && DIP(ip, i_db[loc])== BLK_NOCOPY)
899 				panic("ffs_snapshot: lost direct block");
900 		}
901 	}
902 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(base + loc)),
903 	    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
904 	if (error) {
905 		return (error);
906 	}
907 	indiroff = (base + loc - NDADDR) % NINDIR(fs);
908 	for ( ; loc < len; loc++, indiroff++) {
909 		if (indiroff >= NINDIR(fs)) {
910 			if (passno == 2)
911 				ibp->b_flags |= B_VALIDSUSPWRT;
912 			bawrite(ibp);
913 			error = UFS_BALLOC(vp,
914 			    lblktosize(fs, (off_t)(base + loc)),
915 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
916 			if (error) {
917 				return (error);
918 			}
919 			indiroff = 0;
920 		}
921 		if (ip->i_ump->um_fstype == UFS1) {
922 			if (ffs_isblock(fs, cg_blksfree(cgp), loc))
923 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] =
924 				    BLK_NOCOPY;
925 			else if (passno == 2 && ((ufs1_daddr_t *)(ibp->b_data))
926 			    [indiroff] == BLK_NOCOPY)
927 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] = 0;
928 			else if (passno == 1 && ((ufs1_daddr_t *)(ibp->b_data))
929 			    [indiroff] == BLK_NOCOPY)
930 				panic("ffs_snapshot: lost indirect block");
931 			continue;
932 		}
933 		if (ffs_isblock(fs, cg_blksfree(cgp), loc))
934 			((ufs2_daddr_t *)(ibp->b_data))[indiroff] = BLK_NOCOPY;
935 		else if (passno == 2 &&
936 		    ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY)
937 			((ufs2_daddr_t *)(ibp->b_data))[indiroff] = 0;
938 		else if (passno == 1 &&
939 		    ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY)
940 			panic("ffs_snapshot: lost indirect block");
941 	}
942 	if (passno == 2)
943 		ibp->b_flags |= B_VALIDSUSPWRT;
944 	bdwrite(ibp);
945 	return (0);
946 }
947 
948 /*
949  * Before expunging a snapshot inode, note all the
950  * blocks that it claims with BLK_SNAP so that fsck will
951  * be able to account for those blocks properly and so
952  * that this snapshot knows that it need not copy them
953  * if the other snapshot holding them is freed. This code
954  * is reproduced once each for UFS1 and UFS2.
955  */
956 static int
957 expunge_ufs1(snapvp, cancelip, fs, acctfunc, expungetype)
958 	struct vnode *snapvp;
959 	struct inode *cancelip;
960 	struct fs *fs;
961 	int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
962 	    struct fs *, ufs_lbn_t, int);
963 	int expungetype;
964 {
965 	int i, error, indiroff;
966 	ufs_lbn_t lbn, rlbn;
967 	ufs2_daddr_t len, blkno, numblks, blksperindir;
968 	struct ufs1_dinode *dip;
969 	struct thread *td = curthread;
970 	struct buf *bp;
971 
972 	/*
973 	 * Prepare to expunge the inode. If its inode block has not
974 	 * yet been copied, then allocate and fill the copy.
975 	 */
976 	lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number));
977 	blkno = 0;
978 	if (lbn < NDADDR) {
979 		blkno = VTOI(snapvp)->i_din1->di_db[lbn];
980 	} else {
981 		td->td_pflags |= TDP_COWINPROGRESS;
982 		error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn),
983 		   fs->fs_bsize, KERNCRED, BA_METAONLY, &bp);
984 		td->td_pflags &= ~TDP_COWINPROGRESS;
985 		if (error)
986 			return (error);
987 		indiroff = (lbn - NDADDR) % NINDIR(fs);
988 		blkno = ((ufs1_daddr_t *)(bp->b_data))[indiroff];
989 		bqrelse(bp);
990 	}
991 	if (blkno != 0) {
992 		if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp)))
993 			return (error);
994 	} else {
995 		error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn),
996 		    fs->fs_bsize, KERNCRED, 0, &bp);
997 		if (error)
998 			return (error);
999 		if ((error = readblock(snapvp, bp, lbn)) != 0)
1000 			return (error);
1001 	}
1002 	/*
1003 	 * Set a snapshot inode to be a zero length file, regular files
1004 	 * or unlinked snapshots to be completely unallocated.
1005 	 */
1006 	dip = (struct ufs1_dinode *)bp->b_data +
1007 	    ino_to_fsbo(fs, cancelip->i_number);
1008 	if (expungetype == BLK_NOCOPY || cancelip->i_effnlink == 0)
1009 		dip->di_mode = 0;
1010 	dip->di_size = 0;
1011 	dip->di_blocks = 0;
1012 	dip->di_flags &= ~SF_SNAPSHOT;
1013 	bzero(&dip->di_db[0], (NDADDR + NIADDR) * sizeof(ufs1_daddr_t));
1014 	bdwrite(bp);
1015 	/*
1016 	 * Now go through and expunge all the blocks in the file
1017 	 * using the function requested.
1018 	 */
1019 	numblks = howmany(cancelip->i_size, fs->fs_bsize);
1020 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_db[0],
1021 	    &cancelip->i_din1->di_db[NDADDR], fs, 0, expungetype)))
1022 		return (error);
1023 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_ib[0],
1024 	    &cancelip->i_din1->di_ib[NIADDR], fs, -1, expungetype)))
1025 		return (error);
1026 	blksperindir = 1;
1027 	lbn = -NDADDR;
1028 	len = numblks - NDADDR;
1029 	rlbn = NDADDR;
1030 	for (i = 0; len > 0 && i < NIADDR; i++) {
1031 		error = indiracct_ufs1(snapvp, ITOV(cancelip), i,
1032 		    cancelip->i_din1->di_ib[i], lbn, rlbn, len,
1033 		    blksperindir, fs, acctfunc, expungetype);
1034 		if (error)
1035 			return (error);
1036 		blksperindir *= NINDIR(fs);
1037 		lbn -= blksperindir + 1;
1038 		len -= blksperindir;
1039 		rlbn += blksperindir;
1040 	}
1041 	return (0);
1042 }
1043 
1044 /*
1045  * Descend an indirect block chain for vnode cancelvp accounting for all
1046  * its indirect blocks in snapvp.
1047  */
1048 static int
1049 indiracct_ufs1(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks,
1050 	    blksperindir, fs, acctfunc, expungetype)
1051 	struct vnode *snapvp;
1052 	struct vnode *cancelvp;
1053 	int level;
1054 	ufs1_daddr_t blkno;
1055 	ufs_lbn_t lbn;
1056 	ufs_lbn_t rlbn;
1057 	ufs_lbn_t remblks;
1058 	ufs_lbn_t blksperindir;
1059 	struct fs *fs;
1060 	int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
1061 	    struct fs *, ufs_lbn_t, int);
1062 	int expungetype;
1063 {
1064 	int error, num, i;
1065 	ufs_lbn_t subblksperindir;
1066 	struct indir indirs[NIADDR + 2];
1067 	ufs1_daddr_t last, *bap;
1068 	struct buf *bp;
1069 
1070 	if (blkno == 0) {
1071 		if (expungetype == BLK_NOCOPY)
1072 			return (0);
1073 		panic("indiracct_ufs1: missing indir");
1074 	}
1075 	if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0)
1076 		return (error);
1077 	if (lbn != indirs[num - 1 - level].in_lbn || num < 2)
1078 		panic("indiracct_ufs1: botched params");
1079 	/*
1080 	 * We have to expand bread here since it will deadlock looking
1081 	 * up the block number for any blocks that are not in the cache.
1082 	 */
1083 	bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0);
1084 	bp->b_blkno = fsbtodb(fs, blkno);
1085 	if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0 &&
1086 	    (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) {
1087 		brelse(bp);
1088 		return (error);
1089 	}
1090 	/*
1091 	 * Account for the block pointers in this indirect block.
