xref: /freebsd/sys/ufs/ffs/ffs_snapshot.c (revision db612abe8df3355d1eb23bb3b50fdd97bc21e979)
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 process_deferred_inactive(struct mount *);
171 static void try_free_snapdata(struct vnode *devvp);
172 static struct snapdata *ffs_snapdata_acquire(struct vnode *devvp);
173 static int ffs_bp_snapblk(struct vnode *, struct buf *);
174 
175 /*
176  * To ensure the consistency of snapshots across crashes, we must
177  * synchronously write out copied blocks before allowing the
178  * originals to be modified. Because of the rather severe speed
179  * penalty that this imposes, the following flag allows this
180  * crash persistence to be disabled.
181  */
182 int dopersistence = 0;
183 
184 #ifdef DEBUG
185 #include <sys/sysctl.h>
186 SYSCTL_INT(_debug, OID_AUTO, dopersistence, CTLFLAG_RW, &dopersistence, 0, "");
187 static int snapdebug = 0;
188 SYSCTL_INT(_debug, OID_AUTO, snapdebug, CTLFLAG_RW, &snapdebug, 0, "");
189 int collectsnapstats = 0;
190 SYSCTL_INT(_debug, OID_AUTO, collectsnapstats, CTLFLAG_RW, &collectsnapstats,
191 	0, "");
192 #endif /* DEBUG */
193 
194 /*
195  * Create a snapshot file and initialize it for the filesystem.
196  */
197 int
198 ffs_snapshot(mp, snapfile)
199 	struct mount *mp;
200 	char *snapfile;
201 {
202 	ufs2_daddr_t numblks, blkno, *blkp, *snapblklist;
203 	int error, cg, snaploc;
204 	int i, size, len, loc;
205 	int flag;
206 	struct timespec starttime = {0, 0}, endtime;
207 	char saved_nice = 0;
208 	long redo = 0, snaplistsize = 0;
209 	int32_t *lp;
210 	void *space;
211 	struct fs *copy_fs = NULL, *fs;
212 	struct thread *td = curthread;
213 	struct inode *ip, *xp;
214 	struct buf *bp, *nbp, *ibp, *sbp = NULL;
215 	struct nameidata nd;
216 	struct mount *wrtmp;
217 	struct vattr vat;
218 	struct vnode *vp, *xvp, *mvp, *devvp;
219 	struct uio auio;
220 	struct iovec aiov;
221 	struct snapdata *sn;
222 	struct ufsmount *ump;
223 
224 	ump = VFSTOUFS(mp);
225 	fs = ump->um_fs;
226 	sn = NULL;
227 	MNT_ILOCK(mp);
228 	flag = mp->mnt_flag;
229 	MNT_IUNLOCK(mp);
230 
231 	/*
232 	 * Need to serialize access to snapshot code per filesystem.
233 	 */
234 	/*
235 	 * Assign a snapshot slot in the superblock.
236 	 */
237 	UFS_LOCK(ump);
238 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++)
239 		if (fs->fs_snapinum[snaploc] == 0)
240 			break;
241 	UFS_UNLOCK(ump);
242 	if (snaploc == FSMAXSNAP)
243 		return (ENOSPC);
244 	/*
245 	 * Create the snapshot file.
246 	 */
247 restart:
248 	NDINIT(&nd, CREATE, LOCKPARENT | LOCKLEAF, UIO_SYSSPACE, snapfile, td);
249 	if ((error = namei(&nd)) != 0)
250 		return (error);
251 	if (nd.ni_vp != NULL) {
252 		vput(nd.ni_vp);
253 		error = EEXIST;
254 	}
255 	if (nd.ni_dvp->v_mount != mp)
256 		error = EXDEV;
257 	if (error) {
258 		NDFREE(&nd, NDF_ONLY_PNBUF);
259 		if (nd.ni_dvp == nd.ni_vp)
260 			vrele(nd.ni_dvp);
261 		else
262 			vput(nd.ni_dvp);
263 		return (error);
264 	}
265 	VATTR_NULL(&vat);
266 	vat.va_type = VREG;
267 	vat.va_mode = S_IRUSR;
268 	vat.va_vaflags |= VA_EXCLUSIVE;
269 	if (VOP_GETWRITEMOUNT(nd.ni_dvp, &wrtmp))
270 		wrtmp = NULL;
271 	if (wrtmp != mp)
272 		panic("ffs_snapshot: mount mismatch");
273 	vfs_rel(wrtmp);
274 	if (vn_start_write(NULL, &wrtmp, V_NOWAIT) != 0) {
275 		NDFREE(&nd, NDF_ONLY_PNBUF);
276 		vput(nd.ni_dvp);
277 		if ((error = vn_start_write(NULL, &wrtmp,
278 		    V_XSLEEP | PCATCH)) != 0)
279 			return (error);
280 		goto restart;
281 	}
282 	VOP_LEASE(nd.ni_dvp, td, KERNCRED, LEASE_WRITE);
283 	error = VOP_CREATE(nd.ni_dvp, &nd.ni_vp, &nd.ni_cnd, &vat);
284 	VOP_UNLOCK(nd.ni_dvp, 0);
285 	if (error) {
286 		NDFREE(&nd, NDF_ONLY_PNBUF);
287 		vn_finished_write(wrtmp);
288 		vrele(nd.ni_dvp);
289 		return (error);
290 	}
291 	vp = nd.ni_vp;
292 	vp->v_vflag |= VV_SYSTEM;
293 	ip = VTOI(vp);
294 	devvp = ip->i_devvp;
295 	/*
296 	 * Allocate and copy the last block contents so as to be able
297 	 * to set size to that of the filesystem.
298 	 */
299 	numblks = howmany(fs->fs_size, fs->fs_frag);
300 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)),
301 	    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
302 	if (error)
303 		goto out;
304 	ip->i_size = lblktosize(fs, (off_t)numblks);
305 	DIP_SET(ip, i_size, ip->i_size);
306 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
307 	error = readblock(vp, bp, numblks - 1);
308 	bawrite(bp);
309 	if (error != 0)
310 		goto out;
311 	/*
312 	 * Preallocate critical data structures so that we can copy
313 	 * them in without further allocation after we suspend all
314 	 * operations on the filesystem. We would like to just release
315 	 * the allocated buffers without writing them since they will
316 	 * be filled in below once we are ready to go, but this upsets
317 	 * the soft update code, so we go ahead and write the new buffers.
318 	 *
319 	 * Allocate all indirect blocks and mark all of them as not
320 	 * needing to be copied.
321 	 */
322 	for (blkno = NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
323 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
324 		    fs->fs_bsize, td->td_ucred, BA_METAONLY, &ibp);
325 		if (error)
326 			goto out;
327 		bawrite(ibp);
328 	}
329 	/*
330 	 * Allocate copies for the superblock and its summary information.
331 	 */
332 	error = UFS_BALLOC(vp, fs->fs_sblockloc, fs->fs_sbsize, KERNCRED,
333 	    0, &nbp);
334 	if (error)
335 		goto out;
336 	bawrite(nbp);
337 	blkno = fragstoblks(fs, fs->fs_csaddr);
338 	len = howmany(fs->fs_cssize, fs->fs_bsize);
339 	for (loc = 0; loc < len; loc++) {
340 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(blkno + loc)),
341 		    fs->fs_bsize, KERNCRED, 0, &nbp);
342 		if (error)
343 			goto out;
344 		bawrite(nbp);
345 	}
346 	/*
347 	 * Allocate all cylinder group blocks.
348 	 */
349 	for (cg = 0; cg < fs->fs_ncg; cg++) {
350 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
351 		    fs->fs_bsize, KERNCRED, 0, &nbp);
352 		if (error)
353 			goto out;
354 		bawrite(nbp);
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 	MALLOC(space, void *, 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 (error)
377 			goto out;
378 	}
379 	/*
380 	 * Change inode to snapshot type file.
381 	 */
382 	ip->i_flags |= SF_SNAPSHOT;
383 	DIP_SET(ip, i_flags, ip->i_flags);
384 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
385 	/*
386 	 * Ensure that the snapshot is completely on disk.
387 	 * Since we have marked it as a snapshot it is safe to
388 	 * unlock it as no process will be allowed to write to it.
389 	 */
390 	if ((error = ffs_syncvnode(vp, MNT_WAIT)) != 0)
391 		goto out;
392 	VOP_UNLOCK(vp, 0);
393 	/*
394 	 * All allocations are done, so we can now snapshot the system.
395 	 *
396 	 * Recind nice scheduling while running with the filesystem suspended.
397 	 */
398 	if (td->td_proc->p_nice > 0) {
399 		struct proc *p;
400 
401 		p = td->td_proc;
402 		PROC_LOCK(p);
403 		saved_nice = p->p_nice;
404 		sched_nice(p, 0);
405 		PROC_UNLOCK(p);
406 	}
407 	/*
408 	 * Suspend operation on filesystem.
409 	 */
410 	for (;;) {
411 		vn_finished_write(wrtmp);
412 		if ((error = vfs_write_suspend(vp->v_mount)) != 0) {
413 			vn_start_write(NULL, &wrtmp, V_WAIT);
414 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
415 			goto out;
416 		}
417 		if (mp->mnt_kern_flag & MNTK_SUSPENDED)
418 			break;
419 		vn_start_write(NULL, &wrtmp, V_WAIT);
420 	}
421 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
422 	if (ip->i_effnlink == 0) {
423 		error = ENOENT;		/* Snapshot file unlinked */
424 		goto out1;
425 	}
426 	if (collectsnapstats)
427 		nanotime(&starttime);
428 
429 	/* The last block might have changed.  Copy it again to be sure. */
430 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)),
431 	    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
432 	if (error != 0)
433 		goto out1;
434 	error = readblock(vp, bp, numblks - 1);
435 	bp->b_flags |= B_VALIDSUSPWRT;
436 	bawrite(bp);
437 	if (error != 0)
438 		goto out1;
439 	/*
440 	 * First, copy all the cylinder group maps that have changed.
441 	 */
442 	for (cg = 0; cg < fs->fs_ncg; cg++) {
443 		if ((ACTIVECGNUM(fs, cg) & ACTIVECGOFF(cg)) != 0)
444 			continue;
445 		redo++;
446 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
447 		    fs->fs_bsize, KERNCRED, 0, &nbp);
448 		if (error)
449 			goto out1;
450 		error = cgaccount(cg, vp, nbp, 2);
451 		bawrite(nbp);
452 		if (error)
453 			goto out1;
454 	}
455 	/*
456 	 * Grab a copy of the superblock and its summary information.
457 	 * We delay writing it until the suspension is released below.
458 	 */
459 	error = bread(vp, lblkno(fs, fs->fs_sblockloc), fs->fs_bsize,
460 	    KERNCRED, &sbp);
461 	if (error) {
462 		brelse(sbp);
463 		sbp = NULL;
464 		goto out1;
465 	}
466 	loc = blkoff(fs, fs->fs_sblockloc);
467 	copy_fs = (struct fs *)(sbp->b_data + loc);
468 	bcopy(fs, copy_fs, fs->fs_sbsize);
469 	if ((fs->fs_flags & (FS_UNCLEAN | FS_NEEDSFSCK)) == 0)
470 		copy_fs->fs_clean = 1;
471 	size = fs->fs_bsize < SBLOCKSIZE ? fs->fs_bsize : SBLOCKSIZE;
472 	if (fs->fs_sbsize < size)
473 		bzero(&sbp->b_data[loc + fs->fs_sbsize], size - fs->fs_sbsize);
474 	size = blkroundup(fs, fs->fs_cssize);
475 	if (fs->fs_contigsumsize > 0)
476 		size += fs->fs_ncg * sizeof(int32_t);
477 	space = malloc((u_long)size, M_UFSMNT, M_WAITOK);
478 	copy_fs->fs_csp = space;
479 	bcopy(fs->fs_csp, copy_fs->fs_csp, fs->fs_cssize);
480 	space = (char *)space + fs->fs_cssize;
481 	loc = howmany(fs->fs_cssize, fs->fs_fsize);
482 	i = fs->fs_frag - loc % fs->fs_frag;
483 	len = (i == fs->fs_frag) ? 0 : i * fs->fs_fsize;
484 	if (len > 0) {
485 		if ((error = bread(devvp, fsbtodb(fs, fs->fs_csaddr + loc),
486 		    len, KERNCRED, &bp)) != 0) {
487 			brelse(bp);
488 			free(copy_fs->fs_csp, M_UFSMNT);
489 			bawrite(sbp);
490 			sbp = NULL;
491 			goto out1;
492 		}
493 		bcopy(bp->b_data, space, (u_int)len);
494 		space = (char *)space + len;
495 		bp->b_flags |= B_INVAL | B_NOCACHE;
496 		brelse(bp);
497 	}
498 	if (fs->fs_contigsumsize > 0) {
499 		copy_fs->fs_maxcluster = lp = space;
500 		for (i = 0; i < fs->fs_ncg; i++)
501 			*lp++ = fs->fs_contigsumsize;
502 	}
503 	/*
504 	 * We must check for active files that have been unlinked
505 	 * (e.g., with a zero link count). We have to expunge all
506 	 * trace of these files from the snapshot so that they are
507 	 * not reclaimed prematurely by fsck or unnecessarily dumped.
