xref: /freebsd/sys/ufs/ffs/ffs_snapshot.c (revision 409a390c3341fb4f162cd7de1fd595a323ebbfd8)
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/fcntl.h>
48 #include <sys/proc.h>
49 #include <sys/namei.h>
50 #include <sys/sched.h>
51 #include <sys/stat.h>
52 #include <sys/malloc.h>
53 #include <sys/mount.h>
54 #include <sys/resource.h>
55 #include <sys/resourcevar.h>
56 #include <sys/vnode.h>
57 
58 #include <geom/geom.h>
59 
60 #include <ufs/ufs/extattr.h>
61 #include <ufs/ufs/quota.h>
62 #include <ufs/ufs/ufsmount.h>
63 #include <ufs/ufs/inode.h>
64 #include <ufs/ufs/ufs_extern.h>
65 
66 #include <ufs/ffs/fs.h>
67 #include <ufs/ffs/ffs_extern.h>
68 
69 #define KERNCRED thread0.td_ucred
70 #define DEBUG 1
71 
72 #include "opt_ffs.h"
73 
74 #ifdef NO_FFS_SNAPSHOT
75 int
76 ffs_snapshot(mp, snapfile)
77 	struct mount *mp;
78 	char *snapfile;
79 {
80 	return (EINVAL);
81 }
82 
83 int
84 ffs_snapblkfree(fs, devvp, bno, size, inum)
85 	struct fs *fs;
86 	struct vnode *devvp;
87 	ufs2_daddr_t bno;
88 	long size;
89 	ino_t inum;
90 {
91 	return (EINVAL);
92 }
93 
94 void
95 ffs_snapremove(vp)
96 	struct vnode *vp;
97 {
98 }
99 
100 void
101 ffs_snapshot_mount(mp)
102 	struct mount *mp;
103 {
104 }
105 
106 void
107 ffs_snapshot_unmount(mp)
108 	struct mount *mp;
109 {
110 }
111 
112 void
113 ffs_snapgone(ip)
114 	struct inode *ip;
115 {
116 }
117 
118 int
119 ffs_copyonwrite(devvp, bp)
120 	struct vnode *devvp;
121 	struct buf *bp;
122 {
123 	return (EINVAL);
124 }
125 
126 #else
127 
128 TAILQ_HEAD(snaphead, inode);
129 
130 struct snapdata {
131 	LIST_ENTRY(snapdata) sn_link;
132 	struct snaphead sn_head;
133 	daddr_t sn_listsize;
134 	daddr_t *sn_blklist;
135 	struct lock sn_lock;
136 };
137 
138 LIST_HEAD(, snapdata) snapfree;
139 static struct mtx snapfree_lock;
140 MTX_SYSINIT(ffs_snapfree, &snapfree_lock, "snapdata free list", MTX_DEF);
141 
142 static int cgaccount(int, struct vnode *, struct buf *, int);
143 static int expunge_ufs1(struct vnode *, struct inode *, struct fs *,
144     int (*)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, struct fs *,
145     ufs_lbn_t, int), int);
146 static int indiracct_ufs1(struct vnode *, struct vnode *, int,
147     ufs1_daddr_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, struct fs *,
148     int (*)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, struct fs *,
149     ufs_lbn_t, int), int);
150 static int fullacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
151     struct fs *, ufs_lbn_t, int);
152 static int snapacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
153     struct fs *, ufs_lbn_t, int);
154 static int mapacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
155     struct fs *, ufs_lbn_t, int);
156 static int expunge_ufs2(struct vnode *, struct inode *, struct fs *,
157     int (*)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, struct fs *,
158     ufs_lbn_t, int), int);
159 static int indiracct_ufs2(struct vnode *, struct vnode *, int,
160     ufs2_daddr_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, struct fs *,
161     int (*)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, struct fs *,
162     ufs_lbn_t, int), int);
163 static int fullacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
164     struct fs *, ufs_lbn_t, int);
165 static int snapacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
166     struct fs *, ufs_lbn_t, int);
167 static int mapacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
168     struct fs *, ufs_lbn_t, int);
169 static int readblock(struct vnode *vp, struct buf *, ufs2_daddr_t);
170 static void try_free_snapdata(struct vnode *devvp);
171 static struct snapdata *ffs_snapdata_acquire(struct vnode *devvp);
172 static int ffs_bp_snapblk(struct vnode *, struct buf *);
173 
174 /*
175  * To ensure the consistency of snapshots across crashes, we must
176  * synchronously write out copied blocks before allowing the
177  * originals to be modified. Because of the rather severe speed
178  * penalty that this imposes, the following flag allows this
179  * crash persistence to be disabled.
180  */
181 int dopersistence = 0;
182 
183 #ifdef DEBUG
184 #include <sys/sysctl.h>
185 SYSCTL_INT(_debug, OID_AUTO, dopersistence, CTLFLAG_RW, &dopersistence, 0, "");
186 static int snapdebug = 0;
187 SYSCTL_INT(_debug, OID_AUTO, snapdebug, CTLFLAG_RW, &snapdebug, 0, "");
188 int collectsnapstats = 0;
189 SYSCTL_INT(_debug, OID_AUTO, collectsnapstats, CTLFLAG_RW, &collectsnapstats,
190 	0, "");
191 #endif /* DEBUG */
192 
193 /*
194  * Create a snapshot file and initialize it for the filesystem.
195  */
196 int
197 ffs_snapshot(mp, snapfile)
198 	struct mount *mp;
199 	char *snapfile;
200 {
201 	ufs2_daddr_t numblks, blkno, *blkp, *snapblklist;
202 	int error, cg, snaploc;
203 	int i, size, len, loc;
204 	int flag;
205 	struct timespec starttime = {0, 0}, endtime;
206 	char saved_nice = 0;
207 	long redo = 0, snaplistsize = 0;
208 	int32_t *lp;
209 	void *space;
210 	struct fs *copy_fs = NULL, *fs;
211 	struct thread *td = curthread;
212 	struct inode *ip, *xp;
213 	struct buf *bp, *nbp, *ibp, *sbp = NULL;
214 	struct nameidata nd;
215 	struct mount *wrtmp;
216 	struct vattr vat;
217 	struct vnode *vp, *xvp, *mvp, *devvp;
218 	struct uio auio;
219 	struct iovec aiov;
220 	struct snapdata *sn;
221 	struct ufsmount *ump;
222 
223 	ump = VFSTOUFS(mp);
224 	fs = ump->um_fs;
225 	sn = NULL;
226 	MNT_ILOCK(mp);
227 	flag = mp->mnt_flag;
228 	MNT_IUNLOCK(mp);
229 
230 	/*
231 	 * Need to serialize access to snapshot code per filesystem.
232 	 */
233 	/*
234 	 * Assign a snapshot slot in the superblock.
235 	 */
236 	UFS_LOCK(ump);
237 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++)
238 		if (fs->fs_snapinum[snaploc] == 0)
239 			break;
240 	UFS_UNLOCK(ump);
241 	if (snaploc == FSMAXSNAP)
242 		return (ENOSPC);
243 	/*
244 	 * Create the snapshot file.
245 	 */
246 restart:
247 	NDINIT(&nd, CREATE, LOCKPARENT | LOCKLEAF, UIO_SYSSPACE, snapfile, td);
248 	if ((error = namei(&nd)) != 0)
249 		return (error);
250 	if (nd.ni_vp != NULL) {
251 		vput(nd.ni_vp);
252 		error = EEXIST;
253 	}
254 	if (nd.ni_dvp->v_mount != mp)
255 		error = EXDEV;
256 	if (error) {
257 		NDFREE(&nd, NDF_ONLY_PNBUF);
258 		if (nd.ni_dvp == nd.ni_vp)
259 			vrele(nd.ni_dvp);
260 		else
261 			vput(nd.ni_dvp);
262 		return (error);
263 	}
264 	VATTR_NULL(&vat);
265 	vat.va_type = VREG;
266 	vat.va_mode = S_IRUSR;
267 	vat.va_vaflags |= VA_EXCLUSIVE;
268 	if (VOP_GETWRITEMOUNT(nd.ni_dvp, &wrtmp))
269 		wrtmp = NULL;
270 	if (wrtmp != mp)
271 		panic("ffs_snapshot: mount mismatch");
272 	vfs_rel(wrtmp);
273 	if (vn_start_write(NULL, &wrtmp, V_NOWAIT) != 0) {
274 		NDFREE(&nd, NDF_ONLY_PNBUF);
275 		vput(nd.ni_dvp);
276 		if ((error = vn_start_write(NULL, &wrtmp,
277 		    V_XSLEEP | PCATCH)) != 0)
278 			return (error);
279 		goto restart;
280 	}
281 	error = VOP_CREATE(nd.ni_dvp, &nd.ni_vp, &nd.ni_cnd, &vat);
282 	VOP_UNLOCK(nd.ni_dvp, 0);
283 	if (error) {
284 		NDFREE(&nd, NDF_ONLY_PNBUF);
285 		vn_finished_write(wrtmp);
286 		vrele(nd.ni_dvp);
287 		return (error);
288 	}
289 	vp = nd.ni_vp;
290 	vp->v_vflag |= VV_SYSTEM;
291 	ip = VTOI(vp);
292 	devvp = ip->i_devvp;
293 	/*
294 	 * Allocate and copy the last block contents so as to be able
295 	 * to set size to that of the filesystem.
