xref: /freebsd/sys/ufs/ffs/ffs_snapshot.c (revision 547bc083d614f3639f5632d9e39d79e828519318)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright 2000 Marshall Kirk McKusick. All Rights Reserved.
5  *
6  * Further information about snapshots can be obtained from:
7  *
8  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
9  *	1614 Oxford Street		mckusick@mckusick.com
10  *	Berkeley, CA 94709-1608		+1-510-843-9542
11  *	USA
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  *
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  *
23  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
24  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
25  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
26  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
27  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	@(#)ffs_snapshot.c	8.11 (McKusick) 7/23/00
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_quota.h"
42 
43 #include <sys/param.h>
44 #include <sys/kernel.h>
45 #include <sys/systm.h>
46 #include <sys/conf.h>
47 #include <sys/bio.h>
48 #include <sys/buf.h>
49 #include <sys/fcntl.h>
50 #include <sys/proc.h>
51 #include <sys/namei.h>
52 #include <sys/sched.h>
53 #include <sys/stat.h>
54 #include <sys/malloc.h>
55 #include <sys/mount.h>
56 #include <sys/resource.h>
57 #include <sys/resourcevar.h>
58 #include <sys/rwlock.h>
59 #include <sys/vnode.h>
60 
61 #include <geom/geom.h>
62 
63 #include <ufs/ufs/extattr.h>
64 #include <ufs/ufs/quota.h>
65 #include <ufs/ufs/ufsmount.h>
66 #include <ufs/ufs/inode.h>
67 #include <ufs/ufs/ufs_extern.h>
68 
69 #include <ufs/ffs/fs.h>
70 #include <ufs/ffs/ffs_extern.h>
71 
72 #define KERNCRED thread0.td_ucred
73 #define DEBUG 1
74 
75 #include "opt_ffs.h"
76 
77 #ifdef NO_FFS_SNAPSHOT
78 int
79 ffs_snapshot(mp, snapfile)
80 	struct mount *mp;
81 	char *snapfile;
82 {
83 	return (EINVAL);
84 }
85 
86 int
87 ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, wkhd)
88 	struct fs *fs;
89 	struct vnode *devvp;
90 	ufs2_daddr_t bno;
91 	long size;
92 	ino_t inum;
93 	enum vtype vtype;
94 	struct workhead *wkhd;
95 {
96 	return (EINVAL);
97 }
98 
99 void
100 ffs_snapremove(vp)
101 	struct vnode *vp;
102 {
103 }
104 
105 void
106 ffs_snapshot_mount(mp)
107 	struct mount *mp;
108 {
109 }
110 
111 void
112 ffs_snapshot_unmount(mp)
113 	struct mount *mp;
114 {
115 }
116 
117 void
118 ffs_snapgone(ip)
119 	struct inode *ip;
120 {
121 }
122 
123 int
124 ffs_copyonwrite(devvp, bp)
125 	struct vnode *devvp;
126 	struct buf *bp;
127 {
128 	return (EINVAL);
129 }
130 
131 void
132 ffs_sync_snap(mp, waitfor)
133 	struct mount *mp;
134 	int waitfor;
135 {
136 }
137 
138 #else
139 FEATURE(ffs_snapshot, "FFS snapshot support");
140 
141 LIST_HEAD(, snapdata) snapfree;
142 static struct mtx snapfree_lock;
143 MTX_SYSINIT(ffs_snapfree, &snapfree_lock, "snapdata free list", MTX_DEF);
144 
145 static int cgaccount(int, struct vnode *, struct buf *, int);
146 static int expunge_ufs1(struct vnode *, struct inode *, struct fs *,
147     int (*)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, struct fs *,
148     ufs_lbn_t, int), int, int);
149 static int indiracct_ufs1(struct vnode *, struct vnode *, int,
150     ufs1_daddr_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, struct fs *,
151     int (*)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *, struct fs *,
152     ufs_lbn_t, int), int);
153 static int fullacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
154     struct fs *, ufs_lbn_t, int);
155 static int snapacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
156     struct fs *, ufs_lbn_t, int);
157 static int mapacct_ufs1(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
158     struct fs *, ufs_lbn_t, int);
159 static int expunge_ufs2(struct vnode *, struct inode *, struct fs *,
160     int (*)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, struct fs *,
161     ufs_lbn_t, int), int, int);
162 static int indiracct_ufs2(struct vnode *, struct vnode *, int,
163     ufs2_daddr_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, ufs_lbn_t, struct fs *,
164     int (*)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *, struct fs *,
165     ufs_lbn_t, int), int);
166 static int fullacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
167     struct fs *, ufs_lbn_t, int);
168 static int snapacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
169     struct fs *, ufs_lbn_t, int);
170 static int mapacct_ufs2(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
171     struct fs *, ufs_lbn_t, int);
172 static int readblock(struct vnode *vp, struct buf *, ufs2_daddr_t);
173 static void try_free_snapdata(struct vnode *devvp);
174 static struct snapdata *ffs_snapdata_acquire(struct vnode *devvp);
175 static int ffs_bp_snapblk(struct vnode *, struct buf *);
176 
177 /*
178  * To ensure the consistency of snapshots across crashes, we must
179  * synchronously write out copied blocks before allowing the
180  * originals to be modified. Because of the rather severe speed
181  * penalty that this imposes, the code normally only ensures
182  * persistence for the filesystem metadata contained within a
183  * snapshot. Setting the following flag allows this crash
184  * persistence to be enabled for file contents.
185  */
186 int dopersistence = 0;
187 
188 #ifdef DEBUG
189 #include <sys/sysctl.h>
190 SYSCTL_INT(_debug, OID_AUTO, dopersistence, CTLFLAG_RW, &dopersistence, 0, "");
191 static int snapdebug = 0;
192 SYSCTL_INT(_debug, OID_AUTO, snapdebug, CTLFLAG_RW, &snapdebug, 0, "");
193 int collectsnapstats = 0;
194 SYSCTL_INT(_debug, OID_AUTO, collectsnapstats, CTLFLAG_RW, &collectsnapstats,
195 	0, "");
196 #endif /* DEBUG */
197 
198 /*
199  * Create a snapshot file and initialize it for the filesystem.
200  */
201 int
202 ffs_snapshot(mp, snapfile)
203 	struct mount *mp;
204 	char *snapfile;
205 {
206 	ufs2_daddr_t numblks, blkno, *blkp, *snapblklist;
207 	int error, cg, snaploc;
208 	int i, size, len, loc;
209 	ufs2_daddr_t blockno;
210 	uint64_t flag;
211 	struct timespec starttime = {0, 0}, endtime;
212 	char saved_nice = 0;
213 	long redo = 0, snaplistsize = 0;
214 	int32_t *lp;
215 	void *space;
216 	struct fs *copy_fs = NULL, *fs;
217 	struct thread *td = curthread;
218 	struct inode *ip, *xp;
219 	struct buf *bp, *nbp, *ibp;
220 	struct nameidata nd;
221 	struct mount *wrtmp;
222 	struct vattr vat;
223 	struct vnode *vp, *xvp, *mvp, *devvp;
224 	struct uio auio;
225 	struct iovec aiov;
226 	struct snapdata *sn;
227 	struct ufsmount *ump;
228 
229 	ump = VFSTOUFS(mp);
230 	fs = ump->um_fs;
231 	sn = NULL;
232 	/*
233 	 * At the moment, journaled soft updates cannot support
234 	 * taking snapshots.