1092 	 */
1093 	last = howmany(remblks, blksperindir);
1094 	if (last > NINDIR(fs))
1095 		last = NINDIR(fs);
1096 	MALLOC(bap, ufs1_daddr_t *, fs->fs_bsize, M_DEVBUF, M_WAITOK);
1097 	bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize);
1098 	bqrelse(bp);
1099 	error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs,
1100 	    level == 0 ? rlbn : -1, expungetype);
1101 	if (error || level == 0)
1102 		goto out;
1103 	/*
1104 	 * Account for the block pointers in each of the indirect blocks
1105 	 * in the levels below us.
1106 	 */
1107 	subblksperindir = blksperindir / NINDIR(fs);
1108 	for (lbn++, level--, i = 0; i < last; i++) {
1109 		error = indiracct_ufs1(snapvp, cancelvp, level, bap[i], lbn,
1110 		    rlbn, remblks, subblksperindir, fs, acctfunc, expungetype);
1111 		if (error)
1112 			goto out;
1113 		rlbn += blksperindir;
1114 		lbn -= blksperindir;
1115 		remblks -= blksperindir;
1116 	}
1117 out:
1118 	FREE(bap, M_DEVBUF);
1119 	return (error);
1120 }
1121 
1122 /*
1123  * Do both snap accounting and map accounting.
1124  */
1125 static int
1126 fullacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype)
1127 	struct vnode *vp;
1128 	ufs1_daddr_t *oldblkp, *lastblkp;
1129 	struct fs *fs;
1130 	ufs_lbn_t lblkno;
1131 	int exptype;	/* BLK_SNAP or BLK_NOCOPY */
1132 {
1133 	int error;
1134 
1135 	if ((error = snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype)))
1136 		return (error);
1137 	return (mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype));
1138 }
1139 
1140 /*
1141  * Identify a set of blocks allocated in a snapshot inode.
1142  */
1143 static int
1144 snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1145 	struct vnode *vp;
1146 	ufs1_daddr_t *oldblkp, *lastblkp;
1147 	struct fs *fs;
1148 	ufs_lbn_t lblkno;
1149 	int expungetype;	/* BLK_SNAP or BLK_NOCOPY */
1150 {
1151 	struct inode *ip = VTOI(vp);
1152 	ufs1_daddr_t blkno, *blkp;
1153 	ufs_lbn_t lbn;
1154 	struct buf *ibp;
1155 	int error;
1156 
1157 	for ( ; oldblkp < lastblkp; oldblkp++) {
1158 		blkno = *oldblkp;
1159 		if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP)
1160 			continue;
1161 		lbn = fragstoblks(fs, blkno);
1162 		if (lbn < NDADDR) {
1163 			blkp = &ip->i_din1->di_db[lbn];
1164 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1165 		} else {
1166 			error = ffs_balloc_ufs1(vp, lblktosize(fs, (off_t)lbn),
1167 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1168 			if (error)
1169 				return (error);
1170 			blkp = &((ufs1_daddr_t *)(ibp->b_data))
1171 			    [(lbn - NDADDR) % NINDIR(fs)];
1172 		}
1173 		/*
1174 		 * If we are expunging a snapshot vnode and we
1175 		 * find a block marked BLK_NOCOPY, then it is
1176 		 * one that has been allocated to this snapshot after
1177 		 * we took our current snapshot and can be ignored.
1178 		 */
1179 		if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) {
1180 			if (lbn >= NDADDR)
1181 				brelse(ibp);
1182 		} else {
1183 			if (*blkp != 0)
1184 				panic("snapacct_ufs1: bad block");
1185 			*blkp = expungetype;
1186 			if (lbn >= NDADDR)
1187 				bdwrite(ibp);
1188 		}
1189 	}
1190 	return (0);
1191 }
1192 
1193 /*
1194  * Account for a set of blocks allocated in a snapshot inode.
1195  */
1196 static int
1197 mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1198 	struct vnode *vp;
1199 	ufs1_daddr_t *oldblkp, *lastblkp;
1200 	struct fs *fs;
1201 	ufs_lbn_t lblkno;
1202 	int expungetype;
1203 {
1204 	ufs1_daddr_t blkno;
1205 	struct inode *ip;
1206 	ino_t inum;
1207 	int acctit;
1208 
1209 	ip = VTOI(vp);
1210 	inum = ip->i_number;
1211 	if (lblkno == -1)
1212 		acctit = 0;
1213 	else
1214 		acctit = 1;
1215 	for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) {
1216 		blkno = *oldblkp;
1217 		if (blkno == 0 || blkno == BLK_NOCOPY)
1218 			continue;
1219 		if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP)
1220 			*ip->i_snapblklist++ = lblkno;
1221 		if (blkno == BLK_SNAP)
1222 			blkno = blkstofrags(fs, lblkno);
1223 		ffs_blkfree(ip->i_ump, fs, vp, blkno, fs->fs_bsize, inum);
1224 	}
1225 	return (0);
1226 }
1227 
1228 /*
1229  * Before expunging a snapshot inode, note all the
1230  * blocks that it claims with BLK_SNAP so that fsck will
1231  * be able to account for those blocks properly and so
1232  * that this snapshot knows that it need not copy them
1233  * if the other snapshot holding them is freed. This code
1234  * is reproduced once each for UFS1 and UFS2.
1235  */
1236 static int
1237 expunge_ufs2(snapvp, cancelip, fs, acctfunc, expungetype)
1238 	struct vnode *snapvp;
1239 	struct inode *cancelip;
1240 	struct fs *fs;
1241 	int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
1242 	    struct fs *, ufs_lbn_t, int);
1243 	int expungetype;
1244 {
1245 	int i, error, indiroff;
1246 	ufs_lbn_t lbn, rlbn;
1247 	ufs2_daddr_t len, blkno, numblks, blksperindir;
1248 	struct ufs2_dinode *dip;
1249 	struct thread *td = curthread;
1250 	struct buf *bp;
1251 
1252 	/*
1253 	 * Prepare to expunge the inode. If its inode block has not
1254 	 * yet been copied, then allocate and fill the copy.
1255 	 */
1256 	lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number));
1257 	blkno = 0;
1258 	if (lbn < NDADDR) {
1259 		blkno = VTOI(snapvp)->i_din2->di_db[lbn];
1260 	} else {
1261 		td->td_pflags |= TDP_COWINPROGRESS;
1262 		error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn),
1263 		   fs->fs_bsize, KERNCRED, BA_METAONLY, &bp);
1264 		td->td_pflags &= ~TDP_COWINPROGRESS;
1265 		if (error)
1266 			return (error);
1267 		indiroff = (lbn - NDADDR) % NINDIR(fs);
1268 		blkno = ((ufs2_daddr_t *)(bp->b_data))[indiroff];
1269 		bqrelse(bp);
1270 	}
1271 	if (blkno != 0) {
1272 		if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp)))
1273 			return (error);
1274 	} else {
1275 		error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn),
1276 		    fs->fs_bsize, KERNCRED, 0, &bp);
1277 		if (error)
1278 			return (error);
1279 		if ((error = readblock(snapvp, bp, lbn)) != 0)
1280 			return (error);
1281 	}
1282 	/*
1283 	 * Set a snapshot inode to be a zero length file, regular files
1284 	 * to be completely unallocated.
1285 	 */
1286 	dip = (struct ufs2_dinode *)bp->b_data +
1287 	    ino_to_fsbo(fs, cancelip->i_number);
1288 	if (expungetype == BLK_NOCOPY)
1289 		dip->di_mode = 0;
1290 	dip->di_size = 0;
1291 	dip->di_blocks = 0;
1292 	dip->di_flags &= ~SF_SNAPSHOT;
1293 	bzero(&dip->di_db[0], (NDADDR + NIADDR) * sizeof(ufs2_daddr_t));
1294 	bdwrite(bp);
1295 	/*
1296 	 * Now go through and expunge all the blocks in the file
1297 	 * using the function requested.