296 	 */
297 	numblks = howmany(fs->fs_size, fs->fs_frag);
298 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)),
299 	    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
300 	if (error)
301 		goto out;
302 	ip->i_size = lblktosize(fs, (off_t)numblks);
303 	DIP_SET(ip, i_size, ip->i_size);
304 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
305 	error = readblock(vp, bp, numblks - 1);
306 	bawrite(bp);
307 	if (error != 0)
308 		goto out;
309 	/*
310 	 * Preallocate critical data structures so that we can copy
311 	 * them in without further allocation after we suspend all
312 	 * operations on the filesystem. We would like to just release
313 	 * the allocated buffers without writing them since they will
314 	 * be filled in below once we are ready to go, but this upsets
315 	 * the soft update code, so we go ahead and write the new buffers.
316 	 *
317 	 * Allocate all indirect blocks and mark all of them as not
318 	 * needing to be copied.
319 	 */
320 	for (blkno = NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
321 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
322 		    fs->fs_bsize, td->td_ucred, BA_METAONLY, &ibp);
323 		if (error)
324 			goto out;
325 		bawrite(ibp);
326 	}
327 	/*
328 	 * Allocate copies for the superblock and its summary information.
329 	 */
330 	error = UFS_BALLOC(vp, fs->fs_sblockloc, fs->fs_sbsize, KERNCRED,
331 	    0, &nbp);
332 	if (error)
333 		goto out;
334 	bawrite(nbp);
335 	blkno = fragstoblks(fs, fs->fs_csaddr);
336 	len = howmany(fs->fs_cssize, fs->fs_bsize);
337 	for (loc = 0; loc < len; loc++) {
338 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(blkno + loc)),
339 		    fs->fs_bsize, KERNCRED, 0, &nbp);
340 		if (error)
341 			goto out;
342 		bawrite(nbp);
343 	}
344 	/*
345 	 * Allocate all cylinder group blocks.
346 	 */
347 	for (cg = 0; cg < fs->fs_ncg; cg++) {
348 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
349 		    fs->fs_bsize, KERNCRED, 0, &nbp);
350 		if (error)
351 			goto out;
352 		bawrite(nbp);
353 		if (cg % 10 == 0)
354 			ffs_syncvnode(vp, MNT_WAIT);
355 	}
356 	/*
357 	 * Copy all the cylinder group maps. Although the
358 	 * filesystem is still active, we hope that only a few
359 	 * cylinder groups will change between now and when we
360 	 * suspend operations. Thus, we will be able to quickly
361 	 * touch up the few cylinder groups that changed during
362 	 * the suspension period.
363 	 */
364 	len = howmany(fs->fs_ncg, NBBY);
365 	space = malloc(len, M_DEVBUF, M_WAITOK|M_ZERO);
366 	UFS_LOCK(ump);
367 	fs->fs_active = space;
368 	UFS_UNLOCK(ump);
369 	for (cg = 0; cg < fs->fs_ncg; cg++) {
370 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
371 		    fs->fs_bsize, KERNCRED, 0, &nbp);
372 		if (error)
373 			goto out;
374 		error = cgaccount(cg, vp, nbp, 1);
375 		bawrite(nbp);
376 		if (cg % 10 == 0)
377 			ffs_syncvnode(vp, MNT_WAIT);
378 		if (error)
379 			goto out;
380 	}
381 	/*
382 	 * Change inode to snapshot type file.
383 	 */
384 	ip->i_flags |= SF_SNAPSHOT;
385 	DIP_SET(ip, i_flags, ip->i_flags);
386 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
387 	/*
388 	 * Ensure that the snapshot is completely on disk.
389 	 * Since we have marked it as a snapshot it is safe to
390 	 * unlock it as no process will be allowed to write to it.
391 	 */
392 	if ((error = ffs_syncvnode(vp, MNT_WAIT)) != 0)
393 		goto out;
394 	VOP_UNLOCK(vp, 0);
395 	/*
396 	 * All allocations are done, so we can now snapshot the system.
397 	 *
398 	 * Recind nice scheduling while running with the filesystem suspended.
399 	 */
400 	if (td->td_proc->p_nice > 0) {
401 		struct proc *p;
402 
403 		p = td->td_proc;
404 		PROC_LOCK(p);
405 		saved_nice = p->p_nice;
406 		sched_nice(p, 0);
407 		PROC_UNLOCK(p);
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);
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);
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) != 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);
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) == 0 &&
563 		    vat.va_nlink > 0) {
564 			VOP_UNLOCK(xvp, 0);
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);
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);
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 	 * Acquire a lock on the snapdata structure, creating it if necessary.
616 	 */
617 	sn = ffs_snapdata_acquire(devvp);
618 	/*
619 	 * Change vnode to use shared snapshot lock instead of the original
620 	 * private lock.
621 	 */
622 	vp->v_vnlock = &sn->sn_lock;
623 	lockmgr(&vp->v_lock, LK_RELEASE, NULL);
624 	xp = TAILQ_FIRST(&sn->sn_head);
625 	/*
626 	 * If this is the first snapshot on this filesystem, then we need
627 	 * to allocate the space for the list of preallocated snapshot blocks.
628 	 * This list will be refined below, but this preliminary one will
629 	 * keep us out of deadlock until the full one is ready.
630 	 */
631 	if (xp == NULL) {
632 		snapblklist = malloc(snaplistsize * sizeof(daddr_t),
633 		    M_UFSMNT, M_WAITOK);
634 		blkp = &snapblklist[1];
635 		*blkp++ = lblkno(fs, fs->fs_sblockloc);
636 		blkno = fragstoblks(fs, fs->fs_csaddr);
637 		for (cg = 0; cg < fs->fs_ncg; cg++) {
638 			if (fragstoblks(fs, cgtod(fs, cg) > blkno))
639 				break;
640 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
641 		}
642 		len = howmany(fs->fs_cssize, fs->fs_bsize);
643 		for (loc = 0; loc < len; loc++)
644 			*blkp++ = blkno + loc;
645 		for (; cg < fs->fs_ncg; cg++)
646 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
647 		snapblklist[0] = blkp - snapblklist;
648 		VI_LOCK(devvp);
649 		if (sn->sn_blklist != NULL)
650 			panic("ffs_snapshot: non-empty list");
651 		sn->sn_blklist = snapblklist;
652 		sn->sn_listsize = blkp - snapblklist;
653 		VI_UNLOCK(devvp);
654 	}
655 	/*
656 	 * Record snapshot inode. Since this is the newest snapshot,
657 	 * it must be placed at the end of the list.
658 	 */
659 	VI_LOCK(devvp);
660 	fs->fs_snapinum[snaploc] = ip->i_number;
661 	if (ip->i_nextsnap.tqe_prev != 0)
662 		panic("ffs_snapshot: %d already on list", ip->i_number);
663 	TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
664 	devvp->v_vflag |= VV_COPYONWRITE;
665 	VI_UNLOCK(devvp);
666 	ASSERT_VOP_LOCKED(vp, "ffs_snapshot vp");
667 out1:
668 	KASSERT((sn != NULL && sbp != NULL && error == 0) ||
669 		(sn == NULL && sbp == NULL && error != 0),
670 		("email phk@ and mckusick@"));
671 	/*
672 	 * Resume operation on filesystem.