508 	 * We turn off the MNTK_SUSPENDED flag to avoid a panic from
509 	 * spec_strategy about writing on a suspended filesystem.
510 	 * Note that we skip unlinked snapshot files as they will
511 	 * be handled separately below.
512 	 *
513 	 * We also calculate the needed size for the snapshot list.
514 	 */
515 	snaplistsize = fs->fs_ncg + howmany(fs->fs_cssize, fs->fs_bsize) +
516 	    FSMAXSNAP + 1 /* superblock */ + 1 /* last block */ + 1 /* size */;
517 	MNT_ILOCK(mp);
518 	mp->mnt_kern_flag &= ~MNTK_SUSPENDED;
519 loop:
520 	MNT_VNODE_FOREACH(xvp, mp, mvp) {
521 		VI_LOCK(xvp);
522 		MNT_IUNLOCK(mp);
523 		if ((xvp->v_iflag & VI_DOOMED) ||
524 		    (xvp->v_usecount == 0 &&
525 		     (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) ||
526 		    xvp->v_type == VNON ||
527 		    (VTOI(xvp)->i_flags & SF_SNAPSHOT)) {
528 			VI_UNLOCK(xvp);
529 			MNT_ILOCK(mp);
530 			continue;
531 		}
532 		/*
533 		 * We can skip parent directory vnode because it must have
534 		 * this snapshot file in it.
535 		 */
536 		if (xvp == nd.ni_dvp) {
537 			VI_UNLOCK(xvp);
538 			MNT_ILOCK(mp);
539 			continue;
540 		}
541 		vholdl(xvp);
542 		if (vn_lock(xvp, LK_EXCLUSIVE | LK_INTERLOCK) != 0) {
543 			MNT_ILOCK(mp);
544 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
545 			vdrop(xvp);
546 			goto loop;
547 		}
548 		VI_LOCK(xvp);
549 		if (xvp->v_usecount == 0 &&
550 		    (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) {
551 			VI_UNLOCK(xvp);
552 			VOP_UNLOCK(xvp, 0);
553 			vdrop(xvp);
554 			MNT_ILOCK(mp);
555 			continue;
556 		}
557 		VI_UNLOCK(xvp);
558 		if (snapdebug)
559 			vprint("ffs_snapshot: busy vnode", xvp);
560 		if (VOP_GETATTR(xvp, &vat, td->td_ucred, td) == 0 &&
561 		    vat.va_nlink > 0) {
562 			VOP_UNLOCK(xvp, 0);
563 			vdrop(xvp);
564 			MNT_ILOCK(mp);
565 			continue;
566 		}
567 		xp = VTOI(xvp);
568 		if (ffs_checkfreefile(copy_fs, vp, xp->i_number)) {
569 			VOP_UNLOCK(xvp, 0);
570 			vdrop(xvp);
571 			MNT_ILOCK(mp);
572 			continue;
573 		}
574 		/*
575 		 * If there is a fragment, clear it here.
576 		 */
577 		blkno = 0;
578 		loc = howmany(xp->i_size, fs->fs_bsize) - 1;
579 		if (loc < NDADDR) {
580 			len = fragroundup(fs, blkoff(fs, xp->i_size));
581 			if (len != 0 && len < fs->fs_bsize) {
582 				ffs_blkfree(ump, copy_fs, vp,
583 				    DIP(xp, i_db[loc]), len, xp->i_number);
584 				blkno = DIP(xp, i_db[loc]);
585 				DIP_SET(xp, i_db[loc], 0);
586 			}
587 		}
588 		snaplistsize += 1;
589 		if (xp->i_ump->um_fstype == UFS1)
590 			error = expunge_ufs1(vp, xp, copy_fs, fullacct_ufs1,
591 			    BLK_NOCOPY);
592 		else
593 			error = expunge_ufs2(vp, xp, copy_fs, fullacct_ufs2,
594 			    BLK_NOCOPY);
595 		if (blkno)
596 			DIP_SET(xp, i_db[loc], blkno);
597 		if (!error)
598 			error = ffs_freefile(ump, copy_fs, vp, xp->i_number,
599 			    xp->i_mode);
600 		VOP_UNLOCK(xvp, 0);
601 		vdrop(xvp);
602 		if (error) {
603 			free(copy_fs->fs_csp, M_UFSMNT);
604 			bawrite(sbp);
605 			sbp = NULL;
606 			MNT_VNODE_FOREACH_ABORT(mp, mvp);
607 			goto out1;
608 		}
609 		MNT_ILOCK(mp);
610 	}
611 	MNT_IUNLOCK(mp);
612 	/*
613 	 * Acquire a lock on the snapdata structure, creating it if necessary.
614 	 */
615 	sn = ffs_snapdata_acquire(devvp);
616 	/*
617 	 * Change vnode to use shared snapshot lock instead of the original
618 	 * private lock.
619 	 */
620 	vp->v_vnlock = &sn->sn_lock;
621 	lockmgr(&vp->v_lock, LK_RELEASE, NULL);
622 	xp = TAILQ_FIRST(&sn->sn_head);
623 	/*
624 	 * If this is the first snapshot on this filesystem, then we need
625 	 * to allocate the space for the list of preallocated snapshot blocks.
626 	 * This list will be refined below, but this preliminary one will
627 	 * keep us out of deadlock until the full one is ready.
628 	 */
629 	if (xp == NULL) {
630 		MALLOC(snapblklist, daddr_t *, snaplistsize * sizeof(daddr_t),
631 		    M_UFSMNT, M_WAITOK);
632 		blkp = &snapblklist[1];
633 		*blkp++ = lblkno(fs, fs->fs_sblockloc);
634 		blkno = fragstoblks(fs, fs->fs_csaddr);
635 		for (cg = 0; cg < fs->fs_ncg; cg++) {
636 			if (fragstoblks(fs, cgtod(fs, cg) > blkno))
637 				break;
638 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
639 		}
640 		len = howmany(fs->fs_cssize, fs->fs_bsize);
641 		for (loc = 0; loc < len; loc++)
642 			*blkp++ = blkno + loc;
643 		for (; cg < fs->fs_ncg; cg++)
644 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
645 		snapblklist[0] = blkp - snapblklist;
646 		VI_LOCK(devvp);
647 		if (sn->sn_blklist != NULL)
648 			panic("ffs_snapshot: non-empty list");
649 		sn->sn_blklist = snapblklist;
650 		sn->sn_listsize = blkp - snapblklist;
651 		VI_UNLOCK(devvp);
652 	}
653 	/*
654 	 * Record snapshot inode. Since this is the newest snapshot,
655 	 * it must be placed at the end of the list.
656 	 */
657 	VI_LOCK(devvp);
658 	fs->fs_snapinum[snaploc] = ip->i_number;
659 	if (ip->i_nextsnap.tqe_prev != 0)
660 		panic("ffs_snapshot: %d already on list", ip->i_number);
661 	TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
662 	devvp->v_vflag |= VV_COPYONWRITE;
663 	VI_UNLOCK(devvp);
664 	ASSERT_VOP_LOCKED(vp, "ffs_snapshot vp");
665 out1:
666 	KASSERT((sn != NULL && sbp != NULL && error == 0) ||
667 		(sn == NULL && sbp == NULL && error != 0),
668 		("email phk@ and mckusick@"));
669 	/*
670 	 * Resume operation on filesystem.
671 	 */
672 	vfs_write_resume(vp->v_mount);
673 	vn_start_write(NULL, &wrtmp, V_WAIT);
674 	if (collectsnapstats && starttime.tv_sec > 0) {
675 		nanotime(&endtime);
676 		timespecsub(&endtime, &starttime);
677 		printf("%s: suspended %ld.%03ld sec, redo %ld of %d\n",
678 		    vp->v_mount->mnt_stat.f_mntonname, (long)endtime.tv_sec,
679 		    endtime.tv_nsec / 1000000, redo, fs->fs_ncg);
680 	}
681 	if (sbp == NULL)
682 		goto out;
683 	/*
684 	 * Copy allocation information from all the snapshots in
685 	 * this snapshot and then expunge them from its view.
686 	 */
687 	TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap) {
688 		if (xp == ip)
689 			break;
690 		if (xp->i_ump->um_fstype == UFS1)
691 			error = expunge_ufs1(vp, xp, fs, snapacct_ufs1,
692 			    BLK_SNAP);
693 		else
694 			error = expunge_ufs2(vp, xp, fs, snapacct_ufs2,
695 			    BLK_SNAP);
696 		if (error == 0 && xp->i_effnlink == 0) {
697 			error = ffs_freefile(ump,
698 					     copy_fs,
699 					     vp,
700 					     xp->i_number,
701 					     xp->i_mode);
702 		}
703 		if (error) {
704 			fs->fs_snapinum[snaploc] = 0;
705 			goto done;
706 		}
707 	}
708 	/*
709 	 * Allocate space for the full list of preallocated snapshot blocks.
710 	 */
711 	MALLOC(snapblklist, daddr_t *, snaplistsize * sizeof(daddr_t),
712 	    M_UFSMNT, M_WAITOK);
713 	ip->i_snapblklist = &snapblklist[1];
714 	/*
715 	 * Expunge the blocks used by the snapshots from the set of
716 	 * blocks marked as used in the snapshot bitmaps. Also, collect
717 	 * the list of allocated blocks in i_snapblklist.
718 	 */
719 	if (ip->i_ump->um_fstype == UFS1)
720 		error = expunge_ufs1(vp, ip, copy_fs, mapacct_ufs1, BLK_SNAP);
721 	else
722 		error = expunge_ufs2(vp, ip, copy_fs, mapacct_ufs2, BLK_SNAP);
723 	if (error) {
724 		fs->fs_snapinum[snaploc] = 0;
725 		FREE(snapblklist, M_UFSMNT);
726 		goto done;
727 	}
728 	if (snaplistsize < ip->i_snapblklist - snapblklist)
729 		panic("ffs_snapshot: list too small");
730 	snaplistsize = ip->i_snapblklist - snapblklist;
731 	snapblklist[0] = snaplistsize;
732 	ip->i_snapblklist = 0;
733 	/*
734 	 * Write out the list of allocated blocks to the end of the snapshot.
735 	 */
736 	auio.uio_iov = &aiov;
737 	auio.uio_iovcnt = 1;
738 	aiov.iov_base = (void *)snapblklist;
739 	aiov.iov_len = snaplistsize * sizeof(daddr_t);
740 	auio.uio_resid = aiov.iov_len;;
741 	auio.uio_offset = ip->i_size;
742 	auio.uio_segflg = UIO_SYSSPACE;
743 	auio.uio_rw = UIO_WRITE;
744 	auio.uio_td = td;
745 	if ((error = VOP_WRITE(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
746 		fs->fs_snapinum[snaploc] = 0;
747 		FREE(snapblklist, M_UFSMNT);
748 		goto done;
749 	}
750 	/*
751 	 * Write the superblock and its summary information
752 	 * to the snapshot.