1298 	 */
1299 	numblks = howmany(cancelip->i_size, fs->fs_bsize);
1300 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_db[0],
1301 	    &cancelip->i_din2->di_db[NDADDR], fs, 0, expungetype)))
1302 		return (error);
1303 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_ib[0],
1304 	    &cancelip->i_din2->di_ib[NIADDR], fs, -1, expungetype)))
1305 		return (error);
1306 	blksperindir = 1;
1307 	lbn = -NDADDR;
1308 	len = numblks - NDADDR;
1309 	rlbn = NDADDR;
1310 	for (i = 0; len > 0 && i < NIADDR; i++) {
1311 		error = indiracct_ufs2(snapvp, ITOV(cancelip), i,
1312 		    cancelip->i_din2->di_ib[i], lbn, rlbn, len,
1313 		    blksperindir, fs, acctfunc, expungetype);
1314 		if (error)
1315 			return (error);
1316 		blksperindir *= NINDIR(fs);
1317 		lbn -= blksperindir + 1;
1318 		len -= blksperindir;
1319 		rlbn += blksperindir;
1320 	}
1321 	return (0);
1322 }
1323 
1324 /*
1325  * Descend an indirect block chain for vnode cancelvp accounting for all
1326  * its indirect blocks in snapvp.
1327  */
1328 static int
1329 indiracct_ufs2(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks,
1330 	    blksperindir, fs, acctfunc, expungetype)
1331 	struct vnode *snapvp;
1332 	struct vnode *cancelvp;
1333 	int level;
1334 	ufs2_daddr_t blkno;
1335 	ufs_lbn_t lbn;
1336 	ufs_lbn_t rlbn;
1337 	ufs_lbn_t remblks;
1338 	ufs_lbn_t blksperindir;
1339 	struct fs *fs;
1340 	int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
1341 	    struct fs *, ufs_lbn_t, int);
1342 	int expungetype;
1343 {
1344 	int error, num, i;
1345 	ufs_lbn_t subblksperindir;
1346 	struct indir indirs[NIADDR + 2];
1347 	ufs2_daddr_t last, *bap;
1348 	struct buf *bp;
1349 
1350 	if (blkno == 0) {
1351 		if (expungetype == BLK_NOCOPY)
1352 			return (0);
1353 		panic("indiracct_ufs2: missing indir");
1354 	}
1355 	if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0)
1356 		return (error);
1357 	if (lbn != indirs[num - 1 - level].in_lbn || num < 2)
1358 		panic("indiracct_ufs2: botched params");
1359 	/*
1360 	 * We have to expand bread here since it will deadlock looking
1361 	 * up the block number for any blocks that are not in the cache.
1362 	 */
1363 	bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0);
1364 	bp->b_blkno = fsbtodb(fs, blkno);
1365 	if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0 &&
1366 	    (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) {
1367 		brelse(bp);
1368 		return (error);
1369 	}
1370 	/*
1371 	 * Account for the block pointers in this indirect block.
1372 	 */
1373 	last = howmany(remblks, blksperindir);
1374 	if (last > NINDIR(fs))
1375 		last = NINDIR(fs);
1376 	MALLOC(bap, ufs2_daddr_t *, fs->fs_bsize, M_DEVBUF, M_WAITOK);
1377 	bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize);
1378 	bqrelse(bp);
1379 	error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs,
1380 	    level == 0 ? rlbn : -1, expungetype);
1381 	if (error || level == 0)
1382 		goto out;
1383 	/*
1384 	 * Account for the block pointers in each of the indirect blocks
1385 	 * in the levels below us.
1386 	 */
1387 	subblksperindir = blksperindir / NINDIR(fs);
1388 	for (lbn++, level--, i = 0; i < last; i++) {
1389 		error = indiracct_ufs2(snapvp, cancelvp, level, bap[i], lbn,
1390 		    rlbn, remblks, subblksperindir, fs, acctfunc, expungetype);
1391 		if (error)
1392 			goto out;
1393 		rlbn += blksperindir;
1394 		lbn -= blksperindir;
1395 		remblks -= blksperindir;
1396 	}
1397 out:
1398 	FREE(bap, M_DEVBUF);
1399 	return (error);
1400 }
1401 
1402 /*
1403  * Do both snap accounting and map accounting.
1404  */
1405 static int
1406 fullacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype)
1407 	struct vnode *vp;
1408 	ufs2_daddr_t *oldblkp, *lastblkp;
1409 	struct fs *fs;
1410 	ufs_lbn_t lblkno;
1411 	int exptype;	/* BLK_SNAP or BLK_NOCOPY */
1412 {
1413 	int error;
1414 
1415 	if ((error = snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype)))
1416 		return (error);
1417 	return (mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype));
1418 }
1419 
1420 /*
1421  * Identify a set of blocks allocated in a snapshot inode.
1422  */
1423 static int
1424 snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1425 	struct vnode *vp;
1426 	ufs2_daddr_t *oldblkp, *lastblkp;
1427 	struct fs *fs;
1428 	ufs_lbn_t lblkno;
1429 	int expungetype;	/* BLK_SNAP or BLK_NOCOPY */
1430 {
1431 	struct inode *ip = VTOI(vp);
1432 	ufs2_daddr_t blkno, *blkp;
1433 	ufs_lbn_t lbn;
1434 	struct buf *ibp;
1435 	int error;
1436 
1437 	for ( ; oldblkp < lastblkp; oldblkp++) {
1438 		blkno = *oldblkp;
1439 		if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP)
1440 			continue;
1441 		lbn = fragstoblks(fs, blkno);
1442 		if (lbn < NDADDR) {
1443 			blkp = &ip->i_din2->di_db[lbn];
1444 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1445 		} else {
1446 			error = ffs_balloc_ufs2(vp, lblktosize(fs, (off_t)lbn),
1447 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1448 			if (error)
1449 				return (error);
1450 			blkp = &((ufs2_daddr_t *)(ibp->b_data))
1451 			    [(lbn - NDADDR) % NINDIR(fs)];
1452 		}
1453 		/*
1454 		 * If we are expunging a snapshot vnode and we
1455 		 * find a block marked BLK_NOCOPY, then it is
1456 		 * one that has been allocated to this snapshot after
1457 		 * we took our current snapshot and can be ignored.
1458 		 */
1459 		if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) {
1460 			if (lbn >= NDADDR)
1461 				brelse(ibp);
1462 		} else {
1463 			if (*blkp != 0)
1464 				panic("snapacct_ufs2: bad block");
1465 			*blkp = expungetype;
1466 			if (lbn >= NDADDR)
1467 				bdwrite(ibp);
1468 		}
1469 	}
1470 	return (0);
1471 }
1472 
1473 /*
1474  * Account for a set of blocks allocated in a snapshot inode.
1475  */
1476 static int
1477 mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1478 	struct vnode *vp;
1479 	ufs2_daddr_t *oldblkp, *lastblkp;
1480 	struct fs *fs;
1481 	ufs_lbn_t lblkno;
1482 	int expungetype;
1483 {
1484 	ufs2_daddr_t blkno;
1485 	struct inode *ip;
1486 	ino_t inum;
1487 	int acctit;
1488 
1489 	ip = VTOI(vp);
1490 	inum = ip->i_number;
1491 	if (lblkno == -1)
1492 		acctit = 0;
1493 	else
1494 		acctit = 1;
1495 	for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) {
1496 		blkno = *oldblkp;
1497 		if (blkno == 0 || blkno == BLK_NOCOPY)
1498 			continue;
1499 		if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP)
1500 			*ip->i_snapblklist++ = lblkno;
1501 		if (blkno == BLK_SNAP)
1502 			blkno = blkstofrags(fs, lblkno);
1503 		ffs_blkfree(ip->i_ump, fs, vp, blkno, fs->fs_bsize, inum);
1504 	}
1505 	return (0);
1506 }
1507 
1508 /*
1509  * Decrement extra reference on snapshot when last name is removed.
1510  * It will not be freed until the last open reference goes away.
1511  */
1512 void
1513 ffs_snapgone(ip)
1514 	struct inode *ip;
1515 {
1516 	struct inode *xp;
1517 	struct fs *fs;
1518 	int snaploc;
1519 	struct snapdata *sn;
1520 	struct ufsmount *ump;
1521 
1522 	/*
1523 	 * Find snapshot in incore list.
1524 	 */
1525 	xp = NULL;
1526 	sn = ip->i_devvp->v_rdev->si_snapdata;
1527 	if (sn != NULL)
1528 		TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap)
1529 			if (xp == ip)
1530 				break;
1531 	if (xp != NULL)
1532 		vrele(ITOV(ip));
1533 	else if (snapdebug)
1534 		printf("ffs_snapgone: lost snapshot vnode %d\n",
1535 		    ip->i_number);
1536 	/*
1537 	 * Delete snapshot inode from superblock. Keep list dense.