673 	 */
674 	vfs_write_resume(vp->v_mount);
675 	vn_start_write(NULL, &wrtmp, V_WAIT);
676 	if (collectsnapstats && starttime.tv_sec > 0) {
677 		nanotime(&endtime);
678 		timespecsub(&endtime, &starttime);
679 		printf("%s: suspended %ld.%03ld sec, redo %ld of %d\n",
680 		    vp->v_mount->mnt_stat.f_mntonname, (long)endtime.tv_sec,
681 		    endtime.tv_nsec / 1000000, redo, fs->fs_ncg);
682 	}
683 	if (sbp == NULL)
684 		goto out;
685 	/*
686 	 * Copy allocation information from all the snapshots in
687 	 * this snapshot and then expunge them from its view.
688 	 */
689 	TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap) {
690 		if (xp == ip)
691 			break;
692 		if (xp->i_ump->um_fstype == UFS1)
693 			error = expunge_ufs1(vp, xp, fs, snapacct_ufs1,
694 			    BLK_SNAP);
695 		else
696 			error = expunge_ufs2(vp, xp, fs, snapacct_ufs2,
697 			    BLK_SNAP);
698 		if (error == 0 && xp->i_effnlink == 0) {
699 			error = ffs_freefile(ump,
700 					     copy_fs,
701 					     vp,
702 					     xp->i_number,
703 					     xp->i_mode);
704 		}
705 		if (error) {
706 			fs->fs_snapinum[snaploc] = 0;
707 			goto done;
708 		}
709 	}
710 	/*
711 	 * Allocate space for the full list of preallocated snapshot blocks.
712 	 */
713 	snapblklist = malloc(snaplistsize * sizeof(daddr_t),
714 	    M_UFSMNT, M_WAITOK);
715 	ip->i_snapblklist = &snapblklist[1];
716 	/*
717 	 * Expunge the blocks used by the snapshots from the set of
718 	 * blocks marked as used in the snapshot bitmaps. Also, collect
719 	 * the list of allocated blocks in i_snapblklist.
720 	 */
721 	if (ip->i_ump->um_fstype == UFS1)
722 		error = expunge_ufs1(vp, ip, copy_fs, mapacct_ufs1, BLK_SNAP);
723 	else
724 		error = expunge_ufs2(vp, ip, copy_fs, mapacct_ufs2, BLK_SNAP);
725 	if (error) {
726 		fs->fs_snapinum[snaploc] = 0;
727 		free(snapblklist, M_UFSMNT);
728 		goto done;
729 	}
730 	if (snaplistsize < ip->i_snapblklist - snapblklist)
731 		panic("ffs_snapshot: list too small");
732 	snaplistsize = ip->i_snapblklist - snapblklist;
733 	snapblklist[0] = snaplistsize;
734 	ip->i_snapblklist = 0;
735 	/*
736 	 * Write out the list of allocated blocks to the end of the snapshot.
737 	 */
738 	auio.uio_iov = &aiov;
739 	auio.uio_iovcnt = 1;
740 	aiov.iov_base = (void *)snapblklist;
741 	aiov.iov_len = snaplistsize * sizeof(daddr_t);
742 	auio.uio_resid = aiov.iov_len;
743 	auio.uio_offset = ip->i_size;
744 	auio.uio_segflg = UIO_SYSSPACE;
745 	auio.uio_rw = UIO_WRITE;
746 	auio.uio_td = td;
747 	if ((error = VOP_WRITE(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
748 		fs->fs_snapinum[snaploc] = 0;
749 		free(snapblklist, M_UFSMNT);
750 		goto done;
751 	}
752 	/*
753 	 * Write the superblock and its summary information
754 	 * to the snapshot.
755 	 */
756 	blkno = fragstoblks(fs, fs->fs_csaddr);
757 	len = howmany(fs->fs_cssize, fs->fs_bsize);
758 	space = copy_fs->fs_csp;
759 	for (loc = 0; loc < len; loc++) {
760 		error = bread(vp, blkno + loc, fs->fs_bsize, KERNCRED, &nbp);
761 		if (error) {
762 			brelse(nbp);
763 			fs->fs_snapinum[snaploc] = 0;
764 			free(snapblklist, M_UFSMNT);
765 			goto done;
766 		}
767 		bcopy(space, nbp->b_data, fs->fs_bsize);
768 		space = (char *)space + fs->fs_bsize;
769 		bawrite(nbp);
770 	}
771 	/*
772 	 * As this is the newest list, it is the most inclusive, so
773 	 * should replace the previous list.
774 	 */
775 	VI_LOCK(devvp);
776 	space = sn->sn_blklist;
777 	sn->sn_blklist = snapblklist;
778 	sn->sn_listsize = snaplistsize;
779 	VI_UNLOCK(devvp);
780 	if (space != NULL)
781 		free(space, M_UFSMNT);
782 	/*
783 	 * If another process is currently writing the buffer containing
784 	 * the inode for this snapshot then a deadlock can occur. Drop
785 	 * the snapshot lock until the buffer has been written.
786 	 */
787 	VREF(vp);	/* Protect against ffs_snapgone() */
788 	VOP_UNLOCK(vp, 0);
789 	(void) bread(ip->i_devvp,
790 		     fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
791 		     (int) fs->fs_bsize, NOCRED, &nbp);
792 	brelse(nbp);
793 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
794 	if (ip->i_effnlink == 0)
795 		error = ENOENT;		/* Snapshot file unlinked */
796 	else
797 		vrele(vp);		/* Drop extra reference */
798 done:
799 	free(copy_fs->fs_csp, M_UFSMNT);
800 	bawrite(sbp);
801 out:
802 	NDFREE(&nd, NDF_ONLY_PNBUF);
803 	if (saved_nice > 0) {
804 		struct proc *p;
805 
806 		p = td->td_proc;
807 		PROC_LOCK(p);
808 		sched_nice(td->td_proc, saved_nice);
809 		PROC_UNLOCK(td->td_proc);
810 	}
811 	UFS_LOCK(ump);
812 	if (fs->fs_active != 0) {
813 		free(fs->fs_active, M_DEVBUF);
814 		fs->fs_active = 0;
815 	}
816 	UFS_UNLOCK(ump);
817 	MNT_ILOCK(mp);
818 	mp->mnt_flag = (mp->mnt_flag & MNT_QUOTA) | (flag & ~MNT_QUOTA);
819 	MNT_IUNLOCK(mp);
820 	if (error)
821 		(void) ffs_truncate(vp, (off_t)0, 0, NOCRED, td);
822 	(void) ffs_syncvnode(vp, MNT_WAIT);
823 	if (error)
824 		vput(vp);
825 	else
826 		VOP_UNLOCK(vp, 0);
827 	vrele(nd.ni_dvp);
828 	vn_finished_write(wrtmp);
829 	process_deferred_inactive(mp);
830 	return (error);
831 }
832 
833 /*
834  * Copy a cylinder group map. All the unallocated blocks are marked
835  * BLK_NOCOPY so that the snapshot knows that it need not copy them
836  * if they are later written. If passno is one, then this is a first
837  * pass, so only setting needs to be done. If passno is 2, then this
838  * is a revision to a previous pass which must be undone as the
839  * replacement pass is done.
840  */
841 static int
842 cgaccount(cg, vp, nbp, passno)
843 	int cg;
844 	struct vnode *vp;
845 	struct buf *nbp;
846 	int passno;
847 {
848 	struct buf *bp, *ibp;
849 	struct inode *ip;
850 	struct cg *cgp;
851 	struct fs *fs;
852 	ufs2_daddr_t base, numblks;
853 	int error, len, loc, indiroff;
854 
855 	ip = VTOI(vp);
856 	fs = ip->i_fs;
857 	error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)),
858 		(int)fs->fs_cgsize, KERNCRED, &bp);
859 	if (error) {
860 		brelse(bp);
861 		return (error);
862 	}
863 	cgp = (struct cg *)bp->b_data;
864 	if (!cg_chkmagic(cgp)) {
865 		brelse(bp);
866 		return (EIO);
867 	}
868 	UFS_LOCK(ip->i_ump);
869 	ACTIVESET(fs, cg);
870 	/*
871 	 * Recomputation of summary information might not have been performed
872 	 * at mount time.  Sync up summary information for current cylinder
873 	 * group while data is in memory to ensure that result of background
874 	 * fsck is slightly more consistent.