235 	 */
236 	if (MOUNTEDSUJ(mp)) {
237 		vfs_mount_error(mp, "%s: Snapshots are not yet supported when "
238 		    "running with journaled soft updates", fs->fs_fsmnt);
239 		return (EOPNOTSUPP);
240 	}
241 	MNT_ILOCK(mp);
242 	flag = mp->mnt_flag;
243 	MNT_IUNLOCK(mp);
244 	/*
245 	 * Need to serialize access to snapshot code per filesystem.
246 	 */
247 	/*
248 	 * Assign a snapshot slot in the superblock.
249 	 */
250 	UFS_LOCK(ump);
251 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++)
252 		if (fs->fs_snapinum[snaploc] == 0)
253 			break;
254 	UFS_UNLOCK(ump);
255 	if (snaploc == FSMAXSNAP)
256 		return (ENOSPC);
257 	/*
258 	 * Create the snapshot file.
259 	 */
260 restart:
261 	NDINIT(&nd, CREATE, LOCKPARENT | LOCKLEAF | NOCACHE, UIO_SYSSPACE,
262 	    snapfile, td);
263 	if ((error = namei(&nd)) != 0)
264 		return (error);
265 	if (nd.ni_vp != NULL) {
266 		vput(nd.ni_vp);
267 		error = EEXIST;
268 	}
269 	if (nd.ni_dvp->v_mount != mp)
270 		error = EXDEV;
271 	if (error) {
272 		NDFREE(&nd, NDF_ONLY_PNBUF);
273 		if (nd.ni_dvp == nd.ni_vp)
274 			vrele(nd.ni_dvp);
275 		else
276 			vput(nd.ni_dvp);
277 		return (error);
278 	}
279 	VATTR_NULL(&vat);
280 	vat.va_type = VREG;
281 	vat.va_mode = S_IRUSR;
282 	vat.va_vaflags |= VA_EXCLUSIVE;
283 	if (VOP_GETWRITEMOUNT(nd.ni_dvp, &wrtmp))
284 		wrtmp = NULL;
285 	if (wrtmp != mp)
286 		panic("ffs_snapshot: mount mismatch");
287 	vfs_rel(wrtmp);
288 	if (vn_start_write(NULL, &wrtmp, V_NOWAIT) != 0) {
289 		NDFREE(&nd, NDF_ONLY_PNBUF);
290 		vput(nd.ni_dvp);
291 		if ((error = vn_start_write(NULL, &wrtmp,
292 		    V_XSLEEP | PCATCH)) != 0)
293 			return (error);
294 		goto restart;
295 	}
296 	error = VOP_CREATE(nd.ni_dvp, &nd.ni_vp, &nd.ni_cnd, &vat);
297 	VOP_UNLOCK(nd.ni_dvp, 0);
298 	if (error) {
299 		NDFREE(&nd, NDF_ONLY_PNBUF);
300 		vn_finished_write(wrtmp);
301 		vrele(nd.ni_dvp);
302 		return (error);
303 	}
304 	vp = nd.ni_vp;
305 	vp->v_vflag |= VV_SYSTEM;
306 	ip = VTOI(vp);
307 	devvp = ITODEVVP(ip);
308 	/*
309 	 * Allocate and copy the last block contents so as to be able
310 	 * to set size to that of the filesystem.
311 	 */
312 	numblks = howmany(fs->fs_size, fs->fs_frag);
313 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)),
314 	    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
315 	if (error)
316 		goto out;
317 	ip->i_size = lblktosize(fs, (off_t)numblks);
318 	DIP_SET(ip, i_size, ip->i_size);
319 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
320 	error = readblock(vp, bp, numblks - 1);
321 	bawrite(bp);
322 	if (error != 0)
323 		goto out;
324 	/*
325 	 * Preallocate critical data structures so that we can copy
326 	 * them in without further allocation after we suspend all
327 	 * operations on the filesystem. We would like to just release
328 	 * the allocated buffers without writing them since they will
329 	 * be filled in below once we are ready to go, but this upsets
330 	 * the soft update code, so we go ahead and write the new buffers.
331 	 *
332 	 * Allocate all indirect blocks and mark all of them as not
333 	 * needing to be copied.
334 	 */
335 	for (blkno = UFS_NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
336 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
337 		    fs->fs_bsize, td->td_ucred, BA_METAONLY, &ibp);
338 		if (error)
339 			goto out;
340 		bawrite(ibp);
341 	}
342 	/*
343 	 * Allocate copies for the superblock and its summary information.
344 	 */
345 	error = UFS_BALLOC(vp, fs->fs_sblockloc, fs->fs_sbsize, KERNCRED,
346 	    0, &nbp);
347 	if (error)
348 		goto out;
349 	bawrite(nbp);
350 	blkno = fragstoblks(fs, fs->fs_csaddr);
351 	len = howmany(fs->fs_cssize, fs->fs_bsize);
352 	for (loc = 0; loc < len; loc++) {
353 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(blkno + loc)),
354 		    fs->fs_bsize, KERNCRED, 0, &nbp);
355 		if (error)
356 			goto out;
357 		bawrite(nbp);
358 	}
359 	/*
360 	 * Allocate all cylinder group blocks.
361 	 */
362 	for (cg = 0; cg < fs->fs_ncg; cg++) {
363 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
364 		    fs->fs_bsize, KERNCRED, 0, &nbp);
365 		if (error)
366 			goto out;
367 		bawrite(nbp);
368 		if (cg % 10 == 0)
369 			ffs_syncvnode(vp, MNT_WAIT, 0);
370 	}
371 	/*
372 	 * Copy all the cylinder group maps. Although the
373 	 * filesystem is still active, we hope that only a few
374 	 * cylinder groups will change between now and when we
375 	 * suspend operations. Thus, we will be able to quickly
376 	 * touch up the few cylinder groups that changed during
377 	 * the suspension period.
378 	 */
379 	len = howmany(fs->fs_ncg, NBBY);
380 	space = malloc(len, M_DEVBUF, M_WAITOK|M_ZERO);
381 	UFS_LOCK(ump);
382 	fs->fs_active = space;
383 	UFS_UNLOCK(ump);
384 	for (cg = 0; cg < fs->fs_ncg; cg++) {
385 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
386 		    fs->fs_bsize, KERNCRED, 0, &nbp);
387 		if (error)
388 			goto out;
389 		error = cgaccount(cg, vp, nbp, 1);
390 		bawrite(nbp);
391 		if (cg % 10 == 0)
392 			ffs_syncvnode(vp, MNT_WAIT, 0);
393 		if (error)
394 			goto out;
395 	}
396 	/*
397 	 * Change inode to snapshot type file.
398 	 */
399 	ip->i_flags |= SF_SNAPSHOT;
400 	DIP_SET(ip, i_flags, ip->i_flags);
401 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
402 	/*
403 	 * Ensure that the snapshot is completely on disk.
404 	 * Since we have marked it as a snapshot it is safe to
405 	 * unlock it as no process will be allowed to write to it.