1538 	 */
1539 	fs = ip->i_fs;
1540 	ump = ip->i_ump;
1541 	UFS_LOCK(ump);
1542 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++)
1543 		if (fs->fs_snapinum[snaploc] == ip->i_number)
1544 			break;
1545 	if (snaploc < FSMAXSNAP) {
1546 		for (snaploc++; snaploc < FSMAXSNAP; snaploc++) {
1547 			if (fs->fs_snapinum[snaploc] == 0)
1548 				break;
1549 			fs->fs_snapinum[snaploc - 1] = fs->fs_snapinum[snaploc];
1550 		}
1551 		fs->fs_snapinum[snaploc - 1] = 0;
1552 	}
1553 	UFS_UNLOCK(ump);
1554 }
1555 
1556 /*
1557  * Prepare a snapshot file for being removed.
1558  */
1559 void
1560 ffs_snapremove(vp)
1561 	struct vnode *vp;
1562 {
1563 	struct inode *ip;
1564 	struct vnode *devvp;
1565 	struct buf *ibp;
1566 	struct fs *fs;
1567 	struct thread *td = curthread;
1568 	ufs2_daddr_t numblks, blkno, dblk;
1569 	int error, loc, last;
1570 	struct snapdata *sn;
1571 
1572 	ip = VTOI(vp);
1573 	fs = ip->i_fs;
1574 	devvp = ip->i_devvp;
1575 	/*
1576 	 * If active, delete from incore list (this snapshot may
1577 	 * already have been in the process of being deleted, so
1578 	 * would not have been active).
1579 	 *
1580 	 * Clear copy-on-write flag if last snapshot.
1581 	 */
1582 	VI_LOCK(devvp);
1583 	if (ip->i_nextsnap.tqe_prev != 0) {
1584 		sn = devvp->v_rdev->si_snapdata;
1585 		TAILQ_REMOVE(&sn->sn_head, ip, i_nextsnap);
1586 		ip->i_nextsnap.tqe_prev = 0;
1587 		VI_UNLOCK(devvp);
1588 		lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL, td);
1589 		VI_LOCK(vp);
1590 		KASSERT(vp->v_vnlock == &sn->sn_lock,
1591 			("ffs_snapremove: lost lock mutation"));
1592 		vp->v_vnlock = &vp->v_lock;
1593 		VI_UNLOCK(vp);
1594 		VI_LOCK(devvp);
1595 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL, td);
1596 		try_free_snapdata(devvp, td);
1597 	} else
1598 		VI_UNLOCK(devvp);
1599 	/*
1600 	 * Clear all BLK_NOCOPY fields. Pass any block claims to other
1601 	 * snapshots that want them (see ffs_snapblkfree below).
1602 	 */
1603 	for (blkno = 1; blkno < NDADDR; blkno++) {
1604 		dblk = DIP(ip, i_db[blkno]);
1605 		if (dblk == 0)
1606 			continue;
1607 		if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1608 			DIP_SET(ip, i_db[blkno], 0);
1609 		else if ((dblk == blkstofrags(fs, blkno) &&
1610 		     ffs_snapblkfree(fs, ip->i_devvp, dblk, fs->fs_bsize,
1611 		     ip->i_number))) {
1612 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) -
1613 			    btodb(fs->fs_bsize));
1614 			DIP_SET(ip, i_db[blkno], 0);
1615 		}
1616 	}
1617 	numblks = howmany(ip->i_size, fs->fs_bsize);
1618 	for (blkno = NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
1619 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
1620 		    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1621 		if (error)
1622 			continue;
1623 		if (fs->fs_size - blkno > NINDIR(fs))
1624 			last = NINDIR(fs);
1625 		else
1626 			last = fs->fs_size - blkno;
1627 		for (loc = 0; loc < last; loc++) {
1628 			if (ip->i_ump->um_fstype == UFS1) {
1629 				dblk = ((ufs1_daddr_t *)(ibp->b_data))[loc];
1630 				if (dblk == 0)
1631 					continue;
1632 				if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1633 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1634 				else if ((dblk == blkstofrags(fs, blkno) &&
1635 				     ffs_snapblkfree(fs, ip->i_devvp, dblk,
1636 				     fs->fs_bsize, ip->i_number))) {
1637 					ip->i_din1->di_blocks -=
1638 					    btodb(fs->fs_bsize);
1639 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1640 				}
1641 				continue;
1642 			}
1643 			dblk = ((ufs2_daddr_t *)(ibp->b_data))[loc];
1644 			if (dblk == 0)
1645 				continue;
1646 			if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1647 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1648 			else if ((dblk == blkstofrags(fs, blkno) &&
1649 			     ffs_snapblkfree(fs, ip->i_devvp, dblk,
1650 			     fs->fs_bsize, ip->i_number))) {
1651 				ip->i_din2->di_blocks -= btodb(fs->fs_bsize);
1652 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1653 			}
1654 		}
1655 		bawrite(ibp);
1656 	}
1657 	/*
1658 	 * Clear snapshot flag and drop reference.
1659 	 */
1660 	ip->i_flags &= ~SF_SNAPSHOT;
1661 	DIP_SET(ip, i_flags, ip->i_flags);
1662 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
1663 #ifdef QUOTA
1664 	/*
1665 	 * Reenable disk quotas for ex-snapshot file.
1666 	 */
1667 	if (!getinoquota(ip))
1668 		(void) chkdq(ip, DIP(ip, i_blocks), KERNCRED, FORCE);
1669 #endif
1670 }
1671 
1672 /*
1673  * Notification that a block is being freed. Return zero if the free
1674  * should be allowed to proceed. Return non-zero if the snapshot file
1675  * wants to claim the block. The block will be claimed if it is an
1676  * uncopied part of one of the snapshots. It will be freed if it is
1677  * either a BLK_NOCOPY or has already been copied in all of the snapshots.
1678  * If a fragment is being freed, then all snapshots that care about
1679  * it must make a copy since a snapshot file can only claim full sized
1680  * blocks. Note that if more than one snapshot file maps the block,
1681  * we can pick one at random to claim it. Since none of the snapshots
1682  * can change, we are assurred that they will all see the same unmodified
1683  * image. When deleting a snapshot file (see ffs_snapremove above), we
1684  * must push any of these claimed blocks to one of the other snapshots
1685  * that maps it. These claimed blocks are easily identified as they will
1686  * have a block number equal to their logical block number within the
1687  * snapshot. A copied block can never have this property because they
1688  * must always have been allocated from a BLK_NOCOPY location.
1689  */
1690 int
1691 ffs_snapblkfree(fs, devvp, bno, size, inum)
1692 	struct fs *fs;
1693 	struct vnode *devvp;
1694 	ufs2_daddr_t bno;
1695 	long size;
1696 	ino_t inum;
1697 {
1698 	struct buf *ibp, *cbp, *savedcbp = 0;
1699 	struct thread *td = curthread;
1700 	struct inode *ip;
1701 	struct vnode *vp = NULL;
1702 	ufs_lbn_t lbn;
1703 	ufs2_daddr_t blkno;
1704 	int indiroff = 0, error = 0, claimedblk = 0;
1705 	struct snapdata *sn;
1706 
1707 	lbn = fragstoblks(fs, bno);
1708 retry:
1709 	VI_LOCK(devvp);
1710 	sn = devvp->v_rdev->si_snapdata;
1711 	if (sn == NULL) {
1712 		VI_UNLOCK(devvp);
1713 		return (0);
1714 	}
1715 	if (lockmgr(&sn->sn_lock,
1716 		    LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
1717 		    VI_MTX(devvp), td) != 0)
1718 		goto retry;
1719 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
1720 		vp = ITOV(ip);
1721 		/*
1722 		 * Lookup block being written.