875 	 */
876 	fs->fs_cs(fs, cg) = cgp->cg_cs;
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 	bap = malloc(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 	bap = malloc(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 	ufs2_daddr_t numblks, blkno, dblk;
1568 	int error, loc, last;
1569 	struct snapdata *sn;
1570 
1571 	ip = VTOI(vp);
1572 	fs = ip->i_fs;
1573 	devvp = ip->i_devvp;
1574 	/*
1575 	 * If active, delete from incore list (this snapshot may
1576 	 * already have been in the process of being deleted, so
1577 	 * would not have been active).
1578 	 *
1579 	 * Clear copy-on-write flag if last snapshot.
1580 	 */
1581 	VI_LOCK(devvp);
1582 	if (ip->i_nextsnap.tqe_prev != 0) {
1583 		sn = devvp->v_rdev->si_snapdata;
1584 		TAILQ_REMOVE(&sn->sn_head, ip, i_nextsnap);
1585 		ip->i_nextsnap.tqe_prev = 0;
1586 		VI_UNLOCK(devvp);
1587 		lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
1588 		KASSERT(vp->v_vnlock == &sn->sn_lock,
1589 			("ffs_snapremove: lost lock mutation"));
1590 		vp->v_vnlock = &vp->v_lock;
1591 		VI_LOCK(devvp);
1592 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
1593 		try_free_snapdata(devvp);
1594 	} else
1595 		VI_UNLOCK(devvp);
1596 	/*
1597 	 * Clear all BLK_NOCOPY fields. Pass any block claims to other
1598 	 * snapshots that want them (see ffs_snapblkfree below).
1599 	 */
1600 	for (blkno = 1; blkno < NDADDR; blkno++) {
1601 		dblk = DIP(ip, i_db[blkno]);
1602 		if (dblk == 0)
1603 			continue;
1604 		if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1605 			DIP_SET(ip, i_db[blkno], 0);
1606 		else if ((dblk == blkstofrags(fs, blkno) &&
1607 		     ffs_snapblkfree(fs, ip->i_devvp, dblk, fs->fs_bsize,
1608 		     ip->i_number))) {
1609 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) -
1610 			    btodb(fs->fs_bsize));
1611 			DIP_SET(ip, i_db[blkno], 0);
1612 		}
1613 	}
1614 	numblks = howmany(ip->i_size, fs->fs_bsize);
1615 	for (blkno = NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
1616 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
1617 		    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1618 		if (error)
1619 			continue;
1620 		if (fs->fs_size - blkno > NINDIR(fs))
1621 			last = NINDIR(fs);
1622 		else
1623 			last = fs->fs_size - blkno;
1624 		for (loc = 0; loc < last; loc++) {
1625 			if (ip->i_ump->um_fstype == UFS1) {
1626 				dblk = ((ufs1_daddr_t *)(ibp->b_data))[loc];
1627 				if (dblk == 0)
1628 					continue;
1629 				if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1630 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1631 				else if ((dblk == blkstofrags(fs, blkno) &&
1632 				     ffs_snapblkfree(fs, ip->i_devvp, dblk,
1633 				     fs->fs_bsize, ip->i_number))) {
1634 					ip->i_din1->di_blocks -=
1635 					    btodb(fs->fs_bsize);
1636 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1637 				}
1638 				continue;
1639 			}
1640 			dblk = ((ufs2_daddr_t *)(ibp->b_data))[loc];
1641 			if (dblk == 0)
1642 				continue;
1643 			if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1644 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1645 			else if ((dblk == blkstofrags(fs, blkno) &&
1646 			     ffs_snapblkfree(fs, ip->i_devvp, dblk,
1647 			     fs->fs_bsize, ip->i_number))) {
1648 				ip->i_din2->di_blocks -= btodb(fs->fs_bsize);
1649 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1650 			}
1651 		}
1652 		bawrite(ibp);
1653 	}
1654 	/*
1655 	 * Clear snapshot flag and drop reference.
1656 	 */
1657 	ip->i_flags &= ~SF_SNAPSHOT;
1658 	DIP_SET(ip, i_flags, ip->i_flags);
1659 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
1660 #ifdef QUOTA
1661 	/*
1662 	 * Reenable disk quotas for ex-snapshot file.
1663 	 */
1664 	if (!getinoquota(ip))
1665 		(void) chkdq(ip, DIP(ip, i_blocks), KERNCRED, FORCE);
1666 #endif
1667 }
1668 
1669 /*
1670  * Notification that a block is being freed. Return zero if the free
1671  * should be allowed to proceed. Return non-zero if the snapshot file
1672  * wants to claim the block. The block will be claimed if it is an
1673  * uncopied part of one of the snapshots. It will be freed if it is
1674  * either a BLK_NOCOPY or has already been copied in all of the snapshots.
1675  * If a fragment is being freed, then all snapshots that care about
1676  * it must make a copy since a snapshot file can only claim full sized
1677  * blocks. Note that if more than one snapshot file maps the block,
1678  * we can pick one at random to claim it. Since none of the snapshots
1679  * can change, we are assurred that they will all see the same unmodified
1680  * image. When deleting a snapshot file (see ffs_snapremove above), we
1681  * must push any of these claimed blocks to one of the other snapshots
1682  * that maps it. These claimed blocks are easily identified as they will
1683  * have a block number equal to their logical block number within the
1684  * snapshot. A copied block can never have this property because they
1685  * must always have been allocated from a BLK_NOCOPY location.
1686  */
1687 int
1688 ffs_snapblkfree(fs, devvp, bno, size, inum)
1689 	struct fs *fs;
1690 	struct vnode *devvp;
1691 	ufs2_daddr_t bno;
1692 	long size;
1693 	ino_t inum;
1694 {
1695 	struct buf *ibp, *cbp, *savedcbp = 0;
1696 	struct thread *td = curthread;
1697 	struct inode *ip;
1698 	struct vnode *vp = NULL;
1699 	ufs_lbn_t lbn;
1700 	ufs2_daddr_t blkno;
1701 	int indiroff = 0, error = 0, claimedblk = 0;
1702 	struct snapdata *sn;
1703 
1704 	lbn = fragstoblks(fs, bno);
1705 retry:
1706 	VI_LOCK(devvp);
1707 	sn = devvp->v_rdev->si_snapdata;
1708 	if (sn == NULL) {
1709 		VI_UNLOCK(devvp);
1710 		return (0);
1711 	}
1712 	if (lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
1713 	    VI_MTX(devvp)) != 0)
1714 		goto retry;
1715 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
1716 		vp = ITOV(ip);
1717 		/*
1718 		 * Lookup block being written.
1719 		 */
1720 		if (lbn < NDADDR) {
1721 			blkno = DIP(ip, i_db[lbn]);
1722 		} else {
1723 			td->td_pflags |= TDP_COWINPROGRESS;
1724 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1725 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1726 			td->td_pflags &= ~TDP_COWINPROGRESS;
1727 			if (error)
1728 				break;
1729 			indiroff = (lbn - NDADDR) % NINDIR(fs);
1730 			if (ip->i_ump->um_fstype == UFS1)
1731 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
1732 			else
1733 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
1734 		}
1735 		/*
1736 		 * Check to see if block needs to be copied.
1737 		 */
1738 		if (blkno == 0) {
1739 			/*
1740 			 * A block that we map is being freed. If it has not
1741 			 * been claimed yet, we will claim or copy it (below).
1742 			 */
1743 			claimedblk = 1;
1744 		} else if (blkno == BLK_SNAP) {
1745 			/*
1746 			 * No previous snapshot claimed the block,
1747 			 * so it will be freed and become a BLK_NOCOPY
1748 			 * (don't care) for us.