753 	 */
754 	blkno = fragstoblks(fs, fs->fs_csaddr);
755 	len = howmany(fs->fs_cssize, fs->fs_bsize);
756 	space = copy_fs->fs_csp;
757 	for (loc = 0; loc < len; loc++) {
758 		error = bread(vp, blkno + loc, fs->fs_bsize, KERNCRED, &nbp);
759 		if (error) {
760 			brelse(nbp);
761 			fs->fs_snapinum[snaploc] = 0;
762 			FREE(snapblklist, M_UFSMNT);
763 			goto done;
764 		}
765 		bcopy(space, nbp->b_data, fs->fs_bsize);
766 		space = (char *)space + fs->fs_bsize;
767 		bawrite(nbp);
768 	}
769 	/*
770 	 * As this is the newest list, it is the most inclusive, so
771 	 * should replace the previous list.
772 	 */
773 	VI_LOCK(devvp);
774 	space = sn->sn_blklist;
775 	sn->sn_blklist = snapblklist;
776 	sn->sn_listsize = snaplistsize;
777 	VI_UNLOCK(devvp);
778 	if (space != NULL)
779 		FREE(space, M_UFSMNT);
780 	/*
781 	 * If another process is currently writing the buffer containing
782 	 * the inode for this snapshot then a deadlock can occur. Drop
783 	 * the snapshot lock until the buffer has been written.
784 	 */
785 	VREF(vp);	/* Protect against ffs_snapgone() */
786 	VOP_UNLOCK(vp, 0);
787 	(void) bread(ip->i_devvp,
788 		     fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
789 		     (int) fs->fs_bsize, NOCRED, &nbp);
790 	brelse(nbp);
791 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
792 	if (ip->i_effnlink == 0)
793 		error = ENOENT;		/* Snapshot file unlinked */
794 	else
795 		vrele(vp);		/* Drop extra reference */
796 done:
797 	FREE(copy_fs->fs_csp, M_UFSMNT);
798 	bawrite(sbp);
799 out:
800 	NDFREE(&nd, NDF_ONLY_PNBUF);
801 	if (saved_nice > 0) {
802 		struct proc *p;
803 
804 		p = td->td_proc;
805 		PROC_LOCK(p);
806 		sched_nice(td->td_proc, saved_nice);
807 		PROC_UNLOCK(td->td_proc);
808 	}
809 	UFS_LOCK(ump);
810 	if (fs->fs_active != 0) {
811 		FREE(fs->fs_active, M_DEVBUF);
812 		fs->fs_active = 0;
813 	}
814 	UFS_UNLOCK(ump);
815 	MNT_ILOCK(mp);
816 	mp->mnt_flag = (mp->mnt_flag & MNT_QUOTA) | (flag & ~MNT_QUOTA);
817 	MNT_IUNLOCK(mp);
818 	if (error)
819 		(void) ffs_truncate(vp, (off_t)0, 0, NOCRED, td);
820 	(void) ffs_syncvnode(vp, MNT_WAIT);
821 	if (error)
822 		vput(vp);
823 	else
824 		VOP_UNLOCK(vp, 0);
825 	vrele(nd.ni_dvp);
826 	vn_finished_write(wrtmp);
827 	process_deferred_inactive(mp);
828 	return (error);
829 }
830 
831 /*
832  * Copy a cylinder group map. All the unallocated blocks are marked
833  * BLK_NOCOPY so that the snapshot knows that it need not copy them
834  * if they are later written. If passno is one, then this is a first
835  * pass, so only setting needs to be done. If passno is 2, then this
836  * is a revision to a previous pass which must be undone as the
837  * replacement pass is done.
838  */
839 static int
840 cgaccount(cg, vp, nbp, passno)
841 	int cg;
842 	struct vnode *vp;
843 	struct buf *nbp;
844 	int passno;
845 {
846 	struct buf *bp, *ibp;
847 	struct inode *ip;
848 	struct cg *cgp;
849 	struct fs *fs;
850 	ufs2_daddr_t base, numblks;
851 	int error, len, loc, indiroff;
852 
853 	ip = VTOI(vp);
854 	fs = ip->i_fs;
855 	error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)),
856 		(int)fs->fs_cgsize, KERNCRED, &bp);
857 	if (error) {
858 		brelse(bp);
859 		return (error);
860 	}
861 	cgp = (struct cg *)bp->b_data;
862 	if (!cg_chkmagic(cgp)) {
863 		brelse(bp);
864 		return (EIO);
865 	}
866 	UFS_LOCK(ip->i_ump);
867 	ACTIVESET(fs, cg);
868 	UFS_UNLOCK(ip->i_ump);
869 	bcopy(bp->b_data, nbp->b_data, fs->fs_cgsize);
870 	if (fs->fs_cgsize < fs->fs_bsize)
871 		bzero(&nbp->b_data[fs->fs_cgsize],
872 		    fs->fs_bsize - fs->fs_cgsize);
873 	cgp = (struct cg *)nbp->b_data;
874 	bqrelse(bp);
875 	if (passno == 2)
876 		nbp->b_flags |= B_VALIDSUSPWRT;
877 	numblks = howmany(fs->fs_size, fs->fs_frag);
878 	len = howmany(fs->fs_fpg, fs->fs_frag);
879 	base = cgbase(fs, cg) / fs->fs_frag;
880 	if (base + len >= numblks)
881 		len = numblks - base - 1;
882 	loc = 0;
883 	if (base < NDADDR) {
884 		for ( ; loc < NDADDR; loc++) {
885 			if (ffs_isblock(fs, cg_blksfree(cgp), loc))
886 				DIP_SET(ip, i_db[loc], BLK_NOCOPY);
887 			else if (passno == 2 && DIP(ip, i_db[loc])== BLK_NOCOPY)
888 				DIP_SET(ip, i_db[loc], 0);
889 			else if (passno == 1 && DIP(ip, i_db[loc])== BLK_NOCOPY)
890 				panic("ffs_snapshot: lost direct block");
891 		}
892 	}
893 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(base + loc)),
894 	    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
895 	if (error) {
896 		return (error);
897 	}
898 	indiroff = (base + loc - NDADDR) % NINDIR(fs);
899 	for ( ; loc < len; loc++, indiroff++) {
900 		if (indiroff >= NINDIR(fs)) {
901 			if (passno == 2)
902 				ibp->b_flags |= B_VALIDSUSPWRT;
903 			bawrite(ibp);
904 			error = UFS_BALLOC(vp,
905 			    lblktosize(fs, (off_t)(base + loc)),
906 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
907 			if (error) {
908 				return (error);
909 			}
910 			indiroff = 0;
911 		}
912 		if (ip->i_ump->um_fstype == UFS1) {
913 			if (ffs_isblock(fs, cg_blksfree(cgp), loc))
914 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] =
915 				    BLK_NOCOPY;
916 			else if (passno == 2 && ((ufs1_daddr_t *)(ibp->b_data))
917 			    [indiroff] == BLK_NOCOPY)
918 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] = 0;
919 			else if (passno == 1 && ((ufs1_daddr_t *)(ibp->b_data))
920 			    [indiroff] == BLK_NOCOPY)
921 				panic("ffs_snapshot: lost indirect block");
922 			continue;
923 		}
924 		if (ffs_isblock(fs, cg_blksfree(cgp), loc))
925 			((ufs2_daddr_t *)(ibp->b_data))[indiroff] = BLK_NOCOPY;
926 		else if (passno == 2 &&
927 		    ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY)
928 			((ufs2_daddr_t *)(ibp->b_data))[indiroff] = 0;
929 		else if (passno == 1 &&
930 		    ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY)
931 			panic("ffs_snapshot: lost indirect block");
932 	}
933 	if (passno == 2)
934 		ibp->b_flags |= B_VALIDSUSPWRT;
935 	bdwrite(ibp);
936 	return (0);
937 }
938 
939 /*
940  * Before expunging a snapshot inode, note all the
941  * blocks that it claims with BLK_SNAP so that fsck will
942  * be able to account for those blocks properly and so
943  * that this snapshot knows that it need not copy them
944  * if the other snapshot holding them is freed. This code
945  * is reproduced once each for UFS1 and UFS2.
946  */
947 static int
948 expunge_ufs1(snapvp, cancelip, fs, acctfunc, expungetype)
949 	struct vnode *snapvp;
950 	struct inode *cancelip;
951 	struct fs *fs;
952 	int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
953 	    struct fs *, ufs_lbn_t, int);
954 	int expungetype;
955 {
956 	int i, error, indiroff;
957 	ufs_lbn_t lbn, rlbn;
958 	ufs2_daddr_t len, blkno, numblks, blksperindir;
959 	struct ufs1_dinode *dip;
960 	struct thread *td = curthread;
961 	struct buf *bp;
962 
963 	/*
964 	 * Prepare to expunge the inode. If its inode block has not
965 	 * yet been copied, then allocate and fill the copy.
966 	 */
967 	lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number));
968 	blkno = 0;
969 	if (lbn < NDADDR) {
970 		blkno = VTOI(snapvp)->i_din1->di_db[lbn];
971 	} else {
972 		td->td_pflags |= TDP_COWINPROGRESS;
973 		error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn),
974 		   fs->fs_bsize, KERNCRED, BA_METAONLY, &bp);
975 		td->td_pflags &= ~TDP_COWINPROGRESS;
976 		if (error)
977 			return (error);
978 		indiroff = (lbn - NDADDR) % NINDIR(fs);
979 		blkno = ((ufs1_daddr_t *)(bp->b_data))[indiroff];
980 		bqrelse(bp);
981 	}
982 	if (blkno != 0) {
983 		if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp)))
984 			return (error);
985 	} else {
986 		error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn),
987 		    fs->fs_bsize, KERNCRED, 0, &bp);
988 		if (error)
989 			return (error);
990 		if ((error = readblock(snapvp, bp, lbn)) != 0)
991 			return (error);
992 	}
993 	/*
994 	 * Set a snapshot inode to be a zero length file, regular files
995 	 * or unlinked snapshots to be completely unallocated.
996 	 */
997 	dip = (struct ufs1_dinode *)bp->b_data +
998 	    ino_to_fsbo(fs, cancelip->i_number);
999 	if (expungetype == BLK_NOCOPY || cancelip->i_effnlink == 0)
1000 		dip->di_mode = 0;
1001 	dip->di_size = 0;
1002 	dip->di_blocks = 0;
1003 	dip->di_flags &= ~SF_SNAPSHOT;
1004 	bzero(&dip->di_db[0], (NDADDR + NIADDR) * sizeof(ufs1_daddr_t));
1005 	bdwrite(bp);
1006 	/*
1007 	 * Now go through and expunge all the blocks in the file
1008 	 * using the function requested.
1009 	 */
1010 	numblks = howmany(cancelip->i_size, fs->fs_bsize);
1011 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_db[0],
1012 	    &cancelip->i_din1->di_db[NDADDR], fs, 0, expungetype)))
1013 		return (error);
1014 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_ib[0],
1015 	    &cancelip->i_din1->di_ib[NIADDR], fs, -1, expungetype)))
1016 		return (error);
1017 	blksperindir = 1;
1018 	lbn = -NDADDR;
1019 	len = numblks - NDADDR;
1020 	rlbn = NDADDR;
1021 	for (i = 0; len > 0 && i < NIADDR; i++) {
1022 		error = indiracct_ufs1(snapvp, ITOV(cancelip), i,
1023 		    cancelip->i_din1->di_ib[i], lbn, rlbn, len,
1024 		    blksperindir, fs, acctfunc, expungetype);
1025 		if (error)
1026 			return (error);
1027 		blksperindir *= NINDIR(fs);
1028 		lbn -= blksperindir + 1;
1029 		len -= blksperindir;
1030 		rlbn += blksperindir;
1031 	}
1032 	return (0);
1033 }
1034 
1035 /*
1036  * Descend an indirect block chain for vnode cancelvp accounting for all
1037  * its indirect blocks in snapvp.