406 	 */
407 	if ((error = ffs_syncvnode(vp, MNT_WAIT, 0)) != 0)
408 		goto out;
409 	VOP_UNLOCK(vp, 0);
410 	/*
411 	 * All allocations are done, so we can now snapshot the system.
412 	 *
413 	 * Recind nice scheduling while running with the filesystem suspended.
414 	 */
415 	if (td->td_proc->p_nice > 0) {
416 		struct proc *p;
417 
418 		p = td->td_proc;
419 		PROC_LOCK(p);
420 		saved_nice = p->p_nice;
421 		sched_nice(p, 0);
422 		PROC_UNLOCK(p);
423 	}
424 	/*
425 	 * Suspend operation on filesystem.
426 	 */
427 	for (;;) {
428 		vn_finished_write(wrtmp);
429 		if ((error = vfs_write_suspend(vp->v_mount, 0)) != 0) {
430 			vn_start_write(NULL, &wrtmp, V_WAIT);
431 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
432 			goto out;
433 		}
434 		if (mp->mnt_kern_flag & MNTK_SUSPENDED)
435 			break;
436 		vn_start_write(NULL, &wrtmp, V_WAIT);
437 	}
438 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
439 	if (ip->i_effnlink == 0) {
440 		error = ENOENT;		/* Snapshot file unlinked */
441 		goto out1;
442 	}
443 	if (collectsnapstats)
444 		nanotime(&starttime);
445 
446 	/* The last block might have changed.  Copy it again to be sure. */
447 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(numblks - 1)),
448 	    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
449 	if (error != 0)
450 		goto out1;
451 	error = readblock(vp, bp, numblks - 1);
452 	bp->b_flags |= B_VALIDSUSPWRT;
453 	bawrite(bp);
454 	if (error != 0)
455 		goto out1;
456 	/*
457 	 * First, copy all the cylinder group maps that have changed.
458 	 */
459 	for (cg = 0; cg < fs->fs_ncg; cg++) {
460 		if ((ACTIVECGNUM(fs, cg) & ACTIVECGOFF(cg)) != 0)
461 			continue;
462 		redo++;
463 		error = UFS_BALLOC(vp, lfragtosize(fs, cgtod(fs, cg)),
464 		    fs->fs_bsize, KERNCRED, 0, &nbp);
465 		if (error)
466 			goto out1;
467 		error = cgaccount(cg, vp, nbp, 2);
468 		bawrite(nbp);
469 		if (error)
470 			goto out1;
471 	}
472 	/*
473 	 * Grab a copy of the superblock and its summary information.
474 	 * We delay writing it until the suspension is released below.
475 	 */
476 	copy_fs = malloc((u_long)fs->fs_bsize, M_UFSMNT, M_WAITOK);
477 	bcopy(fs, copy_fs, fs->fs_sbsize);
478 	if ((fs->fs_flags & (FS_UNCLEAN | FS_NEEDSFSCK)) == 0)
479 		copy_fs->fs_clean = 1;
480 	size = fs->fs_bsize < SBLOCKSIZE ? fs->fs_bsize : SBLOCKSIZE;
481 	if (fs->fs_sbsize < size)
482 		bzero(&((char *)copy_fs)[fs->fs_sbsize],
483 		    size - fs->fs_sbsize);
484 	size = blkroundup(fs, fs->fs_cssize);
485 	if (fs->fs_contigsumsize > 0)
486 		size += fs->fs_ncg * sizeof(int32_t);
487 	space = malloc((u_long)size, M_UFSMNT, M_WAITOK);
488 	copy_fs->fs_csp = space;
489 	bcopy(fs->fs_csp, copy_fs->fs_csp, fs->fs_cssize);
490 	space = (char *)space + fs->fs_cssize;
491 	loc = howmany(fs->fs_cssize, fs->fs_fsize);
492 	i = fs->fs_frag - loc % fs->fs_frag;
493 	len = (i == fs->fs_frag) ? 0 : i * fs->fs_fsize;
494 	if (len > 0) {
495 		if ((error = bread(devvp, fsbtodb(fs, fs->fs_csaddr + loc),
496 		    len, KERNCRED, &bp)) != 0) {
497 			brelse(bp);
498 			free(copy_fs->fs_csp, M_UFSMNT);
499 			free(copy_fs, M_UFSMNT);
500 			copy_fs = NULL;
501 			goto out1;
502 		}
503 		bcopy(bp->b_data, space, (u_int)len);
504 		space = (char *)space + len;
505 		bp->b_flags |= B_INVAL | B_NOCACHE;
506 		brelse(bp);
507 	}
508 	if (fs->fs_contigsumsize > 0) {
509 		copy_fs->fs_maxcluster = lp = space;
510 		for (i = 0; i < fs->fs_ncg; i++)
511 			*lp++ = fs->fs_contigsumsize;
512 	}
513 	/*
514 	 * We must check for active files that have been unlinked
515 	 * (e.g., with a zero link count). We have to expunge all
516 	 * trace of these files from the snapshot so that they are
517 	 * not reclaimed prematurely by fsck or unnecessarily dumped.
518 	 * We turn off the MNTK_SUSPENDED flag to avoid a panic from
519 	 * spec_strategy about writing on a suspended filesystem.
520 	 * Note that we skip unlinked snapshot files as they will
521 	 * be handled separately below.
522 	 *
523 	 * We also calculate the needed size for the snapshot list.
524 	 */
525 	snaplistsize = fs->fs_ncg + howmany(fs->fs_cssize, fs->fs_bsize) +
526 	    FSMAXSNAP + 1 /* superblock */ + 1 /* last block */ + 1 /* size */;
527 	MNT_ILOCK(mp);
528 	mp->mnt_kern_flag &= ~MNTK_SUSPENDED;
529 	MNT_IUNLOCK(mp);
530 loop:
531 	MNT_VNODE_FOREACH_ALL(xvp, mp, mvp) {
532 		if ((xvp->v_usecount == 0 &&
533 		     (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) ||
534 		    xvp->v_type == VNON ||
535 		    IS_SNAPSHOT(VTOI(xvp))) {
536 			VI_UNLOCK(xvp);
537 			continue;
538 		}
539 		/*
540 		 * We can skip parent directory vnode because it must have
541 		 * this snapshot file in it.