1749 			 */
1750 			if (claimedblk)
1751 				panic("snapblkfree: inconsistent block type");
1752 			if (lbn < NDADDR) {
1753 				DIP_SET(ip, i_db[lbn], BLK_NOCOPY);
1754 				ip->i_flag |= IN_CHANGE | IN_UPDATE;
1755 			} else if (ip->i_ump->um_fstype == UFS1) {
1756 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] =
1757 				    BLK_NOCOPY;
1758 				bdwrite(ibp);
1759 			} else {
1760 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] =
1761 				    BLK_NOCOPY;
1762 				bdwrite(ibp);
1763 			}
1764 			continue;
1765 		} else /* BLK_NOCOPY or default */ {
1766 			/*
1767 			 * If the snapshot has already copied the block
1768 			 * (default), or does not care about the block,
1769 			 * it is not needed.
1770 			 */
1771 			if (lbn >= NDADDR)
1772 				bqrelse(ibp);
1773 			continue;
1774 		}
1775 		/*
1776 		 * If this is a full size block, we will just grab it
1777 		 * and assign it to the snapshot inode. Otherwise we
1778 		 * will proceed to copy it. See explanation for this
1779 		 * routine as to why only a single snapshot needs to
1780 		 * claim this block.
1781 		 */
1782 		if (size == fs->fs_bsize) {
1783 #ifdef DEBUG
1784 			if (snapdebug)
1785 				printf("%s %d lbn %jd from inum %d\n",
1786 				    "Grabonremove: snapino", ip->i_number,
1787 				    (intmax_t)lbn, inum);
1788 #endif
1789 			if (lbn < NDADDR) {
1790 				DIP_SET(ip, i_db[lbn], bno);
1791 			} else if (ip->i_ump->um_fstype == UFS1) {
1792 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] = bno;
1793 				bdwrite(ibp);
1794 			} else {
1795 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] = bno;
1796 				bdwrite(ibp);
1797 			}
1798 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + btodb(size));
1799 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1800 			lockmgr(vp->v_vnlock, LK_RELEASE, NULL);
1801 			return (1);
1802 		}
1803 		if (lbn >= NDADDR)
1804 			bqrelse(ibp);
1805 		/*
1806 		 * Allocate the block into which to do the copy. Note that this
1807 		 * allocation will never require any additional allocations for
1808 		 * the snapshot inode.
1809 		 */
1810 		td->td_pflags |= TDP_COWINPROGRESS;
1811 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1812 		    fs->fs_bsize, KERNCRED, 0, &cbp);
1813 		td->td_pflags &= ~TDP_COWINPROGRESS;
1814 		if (error)
1815 			break;
1816 #ifdef DEBUG
1817 		if (snapdebug)
1818 			printf("%s%d lbn %jd %s %d size %ld to blkno %jd\n",
1819 			    "Copyonremove: snapino ", ip->i_number,
1820 			    (intmax_t)lbn, "for inum", inum, size,
1821 			    (intmax_t)cbp->b_blkno);
1822 #endif
1823 		/*
1824 		 * If we have already read the old block contents, then
1825 		 * simply copy them to the new block. Note that we need
1826 		 * to synchronously write snapshots that have not been
1827 		 * unlinked, and hence will be visible after a crash,
1828 		 * to ensure their integrity.
1829 		 */
1830 		if (savedcbp != 0) {
1831 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
1832 			bawrite(cbp);
1833 			if (dopersistence && ip->i_effnlink > 0)
1834 				(void) ffs_syncvnode(vp, MNT_WAIT);
1835 			continue;
1836 		}
1837 		/*
1838 		 * Otherwise, read the old block contents into the buffer.
1839 		 */
1840 		if ((error = readblock(vp, cbp, lbn)) != 0) {
1841 			bzero(cbp->b_data, fs->fs_bsize);
1842 			bawrite(cbp);
1843 			if (dopersistence && ip->i_effnlink > 0)
1844 				(void) ffs_syncvnode(vp, MNT_WAIT);
1845 			break;
1846 		}
1847 		savedcbp = cbp;
1848 	}
1849 	/*
1850 	 * Note that we need to synchronously write snapshots that
1851 	 * have not been unlinked, and hence will be visible after
1852 	 * a crash, to ensure their integrity.
1853 	 */
1854 	if (savedcbp) {
1855 		vp = savedcbp->b_vp;
1856 		bawrite(savedcbp);
1857 		if (dopersistence && VTOI(vp)->i_effnlink > 0)
1858 			(void) ffs_syncvnode(vp, MNT_WAIT);
1859 	}
1860 	/*
1861 	 * If we have been unable to allocate a block in which to do
1862 	 * the copy, then return non-zero so that the fragment will
1863 	 * not be freed. Although space will be lost, the snapshot
1864 	 * will stay consistent.
1865 	 */
1866 	lockmgr(vp->v_vnlock, LK_RELEASE, NULL);
1867 	return (error);
1868 }
1869 
1870 /*
1871  * Associate snapshot files when mounting.
1872  */
1873 void
1874 ffs_snapshot_mount(mp)
1875 	struct mount *mp;
1876 {
1877 	struct ufsmount *ump = VFSTOUFS(mp);
1878 	struct vnode *devvp = ump->um_devvp;
1879 	struct fs *fs = ump->um_fs;
1880 	struct thread *td = curthread;
1881 	struct snapdata *sn;
1882 	struct vnode *vp;
1883 	struct vnode *lastvp;
1884 	struct inode *ip;
1885 	struct uio auio;
1886 	struct iovec aiov;
1887 	void *snapblklist;
1888 	char *reason;
1889 	daddr_t snaplistsize;
1890 	int error, snaploc, loc;
1891 
1892 	/*
1893 	 * XXX The following needs to be set before ffs_truncate or
1894 	 * VOP_READ can be called.
1895 	 */
1896 	mp->mnt_stat.f_iosize = fs->fs_bsize;
1897 	/*
1898 	 * Process each snapshot listed in the superblock.
1899 	 */
1900 	vp = NULL;
1901 	lastvp = NULL;
1902 	sn = NULL;
1903 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++) {
1904 		if (fs->fs_snapinum[snaploc] == 0)
1905 			break;
1906 		if ((error = ffs_vget(mp, fs->fs_snapinum[snaploc],
1907 		    LK_EXCLUSIVE, &vp)) != 0){
1908 			printf("ffs_snapshot_mount: vget failed %d\n", error);
1909 			continue;
1910 		}
1911 		ip = VTOI(vp);
1912 		if ((ip->i_flags & SF_SNAPSHOT) == 0 || ip->i_size ==
1913 		    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag))) {
1914 			if ((ip->i_flags & SF_SNAPSHOT) == 0) {
1915 				reason = "non-snapshot";
1916 			} else {
1917 				reason = "old format snapshot";
1918 				(void)ffs_truncate(vp, (off_t)0, 0, NOCRED, td);
1919 				(void)ffs_syncvnode(vp, MNT_WAIT);
1920 			}
1921 			printf("ffs_snapshot_mount: %s inode %d\n",
1922 			    reason, fs->fs_snapinum[snaploc]);
1923 			vput(vp);
1924 			vp = NULL;
1925 			for (loc = snaploc + 1; loc < FSMAXSNAP; loc++) {
1926 				if (fs->fs_snapinum[loc] == 0)
1927 					break;
1928 				fs->fs_snapinum[loc - 1] = fs->fs_snapinum[loc];
1929 			}
1930 			fs->fs_snapinum[loc - 1] = 0;
1931 			snaploc--;
1932 			continue;
1933 		}
1934 		/*
1935 		 * Acquire a lock on the snapdata structure, creating it if
1936 		 * necessary.
1937 		 */
1938 		sn = ffs_snapdata_acquire(devvp);
1939 		/*
1940 		 * Change vnode to use shared snapshot lock instead of the
1941 		 * original private lock.
1942 		 */
1943 		vp->v_vnlock = &sn->sn_lock;
1944 		lockmgr(&vp->v_lock, LK_RELEASE, NULL);
1945 		/*
1946 		 * Link it onto the active snapshot list.
1947 		 */
1948 		VI_LOCK(devvp);
1949 		if (ip->i_nextsnap.tqe_prev != 0)
1950 			panic("ffs_snapshot_mount: %d already on list",
1951 			    ip->i_number);
1952 		else
1953 			TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
1954 		vp->v_vflag |= VV_SYSTEM;
1955 		VI_UNLOCK(devvp);
1956 		VOP_UNLOCK(vp, 0);
1957 		lastvp = vp;
1958 	}
1959 	vp = lastvp;
1960 	/*
1961 	 * No usable snapshots found.