1723 		 */
1724 		if (lbn < NDADDR) {
1725 			blkno = DIP(ip, i_db[lbn]);
1726 		} else {
1727 			td->td_pflags |= TDP_COWINPROGRESS;
1728 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1729 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1730 			td->td_pflags &= ~TDP_COWINPROGRESS;
1731 			if (error)
1732 				break;
1733 			indiroff = (lbn - NDADDR) % NINDIR(fs);
1734 			if (ip->i_ump->um_fstype == UFS1)
1735 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
1736 			else
1737 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
1738 		}
1739 		/*
1740 		 * Check to see if block needs to be copied.
1741 		 */
1742 		if (blkno == 0) {
1743 			/*
1744 			 * A block that we map is being freed. If it has not
1745 			 * been claimed yet, we will claim or copy it (below).
1746 			 */
1747 			claimedblk = 1;
1748 		} else if (blkno == BLK_SNAP) {
1749 			/*
1750 			 * No previous snapshot claimed the block,
1751 			 * so it will be freed and become a BLK_NOCOPY
1752 			 * (don't care) for us.
1753 			 */
1754 			if (claimedblk)
1755 				panic("snapblkfree: inconsistent block type");
1756 			if (lbn < NDADDR) {
1757 				DIP_SET(ip, i_db[lbn], BLK_NOCOPY);
1758 				ip->i_flag |= IN_CHANGE | IN_UPDATE;
1759 			} else if (ip->i_ump->um_fstype == UFS1) {
1760 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] =
1761 				    BLK_NOCOPY;
1762 				bdwrite(ibp);
1763 			} else {
1764 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] =
1765 				    BLK_NOCOPY;
1766 				bdwrite(ibp);
1767 			}
1768 			continue;
1769 		} else /* BLK_NOCOPY or default */ {
1770 			/*
1771 			 * If the snapshot has already copied the block
1772 			 * (default), or does not care about the block,
1773 			 * it is not needed.
1774 			 */
1775 			if (lbn >= NDADDR)
1776 				bqrelse(ibp);
1777 			continue;
1778 		}
1779 		/*
1780 		 * If this is a full size block, we will just grab it
1781 		 * and assign it to the snapshot inode. Otherwise we
1782 		 * will proceed to copy it. See explanation for this
1783 		 * routine as to why only a single snapshot needs to
1784 		 * claim this block.
1785 		 */
1786 		if (size == fs->fs_bsize) {
1787 #ifdef DEBUG
1788 			if (snapdebug)
1789 				printf("%s %d lbn %jd from inum %d\n",
1790 				    "Grabonremove: snapino", ip->i_number,
1791 				    (intmax_t)lbn, inum);
1792 #endif
1793 			if (lbn < NDADDR) {
1794 				DIP_SET(ip, i_db[lbn], bno);
1795 			} else if (ip->i_ump->um_fstype == UFS1) {
1796 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] = bno;
1797 				bdwrite(ibp);
1798 			} else {
1799 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] = bno;
1800 				bdwrite(ibp);
1801 			}
1802 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + btodb(size));
1803 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1804 			lockmgr(vp->v_vnlock, LK_RELEASE, NULL, td);
1805 			return (1);
1806 		}
1807 		if (lbn >= NDADDR)
1808 			bqrelse(ibp);
1809 		/*
1810 		 * Allocate the block into which to do the copy. Note that this
1811 		 * allocation will never require any additional allocations for
1812 		 * the snapshot inode.
1813 		 */
1814 		td->td_pflags |= TDP_COWINPROGRESS;
1815 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1816 		    fs->fs_bsize, KERNCRED, 0, &cbp);
1817 		td->td_pflags &= ~TDP_COWINPROGRESS;
1818 		if (error)
1819 			break;
1820 #ifdef DEBUG
1821 		if (snapdebug)
1822 			printf("%s%d lbn %jd %s %d size %ld to blkno %jd\n",
1823 			    "Copyonremove: snapino ", ip->i_number,
1824 			    (intmax_t)lbn, "for inum", inum, size,
1825 			    (intmax_t)cbp->b_blkno);
1826 #endif
1827 		/*
1828 		 * If we have already read the old block contents, then
1829 		 * simply copy them to the new block. Note that we need
1830 		 * to synchronously write snapshots that have not been
1831 		 * unlinked, and hence will be visible after a crash,
1832 		 * to ensure their integrity.
1833 		 */
1834 		if (savedcbp != 0) {
1835 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
1836 			bawrite(cbp);
1837 			if (dopersistence && ip->i_effnlink > 0)
1838 				(void) ffs_syncvnode(vp, MNT_WAIT);
1839 			continue;
1840 		}
1841 		/*
1842 		 * Otherwise, read the old block contents into the buffer.
1843 		 */
1844 		if ((error = readblock(vp, cbp, lbn)) != 0) {
1845 			bzero(cbp->b_data, fs->fs_bsize);
1846 			bawrite(cbp);
1847 			if (dopersistence && ip->i_effnlink > 0)
1848 				(void) ffs_syncvnode(vp, MNT_WAIT);
1849 			break;
1850 		}
1851 		savedcbp = cbp;
1852 	}
1853 	/*
1854 	 * Note that we need to synchronously write snapshots that
1855 	 * have not been unlinked, and hence will be visible after
1856 	 * a crash, to ensure their integrity.
1857 	 */
1858 	if (savedcbp) {
1859 		vp = savedcbp->b_vp;
1860 		bawrite(savedcbp);
1861 		if (dopersistence && VTOI(vp)->i_effnlink > 0)
1862 			(void) ffs_syncvnode(vp, MNT_WAIT);
1863 	}
1864 	/*
1865 	 * If we have been unable to allocate a block in which to do
1866 	 * the copy, then return non-zero so that the fragment will
1867 	 * not be freed. Although space will be lost, the snapshot
1868 	 * will stay consistent.
1869 	 */
1870 	lockmgr(vp->v_vnlock, LK_RELEASE, NULL, td);
1871 	return (error);
1872 }
1873 
1874 /*
1875  * Associate snapshot files when mounting.
1876  */
1877 void
1878 ffs_snapshot_mount(mp)
1879 	struct mount *mp;
1880 {
1881 	struct ufsmount *ump = VFSTOUFS(mp);
1882 	struct vnode *devvp = ump->um_devvp;
1883 	struct fs *fs = ump->um_fs;
1884 	struct thread *td = curthread;
1885 	struct snapdata *sn;
1886 	struct vnode *vp;
1887 	struct vnode *lastvp;
1888 	struct inode *ip;
1889 	struct uio auio;
1890 	struct iovec aiov;
1891 	void *snapblklist;
1892 	char *reason;
1893 	daddr_t snaplistsize;
1894 	int error, snaploc, loc;
1895 
1896 	/*
1897 	 * XXX The following needs to be set before ffs_truncate or
1898 	 * VOP_READ can be called.
1899 	 */
1900 	mp->mnt_stat.f_iosize = fs->fs_bsize;
1901 	/*
1902 	 * Process each snapshot listed in the superblock.
1903 	 */
1904 	vp = NULL;
1905 	lastvp = NULL;
1906 	sn = devvp->v_rdev->si_snapdata;
1907 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++) {
1908 		if (fs->fs_snapinum[snaploc] == 0)
1909 			break;
1910 		if ((error = ffs_vget(mp, fs->fs_snapinum[snaploc],
1911 		    LK_EXCLUSIVE, &vp)) != 0){
1912 			printf("ffs_snapshot_mount: vget failed %d\n", error);
1913 			continue;
1914 		}
1915 		ip = VTOI(vp);
1916 		if ((ip->i_flags & SF_SNAPSHOT) == 0 || ip->i_size ==
1917 		    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag))) {
1918 			if ((ip->i_flags & SF_SNAPSHOT) == 0) {
1919 				reason = "non-snapshot";
1920 			} else {
1921 				reason = "old format snapshot";
1922 				(void)ffs_truncate(vp, (off_t)0, 0, NOCRED, td);
1923 				(void)ffs_syncvnode(vp, MNT_WAIT);
1924 			}
1925 			printf("ffs_snapshot_mount: %s inode %d\n",
1926 			    reason, fs->fs_snapinum[snaploc]);
1927 			vput(vp);
1928 			vp = NULL;
1929 			for (loc = snaploc + 1; loc < FSMAXSNAP; loc++) {
1930 				if (fs->fs_snapinum[loc] == 0)
1931 					break;
1932 				fs->fs_snapinum[loc - 1] = fs->fs_snapinum[loc];
1933 			}
1934 			fs->fs_snapinum[loc - 1] = 0;
1935 			snaploc--;
1936 			continue;
1937 		}
1938 		/*
1939 		 * If there already exist snapshots on this filesystem, grab a
1940 		 * reference to their shared lock. If this is the first snapshot
1941 		 * on this filesystem, we need to allocate a lock for the
1942 		 * snapshots to share. In either case, acquire the snapshot
1943 		 * lock and give up our original private lock.