542 		 */
543 		if (xvp == nd.ni_dvp) {
544 			VI_UNLOCK(xvp);
545 			continue;
546 		}
547 		vholdl(xvp);
548 		if (vn_lock(xvp, LK_EXCLUSIVE | LK_INTERLOCK) != 0) {
549 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
550 			vdrop(xvp);
551 			goto loop;
552 		}
553 		VI_LOCK(xvp);
554 		if (xvp->v_usecount == 0 &&
555 		    (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) {
556 			VI_UNLOCK(xvp);
557 			VOP_UNLOCK(xvp, 0);
558 			vdrop(xvp);
559 			continue;
560 		}
561 		VI_UNLOCK(xvp);
562 		if (snapdebug)
563 			vn_printf(xvp, "ffs_snapshot: busy vnode ");
564 		if (VOP_GETATTR(xvp, &vat, td->td_ucred) == 0 &&
565 		    vat.va_nlink > 0) {
566 			VOP_UNLOCK(xvp, 0);
567 			vdrop(xvp);
568 			continue;
569 		}
570 		xp = VTOI(xvp);
571 		if (ffs_checkfreefile(copy_fs, vp, xp->i_number)) {
572 			VOP_UNLOCK(xvp, 0);
573 			vdrop(xvp);
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 < UFS_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 				    xvp->v_type, NULL, SINGLETON_KEY);
587 				blkno = DIP(xp, i_db[loc]);
588 				DIP_SET(xp, i_db[loc], 0);
589 			}
590 		}
591 		snaplistsize += 1;
592 		if (I_IS_UFS1(xp))
593 			error = expunge_ufs1(vp, xp, copy_fs, fullacct_ufs1,
594 			    BLK_NOCOPY, 1);
595 		else
596 			error = expunge_ufs2(vp, xp, copy_fs, fullacct_ufs2,
597 			    BLK_NOCOPY, 1);
598 		if (blkno)
599 			DIP_SET(xp, i_db[loc], blkno);
600 		if (!error)
601 			error = ffs_freefile(ump, copy_fs, vp, xp->i_number,
602 			    xp->i_mode, NULL);
603 		VOP_UNLOCK(xvp, 0);
604 		vdrop(xvp);
605 		if (error) {
606 			free(copy_fs->fs_csp, M_UFSMNT);
607 			free(copy_fs, M_UFSMNT);
608 			copy_fs = NULL;
609 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
610 			goto out1;
611 		}
612 	}
613 	/*
614 	 * Erase the journal file from the snapshot.
615 	 */
616 	if (fs->fs_flags & FS_SUJ) {
617 		error = softdep_journal_lookup(mp, &xvp);
618 		if (error) {
619 			free(copy_fs->fs_csp, M_UFSMNT);
620 			free(copy_fs, M_UFSMNT);
621 			copy_fs = NULL;
622 			goto out1;
623 		}
624 		xp = VTOI(xvp);
625 		if (I_IS_UFS1(xp))
626 			error = expunge_ufs1(vp, xp, copy_fs, fullacct_ufs1,
627 			    BLK_NOCOPY, 0);
628 		else
629 			error = expunge_ufs2(vp, xp, copy_fs, fullacct_ufs2,
630 			    BLK_NOCOPY, 0);
631 		vput(xvp);
632 	}
633 	/*
634 	 * Acquire a lock on the snapdata structure, creating it if necessary.
635 	 */
636 	sn = ffs_snapdata_acquire(devvp);
637 	/*
638 	 * Change vnode to use shared snapshot lock instead of the original
639 	 * private lock.
640 	 */
641 	vp->v_vnlock = &sn->sn_lock;
642 	lockmgr(&vp->v_lock, LK_RELEASE, NULL);
643 	xp = TAILQ_FIRST(&sn->sn_head);
644 	/*
645 	 * If this is the first snapshot on this filesystem, then we need
646 	 * to allocate the space for the list of preallocated snapshot blocks.
647 	 * This list will be refined below, but this preliminary one will
648 	 * keep us out of deadlock until the full one is ready.
649 	 */
650 	if (xp == NULL) {
651 		snapblklist = malloc(snaplistsize * sizeof(daddr_t),
652 		    M_UFSMNT, M_WAITOK);
653 		blkp = &snapblklist[1];
654 		*blkp++ = lblkno(fs, fs->fs_sblockloc);
655 		blkno = fragstoblks(fs, fs->fs_csaddr);
656 		for (cg = 0; cg < fs->fs_ncg; cg++) {
657 			if (fragstoblks(fs, cgtod(fs, cg) > blkno))
658 				break;
659 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
660 		}
661 		len = howmany(fs->fs_cssize, fs->fs_bsize);
662 		for (loc = 0; loc < len; loc++)
663 			*blkp++ = blkno + loc;
664 		for (; cg < fs->fs_ncg; cg++)
665 			*blkp++ = fragstoblks(fs, cgtod(fs, cg));
666 		snapblklist[0] = blkp - snapblklist;
667 		VI_LOCK(devvp);
668 		if (sn->sn_blklist != NULL)
669 			panic("ffs_snapshot: non-empty list");
670 		sn->sn_blklist = snapblklist;
671 		sn->sn_listsize = blkp - snapblklist;
672 		VI_UNLOCK(devvp);
673 	}
674 	/*
675 	 * Record snapshot inode. Since this is the newest snapshot,
676 	 * it must be placed at the end of the list.
677 	 */
678 	VI_LOCK(devvp);
679 	fs->fs_snapinum[snaploc] = ip->i_number;
680 	if (ip->i_nextsnap.tqe_prev != 0)
681 		panic("ffs_snapshot: %ju already on list",
682 		    (uintmax_t)ip->i_number);
683 	TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
684 	devvp->v_vflag |= VV_COPYONWRITE;
685 	VI_UNLOCK(devvp);
686 	ASSERT_VOP_LOCKED(vp, "ffs_snapshot vp");
687 out1:
688 	KASSERT((sn != NULL && copy_fs != NULL && error == 0) ||
689 		(sn == NULL && copy_fs == NULL && error != 0),
690 		("email phk@ and mckusick@"));
691 	/*
692 	 * Resume operation on filesystem.
693 	 */
694 	vfs_write_resume(vp->v_mount, VR_START_WRITE | VR_NO_SUSPCLR);
695 	if (collectsnapstats && starttime.tv_sec > 0) {
696 		nanotime(&endtime);
697 		timespecsub(&endtime, &starttime, &endtime);
698 		printf("%s: suspended %ld.%03ld sec, redo %ld of %d\n",
699 		    vp->v_mount->mnt_stat.f_mntonname, (long)endtime.tv_sec,
700 		    endtime.tv_nsec / 1000000, redo, fs->fs_ncg);
701 	}
702 	if (copy_fs == NULL)
703 		goto out;
704 	/*
705 	 * Copy allocation information from all the snapshots in
706 	 * this snapshot and then expunge them from its view.
707 	 */
708 	TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap) {
709 		if (xp == ip)
710 			break;
711 		if (I_IS_UFS1(xp))
712 			error = expunge_ufs1(vp, xp, fs, snapacct_ufs1,
713 			    BLK_SNAP, 0);
714 		else
715 			error = expunge_ufs2(vp, xp, fs, snapacct_ufs2,
716 			    BLK_SNAP, 0);
717 		if (error == 0 && xp->i_effnlink == 0) {
718 			error = ffs_freefile(ump,
719 					     copy_fs,
720 					     vp,
721 					     xp->i_number,
722 					     xp->i_mode, NULL);
723 		}
724 		if (error) {
725 			fs->fs_snapinum[snaploc] = 0;
726 			goto done;
727 		}
728 	}
729 	/*
730 	 * Allocate space for the full list of preallocated snapshot blocks.
731 	 */
732 	snapblklist = malloc(snaplistsize * sizeof(daddr_t),
733 	    M_UFSMNT, M_WAITOK);
734 	ip->i_snapblklist = &snapblklist[1];
735 	/*
736 	 * Expunge the blocks used by the snapshots from the set of
737 	 * blocks marked as used in the snapshot bitmaps. Also, collect
738 	 * the list of allocated blocks in i_snapblklist.