1962 	 */
1963 	if (sn == NULL || vp == NULL)
1964 		return;
1965 	/*
1966 	 * Allocate the space for the block hints list. We always want to
1967 	 * use the list from the newest snapshot.
1968 	 */
1969 	auio.uio_iov = &aiov;
1970 	auio.uio_iovcnt = 1;
1971 	aiov.iov_base = (void *)&snaplistsize;
1972 	aiov.iov_len = sizeof(snaplistsize);
1973 	auio.uio_resid = aiov.iov_len;
1974 	auio.uio_offset =
1975 	    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag));
1976 	auio.uio_segflg = UIO_SYSSPACE;
1977 	auio.uio_rw = UIO_READ;
1978 	auio.uio_td = td;
1979 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1980 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
1981 		printf("ffs_snapshot_mount: read_1 failed %d\n", error);
1982 		VOP_UNLOCK(vp, 0);
1983 		return;
1984 	}
1985 	snapblklist = malloc(snaplistsize * sizeof(daddr_t),
1986 	    M_UFSMNT, M_WAITOK);
1987 	auio.uio_iovcnt = 1;
1988 	aiov.iov_base = snapblklist;
1989 	aiov.iov_len = snaplistsize * sizeof (daddr_t);
1990 	auio.uio_resid = aiov.iov_len;
1991 	auio.uio_offset -= sizeof(snaplistsize);
1992 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
1993 		printf("ffs_snapshot_mount: read_2 failed %d\n", error);
1994 		VOP_UNLOCK(vp, 0);
1995 		free(snapblklist, M_UFSMNT);
1996 		return;
1997 	}
1998 	VOP_UNLOCK(vp, 0);
1999 	VI_LOCK(devvp);
2000 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_mount");
2001 	sn->sn_listsize = snaplistsize;
2002 	sn->sn_blklist = (daddr_t *)snapblklist;
2003 	devvp->v_vflag |= VV_COPYONWRITE;
2004 	VI_UNLOCK(devvp);
2005 }
2006 
2007 /*
2008  * Disassociate snapshot files when unmounting.
2009  */
2010 void
2011 ffs_snapshot_unmount(mp)
2012 	struct mount *mp;
2013 {
2014 	struct vnode *devvp = VFSTOUFS(mp)->um_devvp;
2015 	struct snapdata *sn;
2016 	struct inode *xp;
2017 	struct vnode *vp;
2018 
2019 	VI_LOCK(devvp);
2020 	sn = devvp->v_rdev->si_snapdata;
2021 	while (sn != NULL && (xp = TAILQ_FIRST(&sn->sn_head)) != NULL) {
2022 		vp = ITOV(xp);
2023 		TAILQ_REMOVE(&sn->sn_head, xp, i_nextsnap);
2024 		xp->i_nextsnap.tqe_prev = 0;
2025 		lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE,
2026 		    VI_MTX(devvp));
2027 		lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
2028 		KASSERT(vp->v_vnlock == &sn->sn_lock,
2029 		("ffs_snapshot_unmount: lost lock mutation"));
2030 		vp->v_vnlock = &vp->v_lock;
2031 		lockmgr(&vp->v_lock, LK_RELEASE, NULL);
2032 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
2033 		if (xp->i_effnlink > 0)
2034 			vrele(vp);
2035 		VI_LOCK(devvp);
2036 		sn = devvp->v_rdev->si_snapdata;
2037 	}
2038 	try_free_snapdata(devvp);
2039 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_unmount");
2040 }
2041 
2042 /*
2043  * Check the buffer block to be belong to device buffer that shall be
2044  * locked after snaplk. devvp shall be locked on entry, and will be
2045  * leaved locked upon exit.
2046  */
2047 static int
2048 ffs_bp_snapblk(devvp, bp)
2049 	struct vnode *devvp;
2050 	struct buf *bp;
2051 {
2052 	struct snapdata *sn;
2053 	struct fs *fs;
2054 	ufs2_daddr_t lbn, *snapblklist;
2055 	int lower, upper, mid;
2056 
2057 	ASSERT_VI_LOCKED(devvp, "ffs_bp_snapblk");
2058 	KASSERT(devvp->v_type == VCHR, ("Not a device %p", devvp));
2059 	sn = devvp->v_rdev->si_snapdata;
2060 	if (sn == NULL || TAILQ_FIRST(&sn->sn_head) == NULL)
2061 		return (0);
2062 	fs = TAILQ_FIRST(&sn->sn_head)->i_fs;
2063 	lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno));
2064 	snapblklist = sn->sn_blklist;
2065 	upper = sn->sn_listsize - 1;
2066 	lower = 1;
2067 	while (lower <= upper) {
2068 		mid = (lower + upper) / 2;
2069 		if (snapblklist[mid] == lbn)
2070 			break;
2071 		if (snapblklist[mid] < lbn)
2072 			lower = mid + 1;
2073 		else
2074 			upper = mid - 1;
2075 	}
2076 	if (lower <= upper)
2077 		return (1);
2078 	return (0);
2079 }
2080 
2081 void
2082 ffs_bdflush(bo, bp)
2083 	struct bufobj *bo;
2084 	struct buf *bp;
2085 {
2086 	struct thread *td;
2087 	struct vnode *vp, *devvp;
2088 	struct buf *nbp;
2089 	int bp_bdskip;
2090 
2091 	if (bo->bo_dirty.bv_cnt <= dirtybufthresh)
2092 		return;
2093 
2094 	td = curthread;
2095 	vp = bp->b_vp;
2096 	devvp = bo->__bo_vnode;
2097 	KASSERT(vp == devvp, ("devvp != vp %p %p", bo, bp));
2098 
2099 	VI_LOCK(devvp);
2100 	bp_bdskip = ffs_bp_snapblk(devvp, bp);
2101 	if (bp_bdskip)
2102 		bdwriteskip++;
2103 	VI_UNLOCK(devvp);
2104 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10 && !bp_bdskip) {
2105 		(void) VOP_FSYNC(vp, MNT_NOWAIT, td);
2106 		altbufferflushes++;
2107 	} else {
2108 		BO_LOCK(bo);
2109 		/*
2110 		 * Try to find a buffer to flush.
2111 		 */
2112 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2113 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2114 			    BUF_LOCK(nbp,
2115 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2116 				continue;
2117 			if (bp == nbp)
2118 				panic("bdwrite: found ourselves");
2119 			BO_UNLOCK(bo);
2120 			/*
2121 			 * Don't countdeps with the bo lock
2122 			 * held.
2123 			 */
2124 			if (buf_countdeps(nbp, 0)) {
2125 				BO_LOCK(bo);
2126 				BUF_UNLOCK(nbp);
2127 				continue;
2128 			}
2129 			if (bp_bdskip) {
2130 				VI_LOCK(devvp);
2131 				if (!ffs_bp_snapblk(vp, nbp)) {
2132 					if (BO_MTX(bo) != VI_MTX(vp)) {
2133 						VI_UNLOCK(devvp);
2134 						BO_LOCK(bo);
2135 					}
2136 					BUF_UNLOCK(nbp);
2137 					continue;
2138 				}
2139 				VI_UNLOCK(devvp);
2140 			}
2141 			if (nbp->b_flags & B_CLUSTEROK) {
2142 				vfs_bio_awrite(nbp);
2143 			} else {
2144 				bremfree(nbp);
2145 				bawrite(nbp);
2146 			}
2147 			dirtybufferflushes++;
2148 			break;
2149 		}
2150 		if (nbp == NULL)
2151 			BO_UNLOCK(bo);
2152 	}
2153 }
2154 
2155 /*
2156  * Check for need to copy block that is about to be written,
2157  * copying the block if necessary.
2158  */
2159 int
2160 ffs_copyonwrite(devvp, bp)
2161 	struct vnode *devvp;
2162 	struct buf *bp;
2163 {
2164 	struct snapdata *sn;
2165 	struct buf *ibp, *cbp, *savedcbp = 0;
2166 	struct thread *td = curthread;
2167 	struct fs *fs;
2168 	struct inode *ip;
2169 	struct vnode *vp = 0;
2170 	ufs2_daddr_t lbn, blkno, *snapblklist;
2171 	int lower, upper, mid, indiroff, error = 0;
2172 	int launched_async_io, prev_norunningbuf;
2173 	long saved_runningbufspace;
2174 
2175 	if (devvp != bp->b_vp && (VTOI(bp->b_vp)->i_flags & SF_SNAPSHOT) != 0)
2176 		return (0);		/* Update on a snapshot file */
2177 	if (td->td_pflags & TDP_COWINPROGRESS)
2178 		panic("ffs_copyonwrite: recursive call");
2179 	/*
2180 	 * First check to see if it is in the preallocated list.