1038  */
1039 static int
1040 indiracct_ufs1(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks,
1041 	    blksperindir, fs, acctfunc, expungetype)
1042 	struct vnode *snapvp;
1043 	struct vnode *cancelvp;
1044 	int level;
1045 	ufs1_daddr_t blkno;
1046 	ufs_lbn_t lbn;
1047 	ufs_lbn_t rlbn;
1048 	ufs_lbn_t remblks;
1049 	ufs_lbn_t blksperindir;
1050 	struct fs *fs;
1051 	int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
1052 	    struct fs *, ufs_lbn_t, int);
1053 	int expungetype;
1054 {
1055 	int error, num, i;
1056 	ufs_lbn_t subblksperindir;
1057 	struct indir indirs[NIADDR + 2];
1058 	ufs1_daddr_t last, *bap;
1059 	struct buf *bp;
1060 
1061 	if (blkno == 0) {
1062 		if (expungetype == BLK_NOCOPY)
1063 			return (0);
1064 		panic("indiracct_ufs1: missing indir");
1065 	}
1066 	if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0)
1067 		return (error);
1068 	if (lbn != indirs[num - 1 - level].in_lbn || num < 2)
1069 		panic("indiracct_ufs1: botched params");
1070 	/*
1071 	 * We have to expand bread here since it will deadlock looking
1072 	 * up the block number for any blocks that are not in the cache.
1073 	 */
1074 	bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0);
1075 	bp->b_blkno = fsbtodb(fs, blkno);
1076 	if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0 &&
1077 	    (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) {
1078 		brelse(bp);
1079 		return (error);
1080 	}
1081 	/*
1082 	 * Account for the block pointers in this indirect block.
1083 	 */
1084 	last = howmany(remblks, blksperindir);
1085 	if (last > NINDIR(fs))
1086 		last = NINDIR(fs);
1087 	MALLOC(bap, ufs1_daddr_t *, fs->fs_bsize, M_DEVBUF, M_WAITOK);
1088 	bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize);
1089 	bqrelse(bp);
1090 	error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs,
1091 	    level == 0 ? rlbn : -1, expungetype);
1092 	if (error || level == 0)
1093 		goto out;
1094 	/*
1095 	 * Account for the block pointers in each of the indirect blocks
1096 	 * in the levels below us.
1097 	 */
1098 	subblksperindir = blksperindir / NINDIR(fs);
1099 	for (lbn++, level--, i = 0; i < last; i++) {
1100 		error = indiracct_ufs1(snapvp, cancelvp, level, bap[i], lbn,
1101 		    rlbn, remblks, subblksperindir, fs, acctfunc, expungetype);
1102 		if (error)
1103 			goto out;
1104 		rlbn += blksperindir;
1105 		lbn -= blksperindir;
1106 		remblks -= blksperindir;
1107 	}
1108 out:
1109 	FREE(bap, M_DEVBUF);
1110 	return (error);
1111 }
1112 
1113 /*
1114  * Do both snap accounting and map accounting.
1115  */
1116 static int
1117 fullacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype)
1118 	struct vnode *vp;
1119 	ufs1_daddr_t *oldblkp, *lastblkp;
1120 	struct fs *fs;
1121 	ufs_lbn_t lblkno;
1122 	int exptype;	/* BLK_SNAP or BLK_NOCOPY */
1123 {
1124 	int error;
1125 
1126 	if ((error = snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype)))
1127 		return (error);
1128 	return (mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype));
1129 }
1130 
1131 /*
1132  * Identify a set of blocks allocated in a snapshot inode.
1133  */
1134 static int
1135 snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1136 	struct vnode *vp;
1137 	ufs1_daddr_t *oldblkp, *lastblkp;
1138 	struct fs *fs;
1139 	ufs_lbn_t lblkno;
1140 	int expungetype;	/* BLK_SNAP or BLK_NOCOPY */
1141 {
1142 	struct inode *ip = VTOI(vp);
1143 	ufs1_daddr_t blkno, *blkp;
1144 	ufs_lbn_t lbn;
1145 	struct buf *ibp;
1146 	int error;
1147 
1148 	for ( ; oldblkp < lastblkp; oldblkp++) {
1149 		blkno = *oldblkp;
1150 		if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP)
1151 			continue;
1152 		lbn = fragstoblks(fs, blkno);
1153 		if (lbn < NDADDR) {
1154 			blkp = &ip->i_din1->di_db[lbn];
1155 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1156 		} else {
1157 			error = ffs_balloc_ufs1(vp, lblktosize(fs, (off_t)lbn),
1158 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1159 			if (error)
1160 				return (error);
1161 			blkp = &((ufs1_daddr_t *)(ibp->b_data))
1162 			    [(lbn - NDADDR) % NINDIR(fs)];
1163 		}
1164 		/*
1165 		 * If we are expunging a snapshot vnode and we
1166 		 * find a block marked BLK_NOCOPY, then it is
1167 		 * one that has been allocated to this snapshot after
1168 		 * we took our current snapshot and can be ignored.
1169 		 */
1170 		if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) {
1171 			if (lbn >= NDADDR)
1172 				brelse(ibp);
1173 		} else {
1174 			if (*blkp != 0)
1175 				panic("snapacct_ufs1: bad block");
1176 			*blkp = expungetype;
1177 			if (lbn >= NDADDR)
1178 				bdwrite(ibp);
1179 		}
1180 	}
1181 	return (0);
1182 }
1183 
1184 /*
1185  * Account for a set of blocks allocated in a snapshot inode.
1186  */
1187 static int
1188 mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1189 	struct vnode *vp;
1190 	ufs1_daddr_t *oldblkp, *lastblkp;
1191 	struct fs *fs;
1192 	ufs_lbn_t lblkno;
1193 	int expungetype;
1194 {
1195 	ufs1_daddr_t blkno;
1196 	struct inode *ip;
1197 	ino_t inum;
1198 	int acctit;
1199 
1200 	ip = VTOI(vp);
1201 	inum = ip->i_number;
1202 	if (lblkno == -1)
1203 		acctit = 0;
1204 	else
1205 		acctit = 1;
1206 	for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) {
1207 		blkno = *oldblkp;
1208 		if (blkno == 0 || blkno == BLK_NOCOPY)
1209 			continue;
1210 		if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP)
1211 			*ip->i_snapblklist++ = lblkno;
1212 		if (blkno == BLK_SNAP)
1213 			blkno = blkstofrags(fs, lblkno);
1214 		ffs_blkfree(ip->i_ump, fs, vp, blkno, fs->fs_bsize, inum);
1215 	}
1216 	return (0);
1217 }
1218 
1219 /*
1220  * Before expunging a snapshot inode, note all the
1221  * blocks that it claims with BLK_SNAP so that fsck will
1222  * be able to account for those blocks properly and so
1223  * that this snapshot knows that it need not copy them
1224  * if the other snapshot holding them is freed. This code
1225  * is reproduced once each for UFS1 and UFS2.
1226  */
1227 static int
1228 expunge_ufs2(snapvp, cancelip, fs, acctfunc, expungetype)
1229 	struct vnode *snapvp;
1230 	struct inode *cancelip;
1231 	struct fs *fs;
1232 	int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
1233 	    struct fs *, ufs_lbn_t, int);
1234 	int expungetype;
1235 {
1236 	int i, error, indiroff;
1237 	ufs_lbn_t lbn, rlbn;
1238 	ufs2_daddr_t len, blkno, numblks, blksperindir;
1239 	struct ufs2_dinode *dip;
1240 	struct thread *td = curthread;
1241 	struct buf *bp;
1242 
1243 	/*
1244 	 * Prepare to expunge the inode. If its inode block has not
1245 	 * yet been copied, then allocate and fill the copy.
1246 	 */
1247 	lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number));
1248 	blkno = 0;
1249 	if (lbn < NDADDR) {
1250 		blkno = VTOI(snapvp)->i_din2->di_db[lbn];
1251 	} else {
1252 		td->td_pflags |= TDP_COWINPROGRESS;
1253 		error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn),
1254 		   fs->fs_bsize, KERNCRED, BA_METAONLY, &bp);
1255 		td->td_pflags &= ~TDP_COWINPROGRESS;
1256 		if (error)
1257 			return (error);
1258 		indiroff = (lbn - NDADDR) % NINDIR(fs);
1259 		blkno = ((ufs2_daddr_t *)(bp->b_data))[indiroff];
1260 		bqrelse(bp);
1261 	}
1262 	if (blkno != 0) {
1263 		if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp)))
1264 			return (error);
1265 	} else {
1266 		error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn),
1267 		    fs->fs_bsize, KERNCRED, 0, &bp);
1268 		if (error)
1269 			return (error);
1270 		if ((error = readblock(snapvp, bp, lbn)) != 0)
1271 			return (error);
1272 	}
1273 	/*
1274 	 * Set a snapshot inode to be a zero length file, regular files
1275 	 * to be completely unallocated.
1276 	 */
1277 	dip = (struct ufs2_dinode *)bp->b_data +
1278 	    ino_to_fsbo(fs, cancelip->i_number);
1279 	if (expungetype == BLK_NOCOPY)
1280 		dip->di_mode = 0;
1281 	dip->di_size = 0;
1282 	dip->di_blocks = 0;
1283 	dip->di_flags &= ~SF_SNAPSHOT;
1284 	bzero(&dip->di_db[0], (NDADDR + NIADDR) * sizeof(ufs2_daddr_t));
1285 	bdwrite(bp);
1286 	/*
1287 	 * Now go through and expunge all the blocks in the file
1288 	 * using the function requested.
1289 	 */
1290 	numblks = howmany(cancelip->i_size, fs->fs_bsize);
1291 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_db[0],
1292 	    &cancelip->i_din2->di_db[NDADDR], fs, 0, expungetype)))
1293 		return (error);
1294 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_ib[0],
1295 	    &cancelip->i_din2->di_ib[NIADDR], fs, -1, expungetype)))
1296 		return (error);
1297 	blksperindir = 1;
1298 	lbn = -NDADDR;
1299 	len = numblks - NDADDR;
1300 	rlbn = NDADDR;
1301 	for (i = 0; len > 0 && i < NIADDR; i++) {
1302 		error = indiracct_ufs2(snapvp, ITOV(cancelip), i,
1303 		    cancelip->i_din2->di_ib[i], lbn, rlbn, len,
1304 		    blksperindir, fs, acctfunc, expungetype);
1305 		if (error)
1306 			return (error);
1307 		blksperindir *= NINDIR(fs);
1308 		lbn -= blksperindir + 1;
1309 		len -= blksperindir;
1310 		rlbn += blksperindir;
1311 	}
1312 	return (0);
1313 }
1314 
1315 /*
1316  * Descend an indirect block chain for vnode cancelvp accounting for all
1317  * its indirect blocks in snapvp.
1318  */
1319 static int
1320 indiracct_ufs2(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks,
1321 	    blksperindir, fs, acctfunc, expungetype)
1322 	struct vnode *snapvp;
1323 	struct vnode *cancelvp;
1324 	int level;
1325 	ufs2_daddr_t blkno;
1326 	ufs_lbn_t lbn;
1327 	ufs_lbn_t rlbn;
1328 	ufs_lbn_t remblks;
1329 	ufs_lbn_t blksperindir;
1330 	struct fs *fs;
1331 	int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
1332 	    struct fs *, ufs_lbn_t, int);
1333 	int expungetype;
1334 {
1335 	int error, num, i;
1336 	ufs_lbn_t subblksperindir;
1337 	struct indir indirs[NIADDR + 2];
1338 	ufs2_daddr_t last, *bap;
1339 	struct buf *bp;
1340 
1341 	if (blkno == 0) {
1342 		if (expungetype == BLK_NOCOPY)
1343 			return (0);
1344 		panic("indiracct_ufs2: missing indir");
1345 	}
1346 	if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0)
1347 		return (error);
1348 	if (lbn != indirs[num - 1 - level].in_lbn || num < 2)
1349 		panic("indiracct_ufs2: botched params");
1350 	/*
1351 	 * We have to expand bread here since it will deadlock looking
1352 	 * up the block number for any blocks that are not in the cache.