739 	 */
740 	if (I_IS_UFS1(ip))
741 		error = expunge_ufs1(vp, ip, copy_fs, mapacct_ufs1,
742 		    BLK_SNAP, 0);
743 	else
744 		error = expunge_ufs2(vp, ip, copy_fs, mapacct_ufs2,
745 		    BLK_SNAP, 0);
746 	if (error) {
747 		fs->fs_snapinum[snaploc] = 0;
748 		free(snapblklist, M_UFSMNT);
749 		goto done;
750 	}
751 	if (snaplistsize < ip->i_snapblklist - snapblklist)
752 		panic("ffs_snapshot: list too small");
753 	snaplistsize = ip->i_snapblklist - snapblklist;
754 	snapblklist[0] = snaplistsize;
755 	ip->i_snapblklist = 0;
756 	/*
757 	 * Write out the list of allocated blocks to the end of the snapshot.
758 	 */
759 	auio.uio_iov = &aiov;
760 	auio.uio_iovcnt = 1;
761 	aiov.iov_base = (void *)snapblklist;
762 	aiov.iov_len = snaplistsize * sizeof(daddr_t);
763 	auio.uio_resid = aiov.iov_len;
764 	auio.uio_offset = ip->i_size;
765 	auio.uio_segflg = UIO_SYSSPACE;
766 	auio.uio_rw = UIO_WRITE;
767 	auio.uio_td = td;
768 	if ((error = VOP_WRITE(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
769 		fs->fs_snapinum[snaploc] = 0;
770 		free(snapblklist, M_UFSMNT);
771 		goto done;
772 	}
773 	/*
774 	 * Write the superblock and its summary information
775 	 * to the snapshot.
776 	 */
777 	blkno = fragstoblks(fs, fs->fs_csaddr);
778 	len = howmany(fs->fs_cssize, fs->fs_bsize);
779 	space = copy_fs->fs_csp;
780 	for (loc = 0; loc < len; loc++) {
781 		error = bread(vp, blkno + loc, fs->fs_bsize, KERNCRED, &nbp);
782 		if (error) {
783 			brelse(nbp);
784 			fs->fs_snapinum[snaploc] = 0;
785 			free(snapblklist, M_UFSMNT);
786 			goto done;
787 		}
788 		bcopy(space, nbp->b_data, fs->fs_bsize);
789 		space = (char *)space + fs->fs_bsize;
790 		bawrite(nbp);
791 	}
792 	error = bread(vp, lblkno(fs, fs->fs_sblockloc), fs->fs_bsize,
793 	    KERNCRED, &nbp);
794 	if (error) {
795 		brelse(nbp);
796 	} else {
797 		loc = blkoff(fs, fs->fs_sblockloc);
798 		copy_fs->fs_ckhash = ffs_calc_sbhash(copy_fs);
799 		bcopy((char *)copy_fs, &nbp->b_data[loc], (u_int)fs->fs_sbsize);
800 		bawrite(nbp);
801 	}
802 	/*
803 	 * As this is the newest list, it is the most inclusive, so
804 	 * should replace the previous list.
805 	 */
806 	VI_LOCK(devvp);
807 	space = sn->sn_blklist;
808 	sn->sn_blklist = snapblklist;
809 	sn->sn_listsize = snaplistsize;
810 	VI_UNLOCK(devvp);
811 	if (space != NULL)
812 		free(space, M_UFSMNT);
813 	/*
814 	 * Preallocate all the direct blocks in the snapshot inode so
815 	 * that we never have to write the inode itself to commit an
816 	 * update to the contents of the snapshot. Note that once
817 	 * created, the size of the snapshot will never change, so
818 	 * there will never be a need to write the inode except to
819 	 * update the non-integrity-critical time fields and
820 	 * allocated-block count.
821 	 */
822 	for (blockno = 0; blockno < UFS_NDADDR; blockno++) {
823 		if (DIP(ip, i_db[blockno]) != 0)
824 			continue;
825 		error = UFS_BALLOC(vp, lblktosize(fs, blockno),
826 		    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
827 		if (error)
828 			break;
829 		error = readblock(vp, bp, blockno);
830 		bawrite(bp);
831 		if (error != 0)
832 			break;
833 	}
834 done:
835 	free(copy_fs->fs_csp, M_UFSMNT);
836 	free(copy_fs, M_UFSMNT);
837 	copy_fs = NULL;
838 out:
839 	NDFREE(&nd, NDF_ONLY_PNBUF);
840 	if (saved_nice > 0) {
841 		struct proc *p;
842 
843 		p = td->td_proc;
844 		PROC_LOCK(p);
845 		sched_nice(td->td_proc, saved_nice);
846 		PROC_UNLOCK(td->td_proc);
847 	}
848 	UFS_LOCK(ump);
849 	if (fs->fs_active != 0) {
850 		free(fs->fs_active, M_DEVBUF);
851 		fs->fs_active = 0;
852 	}
853 	UFS_UNLOCK(ump);
854 	MNT_ILOCK(mp);
855 	mp->mnt_flag = (mp->mnt_flag & MNT_QUOTA) | (flag & ~MNT_QUOTA);
856 	MNT_IUNLOCK(mp);
857 	if (error)
858 		(void) ffs_truncate(vp, (off_t)0, 0, NOCRED);
859 	(void) ffs_syncvnode(vp, MNT_WAIT, 0);
860 	if (error)
861 		vput(vp);
862 	else
863 		VOP_UNLOCK(vp, 0);
864 	vrele(nd.ni_dvp);
865 	vn_finished_write(wrtmp);
866 	process_deferred_inactive(mp);
867 	return (error);
868 }
869 
870 /*
871  * Copy a cylinder group map. All the unallocated blocks are marked
872  * BLK_NOCOPY so that the snapshot knows that it need not copy them
873  * if they are later written. If passno is one, then this is a first
874  * pass, so only setting needs to be done. If passno is 2, then this
875  * is a revision to a previous pass which must be undone as the
876  * replacement pass is done.
877  */
878 static int
879 cgaccount(cg, vp, nbp, passno)
880 	int cg;
881 	struct vnode *vp;
882 	struct buf *nbp;
883 	int passno;
884 {
885 	struct buf *bp, *ibp;
886 	struct inode *ip;
887 	struct cg *cgp;
888 	struct fs *fs;
889 	ufs2_daddr_t base, numblks;
890 	int error, len, loc, indiroff;
891 
892 	ip = VTOI(vp);
893 	fs = ITOFS(ip);
894 	if ((error = ffs_getcg(fs, ITODEVVP(ip), cg, &bp, &cgp)) != 0)
895 		return (error);
896 	UFS_LOCK(ITOUMP(ip));
897 	ACTIVESET(fs, cg);
898 	/*
899 	 * Recomputation of summary information might not have been performed
900 	 * at mount time.  Sync up summary information for current cylinder
901 	 * group while data is in memory to ensure that result of background
902 	 * fsck is slightly more consistent.