2181 	 * By doing this check we avoid several potential deadlocks.
2182 	 */
2183 	VI_LOCK(devvp);
2184 	sn = devvp->v_rdev->si_snapdata;
2185 	if (sn == NULL ||
2186 	    TAILQ_EMPTY(&sn->sn_head)) {
2187 		VI_UNLOCK(devvp);
2188 		return (0);		/* No snapshot */
2189 	}
2190 	ip = TAILQ_FIRST(&sn->sn_head);
2191 	fs = ip->i_fs;
2192 	lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno));
2193 	snapblklist = sn->sn_blklist;
2194 	upper = sn->sn_listsize - 1;
2195 	lower = 1;
2196 	while (lower <= upper) {
2197 		mid = (lower + upper) / 2;
2198 		if (snapblklist[mid] == lbn)
2199 			break;
2200 		if (snapblklist[mid] < lbn)
2201 			lower = mid + 1;
2202 		else
2203 			upper = mid - 1;
2204 	}
2205 	if (lower <= upper) {
2206 		VI_UNLOCK(devvp);
2207 		return (0);
2208 	}
2209 	launched_async_io = 0;
2210 	prev_norunningbuf = td->td_pflags & TDP_NORUNNINGBUF;
2211 	/*
2212 	 * Since I/O on bp isn't yet in progress and it may be blocked
2213 	 * for a long time waiting on snaplk, back it out of
2214 	 * runningbufspace, possibly waking other threads waiting for space.
2215 	 */
2216 	saved_runningbufspace = bp->b_runningbufspace;
2217 	if (saved_runningbufspace != 0)
2218 		runningbufwakeup(bp);
2219 	/*
2220 	 * Not in the precomputed list, so check the snapshots.
2221 	 */
2222 	while (lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
2223 	    VI_MTX(devvp)) != 0) {
2224 		VI_LOCK(devvp);
2225 		sn = devvp->v_rdev->si_snapdata;
2226 		if (sn == NULL ||
2227 		    TAILQ_EMPTY(&sn->sn_head)) {
2228 			VI_UNLOCK(devvp);
2229 			if (saved_runningbufspace != 0) {
2230 				bp->b_runningbufspace = saved_runningbufspace;
2231 				atomic_add_long(&runningbufspace,
2232 					       bp->b_runningbufspace);
2233 			}
2234 			return (0);		/* Snapshot gone */
2235 		}
2236 	}
2237 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
2238 		vp = ITOV(ip);
2239 		/*
2240 		 * We ensure that everything of our own that needs to be
2241 		 * copied will be done at the time that ffs_snapshot is
2242 		 * called. Thus we can skip the check here which can
2243 		 * deadlock in doing the lookup in UFS_BALLOC.
2244 		 */
2245 		if (bp->b_vp == vp)
2246 			continue;
2247 		/*
2248 		 * Check to see if block needs to be copied. We do not have
2249 		 * to hold the snapshot lock while doing this lookup as it
2250 		 * will never require any additional allocations for the
2251 		 * snapshot inode.
2252 		 */
2253 		if (lbn < NDADDR) {
2254 			blkno = DIP(ip, i_db[lbn]);
2255 		} else {
2256 			td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2257 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2258 			   fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
2259 			td->td_pflags &= ~TDP_COWINPROGRESS;
2260 			if (error)
2261 				break;
2262 			indiroff = (lbn - NDADDR) % NINDIR(fs);
2263 			if (ip->i_ump->um_fstype == UFS1)
2264 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
2265 			else
2266 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
2267 			bqrelse(ibp);
2268 		}
2269 #ifdef INVARIANTS
2270 		if (blkno == BLK_SNAP && bp->b_lblkno >= 0)
2271 			panic("ffs_copyonwrite: bad copy block");
2272 #endif
2273 		if (blkno != 0)
2274 			continue;
2275 		/*
2276 		 * Allocate the block into which to do the copy. Since
2277 		 * multiple processes may all try to copy the same block,
2278 		 * we have to recheck our need to do a copy if we sleep
2279 		 * waiting for the lock.
2280 		 *
2281 		 * Because all snapshots on a filesystem share a single
2282 		 * lock, we ensure that we will never be in competition
2283 		 * with another process to allocate a block.
2284 		 */
2285 		td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2286 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2287 		    fs->fs_bsize, KERNCRED, 0, &cbp);
2288 		td->td_pflags &= ~TDP_COWINPROGRESS;
2289 		if (error)
2290 			break;
2291 #ifdef DEBUG
2292 		if (snapdebug) {
2293 			printf("Copyonwrite: snapino %d lbn %jd for ",
2294 			    ip->i_number, (intmax_t)lbn);
2295 			if (bp->b_vp == devvp)
2296 				printf("fs metadata");
2297 			else
2298 				printf("inum %d", VTOI(bp->b_vp)->i_number);
2299 			printf(" lblkno %jd to blkno %jd\n",
2300 			    (intmax_t)bp->b_lblkno, (intmax_t)cbp->b_blkno);
2301 		}
2302 #endif
2303 		/*
2304 		 * If we have already read the old block contents, then
2305 		 * simply copy them to the new block. Note that we need
2306 		 * to synchronously write snapshots that have not been
2307 		 * unlinked, and hence will be visible after a crash,
2308 		 * to ensure their integrity.
2309 		 */
2310 		if (savedcbp != 0) {
2311 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
2312 			bawrite(cbp);
2313 			if (dopersistence && ip->i_effnlink > 0)
2314 				(void) ffs_syncvnode(vp, MNT_WAIT);
2315 			else
2316 				launched_async_io = 1;
2317 			continue;
2318 		}
2319 		/*
2320 		 * Otherwise, read the old block contents into the buffer.
2321 		 */
2322 		if ((error = readblock(vp, cbp, lbn)) != 0) {
2323 			bzero(cbp->b_data, fs->fs_bsize);
2324 			bawrite(cbp);
2325 			if (dopersistence && ip->i_effnlink > 0)
2326 				(void) ffs_syncvnode(vp, MNT_WAIT);
2327 			else
2328 				launched_async_io = 1;
2329 			break;
2330 		}
2331 		savedcbp = cbp;
2332 	}
2333 	/*
2334 	 * Note that we need to synchronously write snapshots that
2335 	 * have not been unlinked, and hence will be visible after
2336 	 * a crash, to ensure their integrity.
2337 	 */
2338 	if (savedcbp) {
2339 		vp = savedcbp->b_vp;
2340 		bawrite(savedcbp);
2341 		if (dopersistence && VTOI(vp)->i_effnlink > 0)
2342 			(void) ffs_syncvnode(vp, MNT_WAIT);
2343 		else
2344 			launched_async_io = 1;
2345 	}
2346 	lockmgr(vp->v_vnlock, LK_RELEASE, NULL);
2347 	td->td_pflags = (td->td_pflags & ~TDP_NORUNNINGBUF) |
2348 		prev_norunningbuf;
2349 	if (launched_async_io && (td->td_pflags & TDP_NORUNNINGBUF) == 0)
2350 		waitrunningbufspace();
2351 	/*
2352 	 * I/O on bp will now be started, so count it in runningbufspace.
2353 	 */
2354 	if (saved_runningbufspace != 0) {
2355 		bp->b_runningbufspace = saved_runningbufspace;
2356 		atomic_add_long(&runningbufspace, bp->b_runningbufspace);
2357 	}
2358 	return (error);
2359 }
2360 
2361 /*
2362  * Read the specified block into the given buffer.
2363  * Much of this boiler-plate comes from bwrite().