1353 	 */
1354 	bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0);
1355 	bp->b_blkno = fsbtodb(fs, blkno);
1356 	if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0 &&
1357 	    (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) {
1358 		brelse(bp);
1359 		return (error);
1360 	}
1361 	/*
1362 	 * Account for the block pointers in this indirect block.
1363 	 */
1364 	last = howmany(remblks, blksperindir);
1365 	if (last > NINDIR(fs))
1366 		last = NINDIR(fs);
1367 	MALLOC(bap, ufs2_daddr_t *, fs->fs_bsize, M_DEVBUF, M_WAITOK);
1368 	bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize);
1369 	bqrelse(bp);
1370 	error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs,
1371 	    level == 0 ? rlbn : -1, expungetype);
1372 	if (error || level == 0)
1373 		goto out;
1374 	/*
1375 	 * Account for the block pointers in each of the indirect blocks
1376 	 * in the levels below us.
1377 	 */
1378 	subblksperindir = blksperindir / NINDIR(fs);
1379 	for (lbn++, level--, i = 0; i < last; i++) {
1380 		error = indiracct_ufs2(snapvp, cancelvp, level, bap[i], lbn,
1381 		    rlbn, remblks, subblksperindir, fs, acctfunc, expungetype);
1382 		if (error)
1383 			goto out;
1384 		rlbn += blksperindir;
1385 		lbn -= blksperindir;
1386 		remblks -= blksperindir;
1387 	}
1388 out:
1389 	FREE(bap, M_DEVBUF);
1390 	return (error);
1391 }
1392 
1393 /*
1394  * Do both snap accounting and map accounting.
1395  */
1396 static int
1397 fullacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype)
1398 	struct vnode *vp;
1399 	ufs2_daddr_t *oldblkp, *lastblkp;
1400 	struct fs *fs;
1401 	ufs_lbn_t lblkno;
1402 	int exptype;	/* BLK_SNAP or BLK_NOCOPY */
1403 {
1404 	int error;
1405 
1406 	if ((error = snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype)))
1407 		return (error);
1408 	return (mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype));
1409 }
1410 
1411 /*
1412  * Identify a set of blocks allocated in a snapshot inode.
1413  */
1414 static int
1415 snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1416 	struct vnode *vp;
1417 	ufs2_daddr_t *oldblkp, *lastblkp;
1418 	struct fs *fs;
1419 	ufs_lbn_t lblkno;
1420 	int expungetype;	/* BLK_SNAP or BLK_NOCOPY */
1421 {
1422 	struct inode *ip = VTOI(vp);
1423 	ufs2_daddr_t blkno, *blkp;
1424 	ufs_lbn_t lbn;
1425 	struct buf *ibp;
1426 	int error;
1427 
1428 	for ( ; oldblkp < lastblkp; oldblkp++) {
1429 		blkno = *oldblkp;
1430 		if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP)
1431 			continue;
1432 		lbn = fragstoblks(fs, blkno);
1433 		if (lbn < NDADDR) {
1434 			blkp = &ip->i_din2->di_db[lbn];
1435 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1436 		} else {
1437 			error = ffs_balloc_ufs2(vp, lblktosize(fs, (off_t)lbn),
1438 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1439 			if (error)
1440 				return (error);
1441 			blkp = &((ufs2_daddr_t *)(ibp->b_data))
1442 			    [(lbn - NDADDR) % NINDIR(fs)];
1443 		}
1444 		/*
1445 		 * If we are expunging a snapshot vnode and we
1446 		 * find a block marked BLK_NOCOPY, then it is
1447 		 * one that has been allocated to this snapshot after
1448 		 * we took our current snapshot and can be ignored.
1449 		 */
1450 		if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) {
1451 			if (lbn >= NDADDR)
1452 				brelse(ibp);
1453 		} else {
1454 			if (*blkp != 0)
1455 				panic("snapacct_ufs2: bad block");
1456 			*blkp = expungetype;
1457 			if (lbn >= NDADDR)
1458 				bdwrite(ibp);
1459 		}
1460 	}
1461 	return (0);
1462 }
1463 
1464 /*
1465  * Account for a set of blocks allocated in a snapshot inode.
1466  */
1467 static int
1468 mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1469 	struct vnode *vp;
1470 	ufs2_daddr_t *oldblkp, *lastblkp;
1471 	struct fs *fs;
1472 	ufs_lbn_t lblkno;
1473 	int expungetype;
1474 {
1475 	ufs2_daddr_t blkno;
1476 	struct inode *ip;
1477 	ino_t inum;
1478 	int acctit;
1479 
1480 	ip = VTOI(vp);
1481 	inum = ip->i_number;
1482 	if (lblkno == -1)
1483 		acctit = 0;
1484 	else
1485 		acctit = 1;
1486 	for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) {
1487 		blkno = *oldblkp;
1488 		if (blkno == 0 || blkno == BLK_NOCOPY)
1489 			continue;
1490 		if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP)
1491 			*ip->i_snapblklist++ = lblkno;
1492 		if (blkno == BLK_SNAP)
1493 			blkno = blkstofrags(fs, lblkno);
1494 		ffs_blkfree(ip->i_ump, fs, vp, blkno, fs->fs_bsize, inum);
1495 	}
1496 	return (0);
1497 }
1498 
1499 /*
1500  * Decrement extra reference on snapshot when last name is removed.
1501  * It will not be freed until the last open reference goes away.
1502  */
1503 void
1504 ffs_snapgone(ip)
1505 	struct inode *ip;
1506 {
1507 	struct inode *xp;
1508 	struct fs *fs;
1509 	int snaploc;
1510 	struct snapdata *sn;
1511 	struct ufsmount *ump;
1512 
1513 	/*
1514 	 * Find snapshot in incore list.
1515 	 */
1516 	xp = NULL;
1517 	sn = ip->i_devvp->v_rdev->si_snapdata;
1518 	if (sn != NULL)
1519 		TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap)
1520 			if (xp == ip)
1521 				break;
1522 	if (xp != NULL)
1523 		vrele(ITOV(ip));
1524 	else if (snapdebug)
1525 		printf("ffs_snapgone: lost snapshot vnode %d\n",
1526 		    ip->i_number);
1527 	/*
1528 	 * Delete snapshot inode from superblock. Keep list dense.
1529 	 */
1530 	fs = ip->i_fs;
1531 	ump = ip->i_ump;
1532 	UFS_LOCK(ump);
1533 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++)
1534 		if (fs->fs_snapinum[snaploc] == ip->i_number)
1535 			break;
1536 	if (snaploc < FSMAXSNAP) {
1537 		for (snaploc++; snaploc < FSMAXSNAP; snaploc++) {
1538 			if (fs->fs_snapinum[snaploc] == 0)
1539 				break;
1540 			fs->fs_snapinum[snaploc - 1] = fs->fs_snapinum[snaploc];
1541 		}
1542 		fs->fs_snapinum[snaploc - 1] = 0;
1543 	}
1544 	UFS_UNLOCK(ump);
1545 }
1546 
1547 /*
1548  * Prepare a snapshot file for being removed.
1549  */
1550 void
1551 ffs_snapremove(vp)
1552 	struct vnode *vp;
1553 {
1554 	struct inode *ip;
1555 	struct vnode *devvp;
1556 	struct buf *ibp;
1557 	struct fs *fs;
1558 	ufs2_daddr_t numblks, blkno, dblk;
1559 	int error, loc, last;
1560 	struct snapdata *sn;
1561 
1562 	ip = VTOI(vp);
1563 	fs = ip->i_fs;
1564 	devvp = ip->i_devvp;
1565 	/*
1566 	 * If active, delete from incore list (this snapshot may
1567 	 * already have been in the process of being deleted, so
1568 	 * would not have been active).
1569 	 *
1570 	 * Clear copy-on-write flag if last snapshot.
1571 	 */
1572 	VI_LOCK(devvp);
1573 	if (ip->i_nextsnap.tqe_prev != 0) {
1574 		sn = devvp->v_rdev->si_snapdata;
1575 		TAILQ_REMOVE(&sn->sn_head, ip, i_nextsnap);
1576 		ip->i_nextsnap.tqe_prev = 0;
1577 		VI_UNLOCK(devvp);
1578 		lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
1579 		KASSERT(vp->v_vnlock == &sn->sn_lock,
1580 			("ffs_snapremove: lost lock mutation"));
1581 		vp->v_vnlock = &vp->v_lock;
1582 		VI_LOCK(devvp);
1583 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
1584 		try_free_snapdata(devvp);
1585 	} else
1586 		VI_UNLOCK(devvp);
1587 	/*
1588 	 * Clear all BLK_NOCOPY fields. Pass any block claims to other
1589 	 * snapshots that want them (see ffs_snapblkfree below).
1590 	 */
1591 	for (blkno = 1; blkno < NDADDR; blkno++) {
1592 		dblk = DIP(ip, i_db[blkno]);
1593 		if (dblk == 0)
1594 			continue;
1595 		if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1596 			DIP_SET(ip, i_db[blkno], 0);
1597 		else if ((dblk == blkstofrags(fs, blkno) &&
1598 		     ffs_snapblkfree(fs, ip->i_devvp, dblk, fs->fs_bsize,
1599 		     ip->i_number))) {
1600 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) -
1601 			    btodb(fs->fs_bsize));
1602 			DIP_SET(ip, i_db[blkno], 0);
1603 		}
1604 	}
1605 	numblks = howmany(ip->i_size, fs->fs_bsize);
1606 	for (blkno = NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
1607 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
1608 		    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1609 		if (error)
1610 			continue;
1611 		if (fs->fs_size - blkno > NINDIR(fs))
1612 			last = NINDIR(fs);
1613 		else
1614 			last = fs->fs_size - blkno;
1615 		for (loc = 0; loc < last; loc++) {
1616 			if (ip->i_ump->um_fstype == UFS1) {
1617 				dblk = ((ufs1_daddr_t *)(ibp->b_data))[loc];
1618 				if (dblk == 0)
1619 					continue;
1620 				if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1621 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1622 				else if ((dblk == blkstofrags(fs, blkno) &&
1623 				     ffs_snapblkfree(fs, ip->i_devvp, dblk,
1624 				     fs->fs_bsize, ip->i_number))) {
1625 					ip->i_din1->di_blocks -=
1626 					    btodb(fs->fs_bsize);
1627 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1628 				}
1629 				continue;
1630 			}
1631 			dblk = ((ufs2_daddr_t *)(ibp->b_data))[loc];
1632 			if (dblk == 0)
1633 				continue;
1634 			if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1635 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1636 			else if ((dblk == blkstofrags(fs, blkno) &&
1637 			     ffs_snapblkfree(fs, ip->i_devvp, dblk,
1638 			     fs->fs_bsize, ip->i_number))) {
1639 				ip->i_din2->di_blocks -= btodb(fs->fs_bsize);
1640 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1641 			}
1642 		}
1643 		bawrite(ibp);
1644 	}
1645 	/*
1646 	 * Clear snapshot flag and drop reference.
1647 	 */
1648 	ip->i_flags &= ~SF_SNAPSHOT;
1649 	DIP_SET(ip, i_flags, ip->i_flags);
1650 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
1651 #ifdef QUOTA
1652 	/*
1653 	 * Reenable disk quotas for ex-snapshot file.
1654 	 */
1655 	if (!getinoquota(ip))
1656 		(void) chkdq(ip, DIP(ip, i_blocks), KERNCRED, FORCE);
1657 #endif
1658 }
1659 
1660 /*
1661  * Notification that a block is being freed. Return zero if the free
1662  * should be allowed to proceed. Return non-zero if the snapshot file
1663  * wants to claim the block. The block will be claimed if it is an
1664  * uncopied part of one of the snapshots. It will be freed if it is
1665  * either a BLK_NOCOPY or has already been copied in all of the snapshots.