903 	 */
904 	fs->fs_cs(fs, cg) = cgp->cg_cs;
905 	UFS_UNLOCK(ITOUMP(ip));
906 	bcopy(bp->b_data, nbp->b_data, fs->fs_cgsize);
907 	if (fs->fs_cgsize < fs->fs_bsize)
908 		bzero(&nbp->b_data[fs->fs_cgsize],
909 		    fs->fs_bsize - fs->fs_cgsize);
910 	cgp = (struct cg *)nbp->b_data;
911 	bqrelse(bp);
912 	if (passno == 2)
913 		nbp->b_flags |= B_VALIDSUSPWRT;
914 	numblks = howmany(fs->fs_size, fs->fs_frag);
915 	len = howmany(fs->fs_fpg, fs->fs_frag);
916 	base = cgbase(fs, cg) / fs->fs_frag;
917 	if (base + len >= numblks)
918 		len = numblks - base - 1;
919 	loc = 0;
920 	if (base < UFS_NDADDR) {
921 		for ( ; loc < UFS_NDADDR; loc++) {
922 			if (ffs_isblock(fs, cg_blksfree(cgp), loc))
923 				DIP_SET(ip, i_db[loc], BLK_NOCOPY);
924 			else if (passno == 2 && DIP(ip, i_db[loc])== BLK_NOCOPY)
925 				DIP_SET(ip, i_db[loc], 0);
926 			else if (passno == 1 && DIP(ip, i_db[loc])== BLK_NOCOPY)
927 				panic("ffs_snapshot: lost direct block");
928 		}
929 	}
930 	error = UFS_BALLOC(vp, lblktosize(fs, (off_t)(base + loc)),
931 	    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
932 	if (error) {
933 		goto out;
934 	}
935 	indiroff = (base + loc - UFS_NDADDR) % NINDIR(fs);
936 	for ( ; loc < len; loc++, indiroff++) {
937 		if (indiroff >= NINDIR(fs)) {
938 			if (passno == 2)
939 				ibp->b_flags |= B_VALIDSUSPWRT;
940 			bawrite(ibp);
941 			error = UFS_BALLOC(vp,
942 			    lblktosize(fs, (off_t)(base + loc)),
943 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
944 			if (error) {
945 				goto out;
946 			}
947 			indiroff = 0;
948 		}
949 		if (I_IS_UFS1(ip)) {
950 			if (ffs_isblock(fs, cg_blksfree(cgp), loc))
951 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] =
952 				    BLK_NOCOPY;
953 			else if (passno == 2 && ((ufs1_daddr_t *)(ibp->b_data))
954 			    [indiroff] == BLK_NOCOPY)
955 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] = 0;
956 			else if (passno == 1 && ((ufs1_daddr_t *)(ibp->b_data))
957 			    [indiroff] == BLK_NOCOPY)
958 				panic("ffs_snapshot: lost indirect block");
959 			continue;
960 		}
961 		if (ffs_isblock(fs, cg_blksfree(cgp), loc))
962 			((ufs2_daddr_t *)(ibp->b_data))[indiroff] = BLK_NOCOPY;
963 		else if (passno == 2 &&
964 		    ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY)
965 			((ufs2_daddr_t *)(ibp->b_data))[indiroff] = 0;
966 		else if (passno == 1 &&
967 		    ((ufs2_daddr_t *)(ibp->b_data)) [indiroff] == BLK_NOCOPY)
968 			panic("ffs_snapshot: lost indirect block");
969 	}
970 	if (passno == 2)
971 		ibp->b_flags |= B_VALIDSUSPWRT;
972 	bdwrite(ibp);
973 out:
974 	/*
975 	 * We have to calculate the crc32c here rather than just setting the
976 	 * BX_CYLGRP b_xflags because the allocation of the block for the
977 	 * the cylinder group map will always be a full size block (fs_bsize)
978 	 * even though the cylinder group may be smaller (fs_cgsize). The
979 	 * crc32c must be computed only over fs_cgsize whereas the BX_CYLGRP
980 	 * flag causes it to be computed over the size of the buffer.
981 	 */
982 	if ((fs->fs_metackhash & CK_CYLGRP) != 0) {
983 		((struct cg *)nbp->b_data)->cg_ckhash = 0;
984 		((struct cg *)nbp->b_data)->cg_ckhash =
985 		    calculate_crc32c(~0L, nbp->b_data, fs->fs_cgsize);
986 	}
987 	return (error);
988 }
989 
990 /*
991  * Before expunging a snapshot inode, note all the
992  * blocks that it claims with BLK_SNAP so that fsck will
993  * be able to account for those blocks properly and so
994  * that this snapshot knows that it need not copy them
995  * if the other snapshot holding them is freed. This code
996  * is reproduced once each for UFS1 and UFS2.
997  */
998 static int
999 expunge_ufs1(snapvp, cancelip, fs, acctfunc, expungetype, clearmode)
1000 	struct vnode *snapvp;
1001 	struct inode *cancelip;
1002 	struct fs *fs;
1003 	int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
1004 	    struct fs *, ufs_lbn_t, int);
1005 	int expungetype;
1006 	int clearmode;
1007 {
1008 	int i, error, indiroff;
1009 	ufs_lbn_t lbn, rlbn;
1010 	ufs2_daddr_t len, blkno, numblks, blksperindir;
1011 	struct ufs1_dinode *dip;
1012 	struct thread *td = curthread;
1013 	struct buf *bp;
1014 
1015 	/*
1016 	 * Prepare to expunge the inode. If its inode block has not
1017 	 * yet been copied, then allocate and fill the copy.
1018 	 */
1019 	lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number));
1020 	blkno = 0;
1021 	if (lbn < UFS_NDADDR) {
1022 		blkno = VTOI(snapvp)->i_din1->di_db[lbn];
1023 	} else {
1024 		if (DOINGSOFTDEP(snapvp))
1025 			softdep_prealloc(snapvp, MNT_WAIT);
1026 		td->td_pflags |= TDP_COWINPROGRESS;
1027 		error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn),
1028 		   fs->fs_bsize, KERNCRED, BA_METAONLY, &bp);
1029 		td->td_pflags &= ~TDP_COWINPROGRESS;
1030 		if (error)
1031 			return (error);
1032 		indiroff = (lbn - UFS_NDADDR) % NINDIR(fs);
1033 		blkno = ((ufs1_daddr_t *)(bp->b_data))[indiroff];
1034 		bqrelse(bp);
1035 	}
1036 	if (blkno != 0) {
1037 		if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp)))
1038 			return (error);
1039 	} else {
1040 		error = ffs_balloc_ufs1(snapvp, lblktosize(fs, (off_t)lbn),
1041 		    fs->fs_bsize, KERNCRED, 0, &bp);
1042 		if (error)
1043 			return (error);
1044 		if ((error = readblock(snapvp, bp, lbn)) != 0)
1045 			return (error);
1046 	}
1047 	/*
1048 	 * Set a snapshot inode to be a zero length file, regular files
1049 	 * or unlinked snapshots to be completely unallocated.