1944 		 */
1945 		VI_LOCK(devvp);
1946 		if (sn != NULL) {
1947 
1948 			VI_UNLOCK(devvp);
1949 			VI_LOCK(vp);
1950 			vp->v_vnlock = &sn->sn_lock;
1951 		} else {
1952 			VI_UNLOCK(devvp);
1953 			sn = malloc(sizeof *sn, M_UFSMNT, M_WAITOK | M_ZERO);
1954 			TAILQ_INIT(&sn->sn_head);
1955 			lockinit(&sn->sn_lock, PVFS, "snaplk", VLKTIMEOUT,
1956 			    LK_CANRECURSE | LK_NOSHARE);
1957 			VI_LOCK(vp);
1958 			vp->v_vnlock = &sn->sn_lock;
1959 			devvp->v_rdev->si_snapdata = sn;
1960 		}
1961 		lockmgr(vp->v_vnlock, LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY,
1962 		    VI_MTX(vp), td);
1963 		lockmgr(&vp->v_lock, LK_RELEASE, NULL, td);
1964 		/*
1965 		 * Link it onto the active snapshot list.
1966 		 */
1967 		VI_LOCK(devvp);
1968 		if (ip->i_nextsnap.tqe_prev != 0)
1969 			panic("ffs_snapshot_mount: %d already on list",
1970 			    ip->i_number);
1971 		else
1972 			TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
1973 		vp->v_vflag |= VV_SYSTEM;
1974 		VI_UNLOCK(devvp);
1975 		VOP_UNLOCK(vp, 0, td);
1976 		lastvp = vp;
1977 	}
1978 	vp = lastvp;
1979 	/*
1980 	 * No usable snapshots found.
1981 	 */
1982 	if (vp == NULL)
1983 		return;
1984 	/*
1985 	 * Allocate the space for the block hints list. We always want to
1986 	 * use the list from the newest snapshot.
1987 	 */
1988 	auio.uio_iov = &aiov;
1989 	auio.uio_iovcnt = 1;
1990 	aiov.iov_base = (void *)&snaplistsize;
1991 	aiov.iov_len = sizeof(snaplistsize);
1992 	auio.uio_resid = aiov.iov_len;
1993 	auio.uio_offset =
1994 	    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag));
1995 	auio.uio_segflg = UIO_SYSSPACE;
1996 	auio.uio_rw = UIO_READ;
1997 	auio.uio_td = td;
1998 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1999 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
2000 		printf("ffs_snapshot_mount: read_1 failed %d\n", error);
2001 		VOP_UNLOCK(vp, 0, td);
2002 		return;
2003 	}
2004 	MALLOC(snapblklist, void *, snaplistsize * sizeof(daddr_t),
2005 	    M_UFSMNT, M_WAITOK);
2006 	auio.uio_iovcnt = 1;
2007 	aiov.iov_base = snapblklist;
2008 	aiov.iov_len = snaplistsize * sizeof (daddr_t);
2009 	auio.uio_resid = aiov.iov_len;
2010 	auio.uio_offset -= sizeof(snaplistsize);
2011 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
2012 		printf("ffs_snapshot_mount: read_2 failed %d\n", error);
2013 		VOP_UNLOCK(vp, 0, td);
2014 		FREE(snapblklist, M_UFSMNT);
2015 		return;
2016 	}
2017 	VOP_UNLOCK(vp, 0, td);
2018 	VI_LOCK(devvp);
2019 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_mount");
2020 	sn->sn_listsize = snaplistsize;
2021 	sn->sn_blklist = (daddr_t *)snapblklist;
2022 	devvp->v_vflag |= VV_COPYONWRITE;
2023 	VI_UNLOCK(devvp);
2024 }
2025 
2026 /*
2027  * Disassociate snapshot files when unmounting.
2028  */
2029 void
2030 ffs_snapshot_unmount(mp)
2031 	struct mount *mp;
2032 {
2033 	struct vnode *devvp = VFSTOUFS(mp)->um_devvp;
2034 	struct snapdata *sn;
2035 	struct inode *xp;
2036 	struct vnode *vp;
2037 	struct thread *td = curthread;
2038 
2039 	VI_LOCK(devvp);
2040 	sn = devvp->v_rdev->si_snapdata;
2041 	while (sn != NULL && (xp = TAILQ_FIRST(&sn->sn_head)) != NULL) {
2042 		vp = ITOV(xp);
2043 		TAILQ_REMOVE(&sn->sn_head, xp, i_nextsnap);
2044 		xp->i_nextsnap.tqe_prev = 0;
2045 		lockmgr(&sn->sn_lock,
2046 			LK_INTERLOCK | LK_EXCLUSIVE,
2047 			VI_MTX(devvp),
2048 			td);
2049 		VI_LOCK(vp);
2050 		lockmgr(&vp->v_lock,
2051 			LK_INTERLOCK | LK_EXCLUSIVE,
2052 			VI_MTX(vp), td);
2053 		VI_LOCK(vp);
2054 		KASSERT(vp->v_vnlock == &sn->sn_lock,
2055 		("ffs_snapshot_unmount: lost lock mutation"));
2056 		vp->v_vnlock = &vp->v_lock;
2057 		VI_UNLOCK(vp);
2058 		lockmgr(&vp->v_lock, LK_RELEASE, NULL, td);
2059 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL, td);
2060 		if (xp->i_effnlink > 0)
2061 			vrele(vp);
2062 		VI_LOCK(devvp);
2063 		sn = devvp->v_rdev->si_snapdata;
2064 	}
2065 	try_free_snapdata(devvp, td);
2066 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_unmount");
2067 }
2068 
2069 /*
2070  * Check the buffer block to be belong to device buffer that shall be
2071  * locked after snaplk. devvp shall be locked on entry, and will be
2072  * leaved locked upon exit.
2073  */
2074 static int
2075 ffs_bp_snapblk(devvp, bp)
2076 	struct vnode *devvp;
2077 	struct buf *bp;
2078 {
2079 	struct snapdata *sn;
2080 	struct fs *fs;
2081 	ufs2_daddr_t lbn, *snapblklist;
2082 	int lower, upper, mid;
2083 
2084 	ASSERT_VI_LOCKED(devvp, "ffs_bp_snapblk");
2085 	KASSERT(devvp->v_type == VCHR, ("Not a device %p", devvp));
2086 	sn = devvp->v_rdev->si_snapdata;
2087 	if (sn == NULL || TAILQ_FIRST(&sn->sn_head) == NULL)
2088 		return (0);
2089 	fs = TAILQ_FIRST(&sn->sn_head)->i_fs;
2090 	lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno));
2091 	snapblklist = sn->sn_blklist;
2092 	upper = sn->sn_listsize - 1;
2093 	lower = 1;
2094 	while (lower <= upper) {
2095 		mid = (lower + upper) / 2;
2096 		if (snapblklist[mid] == lbn)
2097 			break;
2098 		if (snapblklist[mid] < lbn)
2099 			lower = mid + 1;
2100 		else
2101 			upper = mid - 1;
2102 	}
2103 	if (lower <= upper)
2104 		return (1);
2105 	return (0);
2106 }
2107 
2108 void
2109 ffs_bdflush(bo, bp)
2110 	struct bufobj *bo;
2111 	struct buf *bp;
2112 {
2113 	struct thread *td;
2114 	struct vnode *vp, *devvp;
2115 	struct buf *nbp;
2116 	int bp_bdskip;
2117 
2118 	if (bo->bo_dirty.bv_cnt <= dirtybufthresh)
2119 		return;
2120 
2121 	td = curthread;
2122 	vp = bp->b_vp;
2123 	devvp = bo->__bo_vnode;
2124 	KASSERT(vp == devvp, ("devvp != vp %p %p", bo, bp));
2125 
2126 	VI_LOCK(devvp);
2127 	bp_bdskip = ffs_bp_snapblk(devvp, bp);
2128 	if (bp_bdskip)
2129 		bdwriteskip++;
2130 	VI_UNLOCK(devvp);
2131 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10 && !bp_bdskip) {
2132 		(void) VOP_FSYNC(vp, MNT_NOWAIT, td);
2133 		altbufferflushes++;
2134 	} else {
2135 		BO_LOCK(bo);
2136 		/*
2137 		 * Try to find a buffer to flush.