2364  */
2365 static int
2366 readblock(vp, bp, lbn)
2367 	struct vnode *vp;
2368 	struct buf *bp;
2369 	ufs2_daddr_t lbn;
2370 {
2371 	struct inode *ip = VTOI(vp);
2372 	struct bio *bip;
2373 
2374 	bip = g_alloc_bio();
2375 	bip->bio_cmd = BIO_READ;
2376 	bip->bio_offset = dbtob(fsbtodb(ip->i_fs, blkstofrags(ip->i_fs, lbn)));
2377 	bip->bio_data = bp->b_data;
2378 	bip->bio_length = bp->b_bcount;
2379 	bip->bio_done = NULL;
2380 
2381 	g_io_request(bip, ip->i_devvp->v_bufobj.bo_private);
2382 	bp->b_error = biowait(bip, "snaprdb");
2383 	g_destroy_bio(bip);
2384 	return (bp->b_error);
2385 }
2386 
2387 #endif
2388 
2389 /*
2390  * Process file deletes that were deferred by ufs_inactive() due to
2391  * the file system being suspended. Transfer IN_LAZYACCESS into
2392  * IN_MODIFIED for vnodes that were accessed during suspension.
2393  */
2394 void
2395 process_deferred_inactive(struct mount *mp)
2396 {
2397 	struct vnode *vp, *mvp;
2398 	struct inode *ip;
2399 	struct thread *td;
2400 	int error;
2401 
2402 	td = curthread;
2403 	(void) vn_start_secondary_write(NULL, &mp, V_WAIT);
2404 	MNT_ILOCK(mp);
2405  loop:
2406 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2407 		VI_LOCK(vp);
2408 		/*
2409 		 * IN_LAZYACCESS is checked here without holding any
2410 		 * vnode lock, but this flag is set only while holding
2411 		 * vnode interlock.
2412 		 */
2413 		if (vp->v_type == VNON || (vp->v_iflag & VI_DOOMED) != 0 ||
2414 		    ((VTOI(vp)->i_flag & IN_LAZYACCESS) == 0 &&
2415 			((vp->v_iflag & VI_OWEINACT) == 0 ||
2416 			vp->v_usecount > 0))) {
2417 			VI_UNLOCK(vp);
2418 			continue;
2419 		}
2420 		MNT_IUNLOCK(mp);
2421 		vholdl(vp);
2422 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2423 		if (error != 0) {
2424 			vdrop(vp);
2425 			MNT_ILOCK(mp);
2426 			if (error == ENOENT)
2427 				continue;	/* vnode recycled */
2428 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2429 			goto loop;
2430 		}
2431 		ip = VTOI(vp);
2432 		if ((ip->i_flag & IN_LAZYACCESS) != 0) {
2433 			ip->i_flag &= ~IN_LAZYACCESS;
2434 			ip->i_flag |= IN_MODIFIED;
2435 		}
2436 		VI_LOCK(vp);
2437 		if ((vp->v_iflag & VI_OWEINACT) == 0 || vp->v_usecount > 0) {
2438 			VI_UNLOCK(vp);
2439 			VOP_UNLOCK(vp, 0);
2440 			vdrop(vp);
2441 			MNT_ILOCK(mp);
2442 			continue;
2443 		}
2444 
2445 		VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2446 			 ("process_deferred_inactive: "
2447 			  "recursed on VI_DOINGINACT"));
2448 		vp->v_iflag |= VI_DOINGINACT;
2449 		vp->v_iflag &= ~VI_OWEINACT;
2450 		VI_UNLOCK(vp);
2451 		(void) VOP_INACTIVE(vp, td);
2452 		VI_LOCK(vp);
2453 		VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2454 			 ("process_deferred_inactive: lost VI_DOINGINACT"));
2455 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2456 			 ("process_deferred_inactive: got VI_OWEINACT"));
2457 		vp->v_iflag &= ~VI_DOINGINACT;
2458 		VI_UNLOCK(vp);
2459 		VOP_UNLOCK(vp, 0);
2460 		vdrop(vp);
2461 		MNT_ILOCK(mp);
2462 	}
2463 	MNT_IUNLOCK(mp);
2464 	vn_finished_secondary_write(mp);
2465 }
2466 
2467 #ifndef NO_FFS_SNAPSHOT
2468 
2469 static struct snapdata *
2470 ffs_snapdata_alloc(void)
2471 {
2472 	struct snapdata *sn;
2473 
2474 	/*
2475 	 * Fetch a snapdata from the free list if there is one available.
2476 	 */
2477 	mtx_lock(&snapfree_lock);
2478 	sn = LIST_FIRST(&snapfree);
2479 	if (sn != NULL)
2480 		LIST_REMOVE(sn, sn_link);
2481 	mtx_unlock(&snapfree_lock);
2482 	if (sn != NULL)
2483 		return (sn);
2484 	/*
2485  	 * If there were no free snapdatas allocate one.
2486 	 */
2487 	sn = malloc(sizeof *sn, M_UFSMNT, M_WAITOK | M_ZERO);
2488 	TAILQ_INIT(&sn->sn_head);
2489 	lockinit(&sn->sn_lock, PVFS, "snaplk", VLKTIMEOUT,
2490 	    LK_CANRECURSE | LK_NOSHARE);
2491 	return (sn);
2492 }
2493 
2494 /*
2495  * The snapdata is never freed because we can not be certain that
2496  * there are no threads sleeping on the snap lock.  Persisting
2497  * them permanently avoids costly synchronization in ffs_lock().
2498  */
2499 static void
2500 ffs_snapdata_free(struct snapdata *sn)
2501 {
2502 	mtx_lock(&snapfree_lock);
2503 	LIST_INSERT_HEAD(&snapfree, sn, sn_link);
2504 	mtx_unlock(&snapfree_lock);
2505 }
2506 
2507 /* Try to free snapdata associated with devvp */
2508 static void
2509 try_free_snapdata(struct vnode *devvp)
2510 {
2511 	struct snapdata *sn;
2512 	ufs2_daddr_t *snapblklist;
2513 
2514 	ASSERT_VI_LOCKED(devvp, "try_free_snapdata");
2515 	sn = devvp->v_rdev->si_snapdata;
2516 
2517 	if (sn == NULL || TAILQ_FIRST(&sn->sn_head) != NULL ||
2518 	    (devvp->v_vflag & VV_COPYONWRITE) == 0) {
2519 		VI_UNLOCK(devvp);
2520 		return;
2521 	}
2522 
2523 	devvp->v_rdev->si_snapdata = NULL;
2524 	devvp->v_vflag &= ~VV_COPYONWRITE;
2525 	lockmgr(&sn->sn_lock, LK_DRAIN|LK_INTERLOCK, VI_MTX(devvp));
2526 	snapblklist = sn->sn_blklist;
2527 	sn->sn_blklist = NULL;
2528 	sn->sn_listsize = 0;
2529 	lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
2530 	if (snapblklist != NULL)
2531 		free(snapblklist, M_UFSMNT);
2532 	ffs_snapdata_free(sn);
2533 }
2534 
2535 static struct snapdata *
2536 ffs_snapdata_acquire(struct vnode *devvp)
2537 {
2538 	struct snapdata *nsn;
2539 	struct snapdata *sn;
2540 
2541 	/*
2542  	 * Allocate a free snapdata.  This is done before acquiring the
2543 	 * devvp lock to avoid allocation while the devvp interlock is
2544 	 * held.
2545 	 */
2546 	nsn = ffs_snapdata_alloc();
2547 	/*
2548 	 * If there snapshots already exist on this filesystem grab a
2549 	 * reference to the shared lock.  Otherwise this is the first
2550 	 * snapshot on this filesystem and we need to use our
2551 	 * pre-allocated snapdata.
2552 	 */
2553 	VI_LOCK(devvp);
2554 	if (devvp->v_rdev->si_snapdata == NULL) {
2555 		devvp->v_rdev->si_snapdata = nsn;
2556 		nsn = NULL;
2557 	}
2558 	sn = devvp->v_rdev->si_snapdata;
2559 	/*
2560 	 * Acquire the snapshot lock.
2561 	 */
2562 	lockmgr(&sn->sn_lock,
2563 	    LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY, VI_MTX(devvp));
2564 	/*
2565 	 * Free any unused snapdata.
2566 	 */
2567 	if (nsn != NULL)
2568 		ffs_snapdata_free(nsn);
2569 
2570 	return (sn);
2571 }
2572 
2573 #endif
2574