1666  * If a fragment is being freed, then all snapshots that care about
1667  * it must make a copy since a snapshot file can only claim full sized
1668  * blocks. Note that if more than one snapshot file maps the block,
1669  * we can pick one at random to claim it. Since none of the snapshots
1670  * can change, we are assurred that they will all see the same unmodified
1671  * image. When deleting a snapshot file (see ffs_snapremove above), we
1672  * must push any of these claimed blocks to one of the other snapshots
1673  * that maps it. These claimed blocks are easily identified as they will
1674  * have a block number equal to their logical block number within the
1675  * snapshot. A copied block can never have this property because they
1676  * must always have been allocated from a BLK_NOCOPY location.
1677  */
1678 int
1679 ffs_snapblkfree(fs, devvp, bno, size, inum)
1680 	struct fs *fs;
1681 	struct vnode *devvp;
1682 	ufs2_daddr_t bno;
1683 	long size;
1684 	ino_t inum;
1685 {
1686 	struct buf *ibp, *cbp, *savedcbp = 0;
1687 	struct thread *td = curthread;
1688 	struct inode *ip;
1689 	struct vnode *vp = NULL;
1690 	ufs_lbn_t lbn;
1691 	ufs2_daddr_t blkno;
1692 	int indiroff = 0, error = 0, claimedblk = 0;
1693 	struct snapdata *sn;
1694 
1695 	lbn = fragstoblks(fs, bno);
1696 retry:
1697 	VI_LOCK(devvp);
1698 	sn = devvp->v_rdev->si_snapdata;
1699 	if (sn == NULL) {
1700 		VI_UNLOCK(devvp);
1701 		return (0);
1702 	}
1703 	if (lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
1704 	    VI_MTX(devvp)) != 0)
1705 		goto retry;
1706 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
1707 		vp = ITOV(ip);
1708 		/*
1709 		 * Lookup block being written.
1710 		 */
1711 		if (lbn < NDADDR) {
1712 			blkno = DIP(ip, i_db[lbn]);
1713 		} else {
1714 			td->td_pflags |= TDP_COWINPROGRESS;
1715 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1716 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1717 			td->td_pflags &= ~TDP_COWINPROGRESS;
1718 			if (error)
1719 				break;
1720 			indiroff = (lbn - NDADDR) % NINDIR(fs);
1721 			if (ip->i_ump->um_fstype == UFS1)
1722 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
1723 			else
1724 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
1725 		}
1726 		/*
1727 		 * Check to see if block needs to be copied.
1728 		 */
1729 		if (blkno == 0) {
1730 			/*
1731 			 * A block that we map is being freed. If it has not
1732 			 * been claimed yet, we will claim or copy it (below).
1733 			 */
1734 			claimedblk = 1;
1735 		} else if (blkno == BLK_SNAP) {
1736 			/*
1737 			 * No previous snapshot claimed the block,
1738 			 * so it will be freed and become a BLK_NOCOPY
1739 			 * (don't care) for us.
1740 			 */
1741 			if (claimedblk)
1742 				panic("snapblkfree: inconsistent block type");
1743 			if (lbn < NDADDR) {
1744 				DIP_SET(ip, i_db[lbn], BLK_NOCOPY);
1745 				ip->i_flag |= IN_CHANGE | IN_UPDATE;
1746 			} else if (ip->i_ump->um_fstype == UFS1) {
1747 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] =
1748 				    BLK_NOCOPY;
1749 				bdwrite(ibp);
1750 			} else {
1751 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] =
1752 				    BLK_NOCOPY;
1753 				bdwrite(ibp);
1754 			}
1755 			continue;
1756 		} else /* BLK_NOCOPY or default */ {
1757 			/*
1758 			 * If the snapshot has already copied the block
1759 			 * (default), or does not care about the block,
1760 			 * it is not needed.
1761 			 */
1762 			if (lbn >= NDADDR)
1763 				bqrelse(ibp);
1764 			continue;
1765 		}
1766 		/*
1767 		 * If this is a full size block, we will just grab it
1768 		 * and assign it to the snapshot inode. Otherwise we
1769 		 * will proceed to copy it. See explanation for this
1770 		 * routine as to why only a single snapshot needs to
1771 		 * claim this block.
1772 		 */
1773 		if (size == fs->fs_bsize) {
1774 #ifdef DEBUG
1775 			if (snapdebug)
1776 				printf("%s %d lbn %jd from inum %d\n",
1777 				    "Grabonremove: snapino", ip->i_number,
1778 				    (intmax_t)lbn, inum);
1779 #endif
1780 			if (lbn < NDADDR) {
1781 				DIP_SET(ip, i_db[lbn], bno);
1782 			} else if (ip->i_ump->um_fstype == UFS1) {
1783 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] = bno;
1784 				bdwrite(ibp);
1785 			} else {
1786 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] = bno;
1787 				bdwrite(ibp);
1788 			}
1789 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + btodb(size));
1790 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1791 			lockmgr(vp->v_vnlock, LK_RELEASE, NULL);
1792 			return (1);
1793 		}
1794 		if (lbn >= NDADDR)
1795 			bqrelse(ibp);
1796 		/*
1797 		 * Allocate the block into which to do the copy. Note that this
1798 		 * allocation will never require any additional allocations for
1799 		 * the snapshot inode.
1800 		 */
1801 		td->td_pflags |= TDP_COWINPROGRESS;
1802 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1803 		    fs->fs_bsize, KERNCRED, 0, &cbp);
1804 		td->td_pflags &= ~TDP_COWINPROGRESS;
1805 		if (error)
1806 			break;
1807 #ifdef DEBUG
1808 		if (snapdebug)
1809 			printf("%s%d lbn %jd %s %d size %ld to blkno %jd\n",
1810 			    "Copyonremove: snapino ", ip->i_number,
1811 			    (intmax_t)lbn, "for inum", inum, size,
1812 			    (intmax_t)cbp->b_blkno);
1813 #endif
1814 		/*
1815 		 * If we have already read the old block contents, then
1816 		 * simply copy them to the new block. Note that we need
1817 		 * to synchronously write snapshots that have not been
1818 		 * unlinked, and hence will be visible after a crash,
1819 		 * to ensure their integrity.
1820 		 */
1821 		if (savedcbp != 0) {
1822 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
1823 			bawrite(cbp);
1824 			if (dopersistence && ip->i_effnlink > 0)
1825 				(void) ffs_syncvnode(vp, MNT_WAIT);
1826 			continue;
1827 		}
1828 		/*
1829 		 * Otherwise, read the old block contents into the buffer.
1830 		 */
1831 		if ((error = readblock(vp, cbp, lbn)) != 0) {
1832 			bzero(cbp->b_data, fs->fs_bsize);
1833 			bawrite(cbp);
1834 			if (dopersistence && ip->i_effnlink > 0)
1835 				(void) ffs_syncvnode(vp, MNT_WAIT);
1836 			break;
1837 		}
1838 		savedcbp = cbp;
1839 	}
1840 	/*
1841 	 * Note that we need to synchronously write snapshots that
1842 	 * have not been unlinked, and hence will be visible after
1843 	 * a crash, to ensure their integrity.
1844 	 */
1845 	if (savedcbp) {
1846 		vp = savedcbp->b_vp;
1847 		bawrite(savedcbp);
1848 		if (dopersistence && VTOI(vp)->i_effnlink > 0)
1849 			(void) ffs_syncvnode(vp, MNT_WAIT);
1850 	}
1851 	/*
1852 	 * If we have been unable to allocate a block in which to do
1853 	 * the copy, then return non-zero so that the fragment will
1854 	 * not be freed. Although space will be lost, the snapshot
1855 	 * will stay consistent.
1856 	 */
1857 	lockmgr(vp->v_vnlock, LK_RELEASE, NULL);
1858 	return (error);
1859 }
1860 
1861 /*
1862  * Associate snapshot files when mounting.
1863  */
1864 void
1865 ffs_snapshot_mount(mp)
1866 	struct mount *mp;
1867 {
1868 	struct ufsmount *ump = VFSTOUFS(mp);
1869 	struct vnode *devvp = ump->um_devvp;
1870 	struct fs *fs = ump->um_fs;
1871 	struct thread *td = curthread;
1872 	struct snapdata *sn;
1873 	struct vnode *vp;
1874 	struct vnode *lastvp;
1875 	struct inode *ip;
1876 	struct uio auio;
1877 	struct iovec aiov;
1878 	void *snapblklist;
1879 	char *reason;
1880 	daddr_t snaplistsize;
1881 	int error, snaploc, loc;
1882 
1883 	/*
1884 	 * XXX The following needs to be set before ffs_truncate or
1885 	 * VOP_READ can be called.
1886 	 */
1887 	mp->mnt_stat.f_iosize = fs->fs_bsize;
1888 	/*
1889 	 * Process each snapshot listed in the superblock.
1890 	 */
1891 	vp = NULL;
1892 	lastvp = NULL;
1893 	sn = NULL;
1894 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++) {
1895 		if (fs->fs_snapinum[snaploc] == 0)
1896 			break;
1897 		if ((error = ffs_vget(mp, fs->fs_snapinum[snaploc],
1898 		    LK_EXCLUSIVE, &vp)) != 0){
1899 			printf("ffs_snapshot_mount: vget failed %d\n", error);
1900 			continue;
1901 		}
1902 		ip = VTOI(vp);
1903 		if ((ip->i_flags & SF_SNAPSHOT) == 0 || ip->i_size ==
1904 		    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag))) {
1905 			if ((ip->i_flags & SF_SNAPSHOT) == 0) {
1906 				reason = "non-snapshot";
1907 			} else {
1908 				reason = "old format snapshot";
1909 				(void)ffs_truncate(vp, (off_t)0, 0, NOCRED, td);
1910 				(void)ffs_syncvnode(vp, MNT_WAIT);
1911 			}
1912 			printf("ffs_snapshot_mount: %s inode %d\n",
1913 			    reason, fs->fs_snapinum[snaploc]);
1914 			vput(vp);
1915 			vp = NULL;
1916 			for (loc = snaploc + 1; loc < FSMAXSNAP; loc++) {
1917 				if (fs->fs_snapinum[loc] == 0)
1918 					break;
1919 				fs->fs_snapinum[loc - 1] = fs->fs_snapinum[loc];
1920 			}
1921 			fs->fs_snapinum[loc - 1] = 0;
1922 			snaploc--;
1923 			continue;
1924 		}
1925 		/*
1926 		 * Acquire a lock on the snapdata structure, creating it if
1927 		 * necessary.
1928 		 */
1929 		sn = ffs_snapdata_acquire(devvp);
1930 		/*
1931 		 * Change vnode to use shared snapshot lock instead of the
1932 		 * original private lock.
1933 		 */
1934 		vp->v_vnlock = &sn->sn_lock;
1935 		lockmgr(&vp->v_lock, LK_RELEASE, NULL);
1936 		/*
1937 		 * Link it onto the active snapshot list.
1938 		 */
1939 		VI_LOCK(devvp);
1940 		if (ip->i_nextsnap.tqe_prev != 0)
1941 			panic("ffs_snapshot_mount: %d already on list",
1942 			    ip->i_number);
1943 		else
1944 			TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
1945 		vp->v_vflag |= VV_SYSTEM;
1946 		VI_UNLOCK(devvp);
1947 		VOP_UNLOCK(vp, 0);
1948 		lastvp = vp;
1949 	}
1950 	vp = lastvp;
1951 	/*
1952 	 * No usable snapshots found.
1953 	 */
1954 	if (sn == NULL || vp == NULL)
1955 		return;
1956 	/*
1957 	 * Allocate the space for the block hints list. We always want to
1958 	 * use the list from the newest snapshot.