2138 		 */
2139 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2140 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2141 			    BUF_LOCK(nbp,
2142 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2143 				continue;
2144 			if (bp == nbp)
2145 				panic("bdwrite: found ourselves");
2146 			BO_UNLOCK(bo);
2147 			/*
2148 			 * Don't countdeps with the bo lock
2149 			 * held.
2150 			 */
2151 			if (buf_countdeps(nbp, 0)) {
2152 				BO_LOCK(bo);
2153 				BUF_UNLOCK(nbp);
2154 				continue;
2155 			}
2156 			if (bp_bdskip) {
2157 				VI_LOCK(devvp);
2158 				if (!ffs_bp_snapblk(vp, nbp)) {
2159 					if (BO_MTX(bo) != VI_MTX(vp)) {
2160 						VI_UNLOCK(devvp);
2161 						BO_LOCK(bo);
2162 					}
2163 					BUF_UNLOCK(nbp);
2164 					continue;
2165 				}
2166 				VI_UNLOCK(devvp);
2167 			}
2168 			if (nbp->b_flags & B_CLUSTEROK) {
2169 				vfs_bio_awrite(nbp);
2170 			} else {
2171 				bremfree(nbp);
2172 				bawrite(nbp);
2173 			}
2174 			dirtybufferflushes++;
2175 			break;
2176 		}
2177 		if (nbp == NULL)
2178 			BO_UNLOCK(bo);
2179 	}
2180 }
2181 
2182 /*
2183  * Check for need to copy block that is about to be written,
2184  * copying the block if necessary.
2185  */
2186 int
2187 ffs_copyonwrite(devvp, bp)
2188 	struct vnode *devvp;
2189 	struct buf *bp;
2190 {
2191 	struct snapdata *sn;
2192 	struct buf *ibp, *cbp, *savedcbp = 0;
2193 	struct thread *td = curthread;
2194 	struct fs *fs;
2195 	struct inode *ip;
2196 	struct vnode *vp = 0;
2197 	ufs2_daddr_t lbn, blkno, *snapblklist;
2198 	int lower, upper, mid, indiroff, error = 0;
2199 	int launched_async_io, prev_norunningbuf;
2200 	long saved_runningbufspace;
2201 
2202 	if (devvp != bp->b_vp && (VTOI(bp->b_vp)->i_flags & SF_SNAPSHOT) != 0)
2203 		return (0);		/* Update on a snapshot file */
2204 	if (td->td_pflags & TDP_COWINPROGRESS)
2205 		panic("ffs_copyonwrite: recursive call");
2206 	/*
2207 	 * First check to see if it is in the preallocated list.
2208 	 * By doing this check we avoid several potential deadlocks.
2209 	 */
2210 	VI_LOCK(devvp);
2211 	sn = devvp->v_rdev->si_snapdata;
2212 	if (sn == NULL ||
2213 	    TAILQ_EMPTY(&sn->sn_head)) {
2214 		VI_UNLOCK(devvp);
2215 		return (0);		/* No snapshot */
2216 	}
2217 	ip = TAILQ_FIRST(&sn->sn_head);
2218 	fs = ip->i_fs;
2219 	lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno));
2220 	snapblklist = sn->sn_blklist;
2221 	upper = sn->sn_listsize - 1;
2222 	lower = 1;
2223 	while (lower <= upper) {
2224 		mid = (lower + upper) / 2;
2225 		if (snapblklist[mid] == lbn)
2226 			break;
2227 		if (snapblklist[mid] < lbn)
2228 			lower = mid + 1;
2229 		else
2230 			upper = mid - 1;
2231 	}
2232 	if (lower <= upper) {
2233 		VI_UNLOCK(devvp);
2234 		return (0);
2235 	}
2236 	launched_async_io = 0;
2237 	prev_norunningbuf = td->td_pflags & TDP_NORUNNINGBUF;
2238 	/*
2239 	 * Since I/O on bp isn't yet in progress and it may be blocked
2240 	 * for a long time waiting on snaplk, back it out of
2241 	 * runningbufspace, possibly waking other threads waiting for space.
2242 	 */
2243 	saved_runningbufspace = bp->b_runningbufspace;
2244 	if (saved_runningbufspace != 0)
2245 		runningbufwakeup(bp);
2246 	/*
2247 	 * Not in the precomputed list, so check the snapshots.
2248 	 */
2249 	while (lockmgr(&sn->sn_lock,
2250 		       LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
2251 		       VI_MTX(devvp), td) != 0) {
2252 		VI_LOCK(devvp);
2253 		sn = devvp->v_rdev->si_snapdata;
2254 		if (sn == NULL ||
2255 		    TAILQ_EMPTY(&sn->sn_head)) {
2256 			VI_UNLOCK(devvp);
2257 			if (saved_runningbufspace != 0) {
2258 				bp->b_runningbufspace = saved_runningbufspace;
2259 				atomic_add_int(&runningbufspace,
2260 					       bp->b_runningbufspace);
2261 			}
2262 			return (0);		/* Snapshot gone */
2263 		}
2264 	}
2265 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
2266 		vp = ITOV(ip);
2267 		/*
2268 		 * We ensure that everything of our own that needs to be
2269 		 * copied will be done at the time that ffs_snapshot is
2270 		 * called. Thus we can skip the check here which can
2271 		 * deadlock in doing the lookup in UFS_BALLOC.
2272 		 */
2273 		if (bp->b_vp == vp)
2274 			continue;
2275 		/*
2276 		 * Check to see if block needs to be copied. We do not have
2277 		 * to hold the snapshot lock while doing this lookup as it
2278 		 * will never require any additional allocations for the
2279 		 * snapshot inode.
2280 		 */
2281 		if (lbn < NDADDR) {
2282 			blkno = DIP(ip, i_db[lbn]);
2283 		} else {
2284 			td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2285 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2286 			   fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
2287 			td->td_pflags &= ~TDP_COWINPROGRESS;
2288 			if (error)
2289 				break;
2290 			indiroff = (lbn - NDADDR) % NINDIR(fs);
2291 			if (ip->i_ump->um_fstype == UFS1)
2292 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
2293 			else
2294 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
2295 			bqrelse(ibp);
2296 		}
2297 #ifdef DIAGNOSTIC
2298 		if (blkno == BLK_SNAP && bp->b_lblkno >= 0)
2299 			panic("ffs_copyonwrite: bad copy block");
2300 #endif
2301 		if (blkno != 0)
2302 			continue;
2303 		/*
2304 		 * Allocate the block into which to do the copy. Since
2305 		 * multiple processes may all try to copy the same block,
2306 		 * we have to recheck our need to do a copy if we sleep
2307 		 * waiting for the lock.
2308 		 *
2309 		 * Because all snapshots on a filesystem share a single
2310 		 * lock, we ensure that we will never be in competition
2311 		 * with another process to allocate a block.
2312 		 */
2313 		td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2314 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2315 		    fs->fs_bsize, KERNCRED, 0, &cbp);
2316 		td->td_pflags &= ~TDP_COWINPROGRESS;
2317 		if (error)
2318 			break;
2319 #ifdef DEBUG
2320 		if (snapdebug) {
2321 			printf("Copyonwrite: snapino %d lbn %jd for ",
2322 			    ip->i_number, (intmax_t)lbn);
2323 			if (bp->b_vp == devvp)
2324 				printf("fs metadata");
2325 			else
2326 				printf("inum %d", VTOI(bp->b_vp)->i_number);
2327 			printf(" lblkno %jd to blkno %jd\n",
2328 			    (intmax_t)bp->b_lblkno, (intmax_t)cbp->b_blkno);
2329 		}
2330 #endif
2331 		/*
2332 		 * If we have already read the old block contents, then
2333 		 * simply copy them to the new block. Note that we need
2334 		 * to synchronously write snapshots that have not been
2335 		 * unlinked, and hence will be visible after a crash,
2336 		 * to ensure their integrity.