1050 	 */
1051 	dip = (struct ufs1_dinode *)bp->b_data +
1052 	    ino_to_fsbo(fs, cancelip->i_number);
1053 	if (clearmode || cancelip->i_effnlink == 0)
1054 		dip->di_mode = 0;
1055 	dip->di_size = 0;
1056 	dip->di_blocks = 0;
1057 	dip->di_flags &= ~SF_SNAPSHOT;
1058 	bzero(&dip->di_db[0], (UFS_NDADDR + UFS_NIADDR) * sizeof(ufs1_daddr_t));
1059 	bdwrite(bp);
1060 	/*
1061 	 * Now go through and expunge all the blocks in the file
1062 	 * using the function requested.
1063 	 */
1064 	numblks = howmany(cancelip->i_size, fs->fs_bsize);
1065 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_db[0],
1066 	    &cancelip->i_din1->di_db[UFS_NDADDR], fs, 0, expungetype)))
1067 		return (error);
1068 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din1->di_ib[0],
1069 	    &cancelip->i_din1->di_ib[UFS_NIADDR], fs, -1, expungetype)))
1070 		return (error);
1071 	blksperindir = 1;
1072 	lbn = -UFS_NDADDR;
1073 	len = numblks - UFS_NDADDR;
1074 	rlbn = UFS_NDADDR;
1075 	for (i = 0; len > 0 && i < UFS_NIADDR; i++) {
1076 		error = indiracct_ufs1(snapvp, ITOV(cancelip), i,
1077 		    cancelip->i_din1->di_ib[i], lbn, rlbn, len,
1078 		    blksperindir, fs, acctfunc, expungetype);
1079 		if (error)
1080 			return (error);
1081 		blksperindir *= NINDIR(fs);
1082 		lbn -= blksperindir + 1;
1083 		len -= blksperindir;
1084 		rlbn += blksperindir;
1085 	}
1086 	return (0);
1087 }
1088 
1089 /*
1090  * Descend an indirect block chain for vnode cancelvp accounting for all
1091  * its indirect blocks in snapvp.
1092  */
1093 static int
1094 indiracct_ufs1(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks,
1095 	    blksperindir, fs, acctfunc, expungetype)
1096 	struct vnode *snapvp;
1097 	struct vnode *cancelvp;
1098 	int level;
1099 	ufs1_daddr_t blkno;
1100 	ufs_lbn_t lbn;
1101 	ufs_lbn_t rlbn;
1102 	ufs_lbn_t remblks;
1103 	ufs_lbn_t blksperindir;
1104 	struct fs *fs;
1105 	int (*acctfunc)(struct vnode *, ufs1_daddr_t *, ufs1_daddr_t *,
1106 	    struct fs *, ufs_lbn_t, int);
1107 	int expungetype;
1108 {
1109 	int error, num, i;
1110 	ufs_lbn_t subblksperindir;
1111 	struct indir indirs[UFS_NIADDR + 2];
1112 	ufs1_daddr_t last, *bap;
1113 	struct buf *bp;
1114 
1115 	if (blkno == 0) {
1116 		if (expungetype == BLK_NOCOPY)
1117 			return (0);
1118 		panic("indiracct_ufs1: missing indir");
1119 	}
1120 	if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0)
1121 		return (error);
1122 	if (lbn != indirs[num - 1 - level].in_lbn || num < 2)
1123 		panic("indiracct_ufs1: botched params");
1124 	/*
1125 	 * We have to expand bread here since it will deadlock looking
1126 	 * up the block number for any blocks that are not in the cache.
1127 	 */
1128 	bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0);
1129 	bp->b_blkno = fsbtodb(fs, blkno);
1130 	if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0 &&
1131 	    (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) {
1132 		brelse(bp);
1133 		return (error);
1134 	}
1135 	/*
1136 	 * Account for the block pointers in this indirect block.
1137 	 */
1138 	last = howmany(remblks, blksperindir);
1139 	if (last > NINDIR(fs))
1140 		last = NINDIR(fs);
1141 	bap = malloc(fs->fs_bsize, M_DEVBUF, M_WAITOK);
1142 	bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize);
1143 	bqrelse(bp);
1144 	error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs,
1145 	    level == 0 ? rlbn : -1, expungetype);
1146 	if (error || level == 0)
1147 		goto out;
1148 	/*
1149 	 * Account for the block pointers in each of the indirect blocks
1150 	 * in the levels below us.
1151 	 */
1152 	subblksperindir = blksperindir / NINDIR(fs);
1153 	for (lbn++, level--, i = 0; i < last; i++) {
1154 		error = indiracct_ufs1(snapvp, cancelvp, level, bap[i], lbn,
1155 		    rlbn, remblks, subblksperindir, fs, acctfunc, expungetype);
1156 		if (error)
1157 			goto out;
1158 		rlbn += blksperindir;
1159 		lbn -= blksperindir;
1160 		remblks -= blksperindir;
1161 	}
1162 out:
1163 	free(bap, M_DEVBUF);
1164 	return (error);
1165 }
1166 
1167 /*
1168  * Do both snap accounting and map accounting.
1169  */
1170 static int
1171 fullacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype)
1172 	struct vnode *vp;
1173 	ufs1_daddr_t *oldblkp, *lastblkp;
1174 	struct fs *fs;
1175 	ufs_lbn_t lblkno;
1176 	int exptype;	/* BLK_SNAP or BLK_NOCOPY */
1177 {
1178 	int error;
1179 
1180 	if ((error = snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype)))
1181 		return (error);
1182 	return (mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, exptype));
1183 }
1184 
1185 /*
1186  * Identify a set of blocks allocated in a snapshot inode.
1187  */
1188 static int
1189 snapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1190 	struct vnode *vp;
1191 	ufs1_daddr_t *oldblkp, *lastblkp;
1192 	struct fs *fs;
1193 	ufs_lbn_t lblkno;
1194 	int expungetype;	/* BLK_SNAP or BLK_NOCOPY */
1195 {
1196 	struct inode *ip = VTOI(vp);
1197 	ufs1_daddr_t blkno, *blkp;
1198 	ufs_lbn_t lbn;
1199 	struct buf *ibp;
1200 	int error;
1201 
1202 	for ( ; oldblkp < lastblkp; oldblkp++) {
1203 		blkno = *oldblkp;
1204 		if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP)
1205 			continue;
1206 		lbn = fragstoblks(fs, blkno);
1207 		if (lbn < UFS_NDADDR) {
1208 			blkp = &ip->i_din1->di_db[lbn];
1209 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1210 		} else {
1211 			error = ffs_balloc_ufs1(vp, lblktosize(fs, (off_t)lbn),
1212 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1213 			if (error)
1214 				return (error);
1215 			blkp = &((ufs1_daddr_t *)(ibp->b_data))
1216 			    [(lbn - UFS_NDADDR) % NINDIR(fs)];
1217 		}
1218 		/*
1219 		 * If we are expunging a snapshot vnode and we
1220 		 * find a block marked BLK_NOCOPY, then it is
1221 		 * one that has been allocated to this snapshot after
1222 		 * we took our current snapshot and can be ignored.
1223 		 */
1224 		if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) {
1225 			if (lbn >= UFS_NDADDR)
1226 				brelse(ibp);
1227 		} else {
1228 			if (*blkp != 0)
1229 				panic("snapacct_ufs1: bad block");
1230 			*blkp = expungetype;
1231 			if (lbn >= UFS_NDADDR)
1232 				bdwrite(ibp);
1233 		}
1234 	}
1235 	return (0);
1236 }
1237 
1238 /*
1239  * Account for a set of blocks allocated in a snapshot inode.