1959 	 */
1960 	auio.uio_iov = &aiov;
1961 	auio.uio_iovcnt = 1;
1962 	aiov.iov_base = (void *)&snaplistsize;
1963 	aiov.iov_len = sizeof(snaplistsize);
1964 	auio.uio_resid = aiov.iov_len;
1965 	auio.uio_offset =
1966 	    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag));
1967 	auio.uio_segflg = UIO_SYSSPACE;
1968 	auio.uio_rw = UIO_READ;
1969 	auio.uio_td = td;
1970 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1971 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
1972 		printf("ffs_snapshot_mount: read_1 failed %d\n", error);
1973 		VOP_UNLOCK(vp, 0);
1974 		return;
1975 	}
1976 	MALLOC(snapblklist, void *, snaplistsize * sizeof(daddr_t),
1977 	    M_UFSMNT, M_WAITOK);
1978 	auio.uio_iovcnt = 1;
1979 	aiov.iov_base = snapblklist;
1980 	aiov.iov_len = snaplistsize * sizeof (daddr_t);
1981 	auio.uio_resid = aiov.iov_len;
1982 	auio.uio_offset -= sizeof(snaplistsize);
1983 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
1984 		printf("ffs_snapshot_mount: read_2 failed %d\n", error);
1985 		VOP_UNLOCK(vp, 0);
1986 		FREE(snapblklist, M_UFSMNT);
1987 		return;
1988 	}
1989 	VOP_UNLOCK(vp, 0);
1990 	VI_LOCK(devvp);
1991 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_mount");
1992 	sn->sn_listsize = snaplistsize;
1993 	sn->sn_blklist = (daddr_t *)snapblklist;
1994 	devvp->v_vflag |= VV_COPYONWRITE;
1995 	VI_UNLOCK(devvp);
1996 }
1997 
1998 /*
1999  * Disassociate snapshot files when unmounting.
2000  */
2001 void
2002 ffs_snapshot_unmount(mp)
2003 	struct mount *mp;
2004 {
2005 	struct vnode *devvp = VFSTOUFS(mp)->um_devvp;
2006 	struct snapdata *sn;
2007 	struct inode *xp;
2008 	struct vnode *vp;
2009 
2010 	VI_LOCK(devvp);
2011 	sn = devvp->v_rdev->si_snapdata;
2012 	while (sn != NULL && (xp = TAILQ_FIRST(&sn->sn_head)) != NULL) {
2013 		vp = ITOV(xp);
2014 		TAILQ_REMOVE(&sn->sn_head, xp, i_nextsnap);
2015 		xp->i_nextsnap.tqe_prev = 0;
2016 		lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE,
2017 		    VI_MTX(devvp));
2018 		lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
2019 		KASSERT(vp->v_vnlock == &sn->sn_lock,
2020 		("ffs_snapshot_unmount: lost lock mutation"));
2021 		vp->v_vnlock = &vp->v_lock;
2022 		lockmgr(&vp->v_lock, LK_RELEASE, NULL);
2023 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
2024 		if (xp->i_effnlink > 0)
2025 			vrele(vp);
2026 		VI_LOCK(devvp);
2027 		sn = devvp->v_rdev->si_snapdata;
2028 	}
2029 	try_free_snapdata(devvp);
2030 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_unmount");
2031 }
2032 
2033 /*
2034  * Check the buffer block to be belong to device buffer that shall be
2035  * locked after snaplk. devvp shall be locked on entry, and will be
2036  * leaved locked upon exit.
2037  */
2038 static int
2039 ffs_bp_snapblk(devvp, bp)
2040 	struct vnode *devvp;
2041 	struct buf *bp;
2042 {
2043 	struct snapdata *sn;
2044 	struct fs *fs;
2045 	ufs2_daddr_t lbn, *snapblklist;
2046 	int lower, upper, mid;
2047 
2048 	ASSERT_VI_LOCKED(devvp, "ffs_bp_snapblk");
2049 	KASSERT(devvp->v_type == VCHR, ("Not a device %p", devvp));
2050 	sn = devvp->v_rdev->si_snapdata;
2051 	if (sn == NULL || TAILQ_FIRST(&sn->sn_head) == NULL)
2052 		return (0);
2053 	fs = TAILQ_FIRST(&sn->sn_head)->i_fs;
2054 	lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno));
2055 	snapblklist = sn->sn_blklist;
2056 	upper = sn->sn_listsize - 1;
2057 	lower = 1;
2058 	while (lower <= upper) {
2059 		mid = (lower + upper) / 2;
2060 		if (snapblklist[mid] == lbn)
2061 			break;
2062 		if (snapblklist[mid] < lbn)
2063 			lower = mid + 1;
2064 		else
2065 			upper = mid - 1;
2066 	}
2067 	if (lower <= upper)
2068 		return (1);
2069 	return (0);
2070 }
2071 
2072 void
2073 ffs_bdflush(bo, bp)
2074 	struct bufobj *bo;
2075 	struct buf *bp;
2076 {
2077 	struct thread *td;
2078 	struct vnode *vp, *devvp;
2079 	struct buf *nbp;
2080 	int bp_bdskip;
2081 
2082 	if (bo->bo_dirty.bv_cnt <= dirtybufthresh)
2083 		return;
2084 
2085 	td = curthread;
2086 	vp = bp->b_vp;
2087 	devvp = bo->__bo_vnode;
2088 	KASSERT(vp == devvp, ("devvp != vp %p %p", bo, bp));
2089 
2090 	VI_LOCK(devvp);
2091 	bp_bdskip = ffs_bp_snapblk(devvp, bp);
2092 	if (bp_bdskip)
2093 		bdwriteskip++;
2094 	VI_UNLOCK(devvp);
2095 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10 && !bp_bdskip) {
2096 		(void) VOP_FSYNC(vp, MNT_NOWAIT, td);
2097 		altbufferflushes++;
2098 	} else {
2099 		BO_LOCK(bo);
2100 		/*
2101 		 * Try to find a buffer to flush.
2102 		 */
2103 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2104 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2105 			    BUF_LOCK(nbp,
2106 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2107 				continue;
2108 			if (bp == nbp)
2109 				panic("bdwrite: found ourselves");
2110 			BO_UNLOCK(bo);
2111 			/*
2112 			 * Don't countdeps with the bo lock
2113 			 * held.
2114 			 */
2115 			if (buf_countdeps(nbp, 0)) {
2116 				BO_LOCK(bo);
2117 				BUF_UNLOCK(nbp);
2118 				continue;
2119 			}
2120 			if (bp_bdskip) {
2121 				VI_LOCK(devvp);
2122 				if (!ffs_bp_snapblk(vp, nbp)) {
2123 					if (BO_MTX(bo) != VI_MTX(vp)) {
2124 						VI_UNLOCK(devvp);
2125 						BO_LOCK(bo);
2126 					}
2127 					BUF_UNLOCK(nbp);
2128 					continue;
2129 				}
2130 				VI_UNLOCK(devvp);
2131 			}
2132 			if (nbp->b_flags & B_CLUSTEROK) {
2133 				vfs_bio_awrite(nbp);
2134 			} else {
2135 				bremfree(nbp);
2136 				bawrite(nbp);
2137 			}
2138 			dirtybufferflushes++;
2139 			break;
2140 		}
2141 		if (nbp == NULL)
2142 			BO_UNLOCK(bo);
2143 	}
2144 }
2145 
2146 /*
2147  * Check for need to copy block that is about to be written,
2148  * copying the block if necessary.
2149  */
2150 int
2151 ffs_copyonwrite(devvp, bp)
2152 	struct vnode *devvp;
2153 	struct buf *bp;
2154 {
2155 	struct snapdata *sn;
2156 	struct buf *ibp, *cbp, *savedcbp = 0;
2157 	struct thread *td = curthread;
2158 	struct fs *fs;
2159 	struct inode *ip;
2160 	struct vnode *vp = 0;
2161 	ufs2_daddr_t lbn, blkno, *snapblklist;
2162 	int lower, upper, mid, indiroff, error = 0;
2163 	int launched_async_io, prev_norunningbuf;
2164 	long saved_runningbufspace;
2165 
2166 	if (devvp != bp->b_vp && (VTOI(bp->b_vp)->i_flags & SF_SNAPSHOT) != 0)
2167 		return (0);		/* Update on a snapshot file */
2168 	if (td->td_pflags & TDP_COWINPROGRESS)
2169 		panic("ffs_copyonwrite: recursive call");
2170 	/*
2171 	 * First check to see if it is in the preallocated list.
2172 	 * By doing this check we avoid several potential deadlocks.
2173 	 */
2174 	VI_LOCK(devvp);
2175 	sn = devvp->v_rdev->si_snapdata;
2176 	if (sn == NULL ||
2177 	    TAILQ_EMPTY(&sn->sn_head)) {
2178 		VI_UNLOCK(devvp);
2179 		return (0);		/* No snapshot */
2180 	}
2181 	ip = TAILQ_FIRST(&sn->sn_head);
2182 	fs = ip->i_fs;
2183 	lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno));
2184 	snapblklist = sn->sn_blklist;
2185 	upper = sn->sn_listsize - 1;
2186 	lower = 1;
2187 	while (lower <= upper) {
2188 		mid = (lower + upper) / 2;
2189 		if (snapblklist[mid] == lbn)
2190 			break;
2191 		if (snapblklist[mid] < lbn)
2192 			lower = mid + 1;
2193 		else
2194 			upper = mid - 1;
2195 	}
2196 	if (lower <= upper) {
2197 		VI_UNLOCK(devvp);
2198 		return (0);
2199 	}
2200 	launched_async_io = 0;
2201 	prev_norunningbuf = td->td_pflags & TDP_NORUNNINGBUF;
2202 	/*
2203 	 * Since I/O on bp isn't yet in progress and it may be blocked
2204 	 * for a long time waiting on snaplk, back it out of
2205 	 * runningbufspace, possibly waking other threads waiting for space.
2206 	 */
2207 	saved_runningbufspace = bp->b_runningbufspace;
2208 	if (saved_runningbufspace != 0)
2209 		runningbufwakeup(bp);
2210 	/*
2211 	 * Not in the precomputed list, so check the snapshots.
2212 	 */
2213 	while (lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
2214 	    VI_MTX(devvp)) != 0) {
2215 		VI_LOCK(devvp);
2216 		sn = devvp->v_rdev->si_snapdata;
2217 		if (sn == NULL ||
2218 		    TAILQ_EMPTY(&sn->sn_head)) {
2219 			VI_UNLOCK(devvp);
2220 			if (saved_runningbufspace != 0) {
2221 				bp->b_runningbufspace = saved_runningbufspace;
2222 				atomic_add_int(&runningbufspace,
2223 					       bp->b_runningbufspace);
2224 			}
2225 			return (0);		/* Snapshot gone */
2226 		}
2227 	}
2228 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
2229 		vp = ITOV(ip);
2230 		/*
2231 		 * We ensure that everything of our own that needs to be
2232 		 * copied will be done at the time that ffs_snapshot is
2233 		 * called. Thus we can skip the check here which can
2234 		 * deadlock in doing the lookup in UFS_BALLOC.
2235 		 */
2236 		if (bp->b_vp == vp)
2237 			continue;
2238 		/*
2239 		 * Check to see if block needs to be copied. We do not have
2240 		 * to hold the snapshot lock while doing this lookup as it
2241 		 * will never require any additional allocations for the
2242 		 * snapshot inode.
2243 		 */
2244 		if (lbn < NDADDR) {
2245 			blkno = DIP(ip, i_db[lbn]);
2246 		} else {
2247 			td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2248 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2249 			   fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
2250 			td->td_pflags &= ~TDP_COWINPROGRESS;
2251 			if (error)
2252 				break;
2253 			indiroff = (lbn - NDADDR) % NINDIR(fs);
2254 			if (ip->i_ump->um_fstype == UFS1)
2255 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
2256 			else
2257 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
2258 			bqrelse(ibp);
2259 		}
2260 #ifdef INVARIANTS
2261 		if (blkno == BLK_SNAP && bp->b_lblkno >= 0)
2262 			panic("ffs_copyonwrite: bad copy block");
2263 #endif
2264 		if (blkno != 0)
2265 			continue;
2266 		/*
2267 		 * Allocate the block into which to do the copy. Since
2268 		 * multiple processes may all try to copy the same block,
2269 		 * we have to recheck our need to do a copy if we sleep
2270 		 * waiting for the lock.