2337 		 */
2338 		if (savedcbp != 0) {
2339 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
2340 			bawrite(cbp);
2341 			if (dopersistence && ip->i_effnlink > 0)
2342 				(void) ffs_syncvnode(vp, MNT_WAIT);
2343 			else
2344 				launched_async_io = 1;
2345 			continue;
2346 		}
2347 		/*
2348 		 * Otherwise, read the old block contents into the buffer.
2349 		 */
2350 		if ((error = readblock(vp, cbp, lbn)) != 0) {
2351 			bzero(cbp->b_data, fs->fs_bsize);
2352 			bawrite(cbp);
2353 			if (dopersistence && ip->i_effnlink > 0)
2354 				(void) ffs_syncvnode(vp, MNT_WAIT);
2355 			else
2356 				launched_async_io = 1;
2357 			break;
2358 		}
2359 		savedcbp = cbp;
2360 	}
2361 	/*
2362 	 * Note that we need to synchronously write snapshots that
2363 	 * have not been unlinked, and hence will be visible after
2364 	 * a crash, to ensure their integrity.
2365 	 */
2366 	if (savedcbp) {
2367 		vp = savedcbp->b_vp;
2368 		bawrite(savedcbp);
2369 		if (dopersistence && VTOI(vp)->i_effnlink > 0)
2370 			(void) ffs_syncvnode(vp, MNT_WAIT);
2371 		else
2372 			launched_async_io = 1;
2373 	}
2374 	lockmgr(vp->v_vnlock, LK_RELEASE, NULL, td);
2375 	td->td_pflags = (td->td_pflags & ~TDP_NORUNNINGBUF) |
2376 		prev_norunningbuf;
2377 	if (launched_async_io && (td->td_pflags & TDP_NORUNNINGBUF) == 0)
2378 		waitrunningbufspace();
2379 	/*
2380 	 * I/O on bp will now be started, so count it in runningbufspace.
2381 	 */
2382 	if (saved_runningbufspace != 0) {
2383 		bp->b_runningbufspace = saved_runningbufspace;
2384 		atomic_add_int(&runningbufspace, bp->b_runningbufspace);
2385 	}
2386 	return (error);
2387 }
2388 
2389 /*
2390  * Read the specified block into the given buffer.
2391  * Much of this boiler-plate comes from bwrite().
2392  */
2393 static int
2394 readblock(vp, bp, lbn)
2395 	struct vnode *vp;
2396 	struct buf *bp;
2397 	ufs2_daddr_t lbn;
2398 {
2399 	struct inode *ip = VTOI(vp);
2400 	struct bio *bip;
2401 
2402 	bip = g_alloc_bio();
2403 	bip->bio_cmd = BIO_READ;
2404 	bip->bio_offset = dbtob(fsbtodb(ip->i_fs, blkstofrags(ip->i_fs, lbn)));
2405 	bip->bio_data = bp->b_data;
2406 	bip->bio_length = bp->b_bcount;
2407 	bip->bio_done = NULL;
2408 
2409 	g_io_request(bip, ip->i_devvp->v_bufobj.bo_private);
2410 	bp->b_error = biowait(bip, "snaprdb");
2411 	g_destroy_bio(bip);
2412 	return (bp->b_error);
2413 }
2414 
2415 /*
2416  * Process file deletes that were deferred by ufs_inactive() due to
2417  * the file system being suspended. Transfer IN_LAZYACCESS into
2418  * IN_MODIFIED for vnodes that were accessed during suspension.
2419  */
2420 static void
2421 process_deferred_inactive(struct mount *mp)
2422 {
2423 	struct vnode *vp, *mvp;
2424 	struct inode *ip;
2425 	struct thread *td;
2426 	int error;
2427 
2428 	td = curthread;
2429 	(void) vn_start_secondary_write(NULL, &mp, V_WAIT);
2430 	MNT_ILOCK(mp);
2431  loop:
2432 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2433 		VI_LOCK(vp);
2434 		/*
2435 		 * IN_LAZYACCESS is checked here without holding any
2436 		 * vnode lock, but this flag is set only while holding
2437 		 * vnode interlock.
2438 		 */
2439 		if (vp->v_type == VNON || (vp->v_iflag & VI_DOOMED) != 0 ||
2440 		    ((VTOI(vp)->i_flag & IN_LAZYACCESS) == 0 &&
2441 			((vp->v_iflag & VI_OWEINACT) == 0 ||
2442 			vp->v_usecount > 0))) {
2443 			VI_UNLOCK(vp);
2444 			continue;
2445 		}
2446 		MNT_IUNLOCK(mp);
2447 		vholdl(vp);
2448 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td);
2449 		if (error != 0) {
2450 			vdrop(vp);
2451 			MNT_ILOCK(mp);
2452 			if (error == ENOENT)
2453 				continue;	/* vnode recycled */
2454 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2455 			goto loop;
2456 		}
2457 		ip = VTOI(vp);
2458 		if ((ip->i_flag & IN_LAZYACCESS) != 0) {
2459 			ip->i_flag &= ~IN_LAZYACCESS;
2460 			ip->i_flag |= IN_MODIFIED;
2461 		}
2462 		VI_LOCK(vp);
2463 		if ((vp->v_iflag & VI_OWEINACT) == 0 || vp->v_usecount > 0) {
2464 			VI_UNLOCK(vp);
2465 			VOP_UNLOCK(vp, 0, td);
2466 			vdrop(vp);
2467 			MNT_ILOCK(mp);
2468 			continue;
2469 		}
2470 
2471 		VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2472 			 ("process_deferred_inactive: "
2473 			  "recursed on VI_DOINGINACT"));
2474 		vp->v_iflag |= VI_DOINGINACT;
2475 		vp->v_iflag &= ~VI_OWEINACT;
2476 		VI_UNLOCK(vp);
2477 		(void) VOP_INACTIVE(vp, td);
2478 		VI_LOCK(vp);
2479 		VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2480 			 ("process_deferred_inactive: lost VI_DOINGINACT"));
2481 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2482 			 ("process_deferred_inactive: got VI_OWEINACT"));
2483 		vp->v_iflag &= ~VI_DOINGINACT;
2484 		VI_UNLOCK(vp);
2485 		VOP_UNLOCK(vp, 0, td);
2486 		vdrop(vp);
2487 		MNT_ILOCK(mp);
2488 	}
2489 	MNT_IUNLOCK(mp);
2490 	vn_finished_secondary_write(mp);
2491 }
2492 
2493 /* Try to free snapdata associated with devvp */
2494 static void
2495 try_free_snapdata(struct vnode *devvp,
2496 		  struct thread *td)
2497 {
2498 	struct snapdata *sn;
2499 	ufs2_daddr_t *snapblklist;
2500 
2501 	sn = devvp->v_rdev->si_snapdata;
2502 
2503 	if (sn == NULL || TAILQ_FIRST(&sn->sn_head) != NULL ||
2504 	    (devvp->v_vflag & VV_COPYONWRITE) == 0) {
2505 		VI_UNLOCK(devvp);
2506 		return;
2507 	}
2508 
2509 	devvp->v_rdev->si_snapdata = NULL;
2510 	devvp->v_vflag &= ~VV_COPYONWRITE;
2511 	snapblklist = sn->sn_blklist;
2512 	sn->sn_blklist = NULL;
2513 	sn->sn_listsize = 0;
2514 	lockmgr(&sn->sn_lock, LK_DRAIN|LK_INTERLOCK, VI_MTX(devvp), td);
2515 	lockmgr(&sn->sn_lock, LK_RELEASE, NULL, td);
2516 	lockdestroy(&sn->sn_lock);
2517 	free(sn, M_UFSMNT);
2518 	if (snapblklist != NULL)
2519 		FREE(snapblklist, M_UFSMNT);
2520 }
2521 #endif
2522