2271 		 *
2272 		 * Because all snapshots on a filesystem share a single
2273 		 * lock, we ensure that we will never be in competition
2274 		 * with another process to allocate a block.
2275 		 */
2276 		td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2277 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2278 		    fs->fs_bsize, KERNCRED, 0, &cbp);
2279 		td->td_pflags &= ~TDP_COWINPROGRESS;
2280 		if (error)
2281 			break;
2282 #ifdef DEBUG
2283 		if (snapdebug) {
2284 			printf("Copyonwrite: snapino %d lbn %jd for ",
2285 			    ip->i_number, (intmax_t)lbn);
2286 			if (bp->b_vp == devvp)
2287 				printf("fs metadata");
2288 			else
2289 				printf("inum %d", VTOI(bp->b_vp)->i_number);
2290 			printf(" lblkno %jd to blkno %jd\n",
2291 			    (intmax_t)bp->b_lblkno, (intmax_t)cbp->b_blkno);
2292 		}
2293 #endif
2294 		/*
2295 		 * If we have already read the old block contents, then
2296 		 * simply copy them to the new block. Note that we need
2297 		 * to synchronously write snapshots that have not been
2298 		 * unlinked, and hence will be visible after a crash,
2299 		 * to ensure their integrity.
2300 		 */
2301 		if (savedcbp != 0) {
2302 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
2303 			bawrite(cbp);
2304 			if (dopersistence && ip->i_effnlink > 0)
2305 				(void) ffs_syncvnode(vp, MNT_WAIT);
2306 			else
2307 				launched_async_io = 1;
2308 			continue;
2309 		}
2310 		/*
2311 		 * Otherwise, read the old block contents into the buffer.
2312 		 */
2313 		if ((error = readblock(vp, cbp, lbn)) != 0) {
2314 			bzero(cbp->b_data, fs->fs_bsize);
2315 			bawrite(cbp);
2316 			if (dopersistence && ip->i_effnlink > 0)
2317 				(void) ffs_syncvnode(vp, MNT_WAIT);
2318 			else
2319 				launched_async_io = 1;
2320 			break;
2321 		}
2322 		savedcbp = cbp;
2323 	}
2324 	/*
2325 	 * Note that we need to synchronously write snapshots that
2326 	 * have not been unlinked, and hence will be visible after
2327 	 * a crash, to ensure their integrity.
2328 	 */
2329 	if (savedcbp) {
2330 		vp = savedcbp->b_vp;
2331 		bawrite(savedcbp);
2332 		if (dopersistence && VTOI(vp)->i_effnlink > 0)
2333 			(void) ffs_syncvnode(vp, MNT_WAIT);
2334 		else
2335 			launched_async_io = 1;
2336 	}
2337 	lockmgr(vp->v_vnlock, LK_RELEASE, NULL);
2338 	td->td_pflags = (td->td_pflags & ~TDP_NORUNNINGBUF) |
2339 		prev_norunningbuf;
2340 	if (launched_async_io && (td->td_pflags & TDP_NORUNNINGBUF) == 0)
2341 		waitrunningbufspace();
2342 	/*
2343 	 * I/O on bp will now be started, so count it in runningbufspace.
2344 	 */
2345 	if (saved_runningbufspace != 0) {
2346 		bp->b_runningbufspace = saved_runningbufspace;
2347 		atomic_add_int(&runningbufspace, bp->b_runningbufspace);
2348 	}
2349 	return (error);
2350 }
2351 
2352 /*
2353  * Read the specified block into the given buffer.
2354  * Much of this boiler-plate comes from bwrite().
2355  */
2356 static int
2357 readblock(vp, bp, lbn)
2358 	struct vnode *vp;
2359 	struct buf *bp;
2360 	ufs2_daddr_t lbn;
2361 {
2362 	struct inode *ip = VTOI(vp);
2363 	struct bio *bip;
2364 
2365 	bip = g_alloc_bio();
2366 	bip->bio_cmd = BIO_READ;
2367 	bip->bio_offset = dbtob(fsbtodb(ip->i_fs, blkstofrags(ip->i_fs, lbn)));
2368 	bip->bio_data = bp->b_data;
2369 	bip->bio_length = bp->b_bcount;
2370 	bip->bio_done = NULL;
2371 
2372 	g_io_request(bip, ip->i_devvp->v_bufobj.bo_private);
2373 	bp->b_error = biowait(bip, "snaprdb");
2374 	g_destroy_bio(bip);
2375 	return (bp->b_error);
2376 }
2377 
2378 /*
2379  * Process file deletes that were deferred by ufs_inactive() due to
2380  * the file system being suspended. Transfer IN_LAZYACCESS into
2381  * IN_MODIFIED for vnodes that were accessed during suspension.
2382  */
2383 static void
2384 process_deferred_inactive(struct mount *mp)
2385 {
2386 	struct vnode *vp, *mvp;
2387 	struct inode *ip;
2388 	struct thread *td;
2389 	int error;
2390 
2391 	td = curthread;
2392 	(void) vn_start_secondary_write(NULL, &mp, V_WAIT);
2393 	MNT_ILOCK(mp);
2394  loop:
2395 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2396 		VI_LOCK(vp);
2397 		/*
2398 		 * IN_LAZYACCESS is checked here without holding any
2399 		 * vnode lock, but this flag is set only while holding
2400 		 * vnode interlock.
2401 		 */
2402 		if (vp->v_type == VNON || (vp->v_iflag & VI_DOOMED) != 0 ||
2403 		    ((VTOI(vp)->i_flag & IN_LAZYACCESS) == 0 &&
2404 			((vp->v_iflag & VI_OWEINACT) == 0 ||
2405 			vp->v_usecount > 0))) {
2406 			VI_UNLOCK(vp);
2407 			continue;
2408 		}
2409 		MNT_IUNLOCK(mp);
2410 		vholdl(vp);
2411 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2412 		if (error != 0) {
2413 			vdrop(vp);
2414 			MNT_ILOCK(mp);
2415 			if (error == ENOENT)
2416 				continue;	/* vnode recycled */
2417 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2418 			goto loop;
2419 		}
2420 		ip = VTOI(vp);
2421 		if ((ip->i_flag & IN_LAZYACCESS) != 0) {
2422 			ip->i_flag &= ~IN_LAZYACCESS;
2423 			ip->i_flag |= IN_MODIFIED;
2424 		}
2425 		VI_LOCK(vp);
2426 		if ((vp->v_iflag & VI_OWEINACT) == 0 || vp->v_usecount > 0) {
2427 			VI_UNLOCK(vp);
2428 			VOP_UNLOCK(vp, 0);
2429 			vdrop(vp);
2430 			MNT_ILOCK(mp);
2431 			continue;
2432 		}
2433 
2434 		VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2435 			 ("process_deferred_inactive: "
2436 			  "recursed on VI_DOINGINACT"));
2437 		vp->v_iflag |= VI_DOINGINACT;
2438 		vp->v_iflag &= ~VI_OWEINACT;
2439 		VI_UNLOCK(vp);
2440 		(void) VOP_INACTIVE(vp, td);
2441 		VI_LOCK(vp);
2442 		VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2443 			 ("process_deferred_inactive: lost VI_DOINGINACT"));
2444 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2445 			 ("process_deferred_inactive: got VI_OWEINACT"));
2446 		vp->v_iflag &= ~VI_DOINGINACT;
2447 		VI_UNLOCK(vp);
2448 		VOP_UNLOCK(vp, 0);
2449 		vdrop(vp);
2450 		MNT_ILOCK(mp);
2451 	}
2452 	MNT_IUNLOCK(mp);
2453 	vn_finished_secondary_write(mp);
2454 }
2455 
2456 static struct snapdata *
2457 ffs_snapdata_alloc(void)
2458 {
2459 	struct snapdata *sn;
2460 
2461 	/*
2462 	 * Fetch a snapdata from the free list if there is one available.
2463 	 */
2464 	mtx_lock(&snapfree_lock);
2465 	sn = LIST_FIRST(&snapfree);
2466 	if (sn != NULL)
2467 		LIST_REMOVE(sn, sn_link);
2468 	mtx_unlock(&snapfree_lock);
2469 	if (sn != NULL)
2470 		return (sn);
2471 	/*
2472  	 * If there were no free snapdatas allocate one.
2473 	 */
2474 	sn = malloc(sizeof *sn, M_UFSMNT, M_WAITOK | M_ZERO);
2475 	TAILQ_INIT(&sn->sn_head);
2476 	lockinit(&sn->sn_lock, PVFS, "snaplk", VLKTIMEOUT,
2477 	    LK_CANRECURSE | LK_NOSHARE);
2478 	return (sn);
2479 }
2480 
2481 /*
2482  * The snapdata is never freed because we can not be certain that
2483  * there are no threads sleeping on the snap lock.  Persisting
2484  * them permanently avoids costly synchronization in ffs_lock().
2485  */
2486 static void
2487 ffs_snapdata_free(struct snapdata *sn)
2488 {
2489 	mtx_lock(&snapfree_lock);
2490 	LIST_INSERT_HEAD(&snapfree, sn, sn_link);
2491 	mtx_unlock(&snapfree_lock);
2492 }
2493 
2494 /* Try to free snapdata associated with devvp */
2495 static void
2496 try_free_snapdata(struct vnode *devvp)
2497 {
2498 	struct snapdata *sn;
2499 	ufs2_daddr_t *snapblklist;
2500 
2501 	ASSERT_VI_LOCKED(devvp, "try_free_snapdata");
2502 	sn = devvp->v_rdev->si_snapdata;
2503 
2504 	if (sn == NULL || TAILQ_FIRST(&sn->sn_head) != NULL ||
2505 	    (devvp->v_vflag & VV_COPYONWRITE) == 0) {
2506 		VI_UNLOCK(devvp);
2507 		return;
2508 	}
2509 
2510 	devvp->v_rdev->si_snapdata = NULL;
2511 	devvp->v_vflag &= ~VV_COPYONWRITE;
2512 	lockmgr(&sn->sn_lock, LK_DRAIN|LK_INTERLOCK, VI_MTX(devvp));
2513 	snapblklist = sn->sn_blklist;
2514 	sn->sn_blklist = NULL;
2515 	sn->sn_listsize = 0;
2516 	lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
2517 	if (snapblklist != NULL)
2518 		FREE(snapblklist, M_UFSMNT);
2519 	ffs_snapdata_free(sn);
2520 }
2521 
2522 static struct snapdata *
2523 ffs_snapdata_acquire(struct vnode *devvp)
2524 {
2525 	struct snapdata *nsn;
2526 	struct snapdata *sn;
2527 
2528 	/*
2529  	 * Allocate a free snapdata.  This is done before acquiring the
2530 	 * devvp lock to avoid allocation while the devvp interlock is
2531 	 * held.
2532 	 */
2533 	nsn = ffs_snapdata_alloc();
2534 	/*
2535 	 * If there snapshots already exist on this filesystem grab a
2536 	 * reference to the shared lock.  Otherwise this is the first
2537 	 * snapshot on this filesystem and we need to use our
2538 	 * pre-allocated snapdata.
2539 	 */
2540 	VI_LOCK(devvp);
2541 	if (devvp->v_rdev->si_snapdata == NULL) {
2542 		devvp->v_rdev->si_snapdata = nsn;
2543 		nsn = NULL;
2544 	}
2545 	sn = devvp->v_rdev->si_snapdata;
2546 	/*
2547 	 * Acquire the snapshot lock.
2548 	 */
2549 	lockmgr(&sn->sn_lock,
2550 	    LK_INTERLOCK | LK_EXCLUSIVE | LK_RETRY, VI_MTX(devvp));
2551 	/*
2552 	 * Free any unused snapdata.
2553 	 */
2554 	if (nsn != NULL)
2555 		ffs_snapdata_free(nsn);
2556 
2557 	return (sn);
2558 }
2559 
2560 #endif
2561