1240  */
1241 static int
1242 mapacct_ufs1(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1243 	struct vnode *vp;
1244 	ufs1_daddr_t *oldblkp, *lastblkp;
1245 	struct fs *fs;
1246 	ufs_lbn_t lblkno;
1247 	int expungetype;
1248 {
1249 	ufs1_daddr_t blkno;
1250 	struct inode *ip;
1251 	ino_t inum;
1252 	int acctit;
1253 
1254 	ip = VTOI(vp);
1255 	inum = ip->i_number;
1256 	if (lblkno == -1)
1257 		acctit = 0;
1258 	else
1259 		acctit = 1;
1260 	for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) {
1261 		blkno = *oldblkp;
1262 		if (blkno == 0 || blkno == BLK_NOCOPY)
1263 			continue;
1264 		if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP)
1265 			*ip->i_snapblklist++ = lblkno;
1266 		if (blkno == BLK_SNAP)
1267 			blkno = blkstofrags(fs, lblkno);
1268 		ffs_blkfree(ITOUMP(ip), fs, vp, blkno, fs->fs_bsize, inum,
1269 		    vp->v_type, NULL, SINGLETON_KEY);
1270 	}
1271 	return (0);
1272 }
1273 
1274 /*
1275  * Before expunging a snapshot inode, note all the
1276  * blocks that it claims with BLK_SNAP so that fsck will
1277  * be able to account for those blocks properly and so
1278  * that this snapshot knows that it need not copy them
1279  * if the other snapshot holding them is freed. This code
1280  * is reproduced once each for UFS1 and UFS2.
1281  */
1282 static int
1283 expunge_ufs2(snapvp, cancelip, fs, acctfunc, expungetype, clearmode)
1284 	struct vnode *snapvp;
1285 	struct inode *cancelip;
1286 	struct fs *fs;
1287 	int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
1288 	    struct fs *, ufs_lbn_t, int);
1289 	int expungetype;
1290 	int clearmode;
1291 {
1292 	int i, error, indiroff;
1293 	ufs_lbn_t lbn, rlbn;
1294 	ufs2_daddr_t len, blkno, numblks, blksperindir;
1295 	struct ufs2_dinode *dip;
1296 	struct thread *td = curthread;
1297 	struct buf *bp;
1298 
1299 	/*
1300 	 * Prepare to expunge the inode. If its inode block has not
1301 	 * yet been copied, then allocate and fill the copy.
1302 	 */
1303 	lbn = fragstoblks(fs, ino_to_fsba(fs, cancelip->i_number));
1304 	blkno = 0;
1305 	if (lbn < UFS_NDADDR) {
1306 		blkno = VTOI(snapvp)->i_din2->di_db[lbn];
1307 	} else {
1308 		if (DOINGSOFTDEP(snapvp))
1309 			softdep_prealloc(snapvp, MNT_WAIT);
1310 		td->td_pflags |= TDP_COWINPROGRESS;
1311 		error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn),
1312 		   fs->fs_bsize, KERNCRED, BA_METAONLY, &bp);
1313 		td->td_pflags &= ~TDP_COWINPROGRESS;
1314 		if (error)
1315 			return (error);
1316 		indiroff = (lbn - UFS_NDADDR) % NINDIR(fs);
1317 		blkno = ((ufs2_daddr_t *)(bp->b_data))[indiroff];
1318 		bqrelse(bp);
1319 	}
1320 	if (blkno != 0) {
1321 		if ((error = bread(snapvp, lbn, fs->fs_bsize, KERNCRED, &bp)))
1322 			return (error);
1323 	} else {
1324 		error = ffs_balloc_ufs2(snapvp, lblktosize(fs, (off_t)lbn),
1325 		    fs->fs_bsize, KERNCRED, 0, &bp);
1326 		if (error)
1327 			return (error);
1328 		if ((error = readblock(snapvp, bp, lbn)) != 0)
1329 			return (error);
1330 	}
1331 	/*
1332 	 * Set a snapshot inode to be a zero length file, regular files
1333 	 * to be completely unallocated.
1334 	 */
1335 	dip = (struct ufs2_dinode *)bp->b_data +
1336 	    ino_to_fsbo(fs, cancelip->i_number);
1337 	dip->di_size = 0;
1338 	dip->di_blocks = 0;
1339 	dip->di_flags &= ~SF_SNAPSHOT;
1340 	bzero(&dip->di_db[0], (UFS_NDADDR + UFS_NIADDR) * sizeof(ufs2_daddr_t));
1341 	if (clearmode || cancelip->i_effnlink == 0)
1342 		dip->di_mode = 0;
1343 	else
1344 		ffs_update_dinode_ckhash(fs, dip);
1345 	bdwrite(bp);
1346 	/*
1347 	 * Now go through and expunge all the blocks in the file
1348 	 * using the function requested.
1349 	 */
1350 	numblks = howmany(cancelip->i_size, fs->fs_bsize);
1351 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_db[0],
1352 	    &cancelip->i_din2->di_db[UFS_NDADDR], fs, 0, expungetype)))
1353 		return (error);
1354 	if ((error = (*acctfunc)(snapvp, &cancelip->i_din2->di_ib[0],
1355 	    &cancelip->i_din2->di_ib[UFS_NIADDR], fs, -1, expungetype)))
1356 		return (error);
1357 	blksperindir = 1;
1358 	lbn = -UFS_NDADDR;
1359 	len = numblks - UFS_NDADDR;
1360 	rlbn = UFS_NDADDR;
1361 	for (i = 0; len > 0 && i < UFS_NIADDR; i++) {
1362 		error = indiracct_ufs2(snapvp, ITOV(cancelip), i,
1363 		    cancelip->i_din2->di_ib[i], lbn, rlbn, len,
1364 		    blksperindir, fs, acctfunc, expungetype);
1365 		if (error)
1366 			return (error);
1367 		blksperindir *= NINDIR(fs);
1368 		lbn -= blksperindir + 1;
1369 		len -= blksperindir;
1370 		rlbn += blksperindir;
1371 	}
1372 	return (0);
1373 }
1374 
1375 /*
1376  * Descend an indirect block chain for vnode cancelvp accounting for all
1377  * its indirect blocks in snapvp.
1378  */
1379 static int
1380 indiracct_ufs2(snapvp, cancelvp, level, blkno, lbn, rlbn, remblks,
1381 	    blksperindir, fs, acctfunc, expungetype)
1382 	struct vnode *snapvp;
1383 	struct vnode *cancelvp;
1384 	int level;
1385 	ufs2_daddr_t blkno;
1386 	ufs_lbn_t lbn;
1387 	ufs_lbn_t rlbn;
1388 	ufs_lbn_t remblks;
1389 	ufs_lbn_t blksperindir;
1390 	struct fs *fs;
1391 	int (*acctfunc)(struct vnode *, ufs2_daddr_t *, ufs2_daddr_t *,
1392 	    struct fs *, ufs_lbn_t, int);
1393 	int expungetype;
1394 {
1395 	int error, num, i;
1396 	ufs_lbn_t subblksperindir;
1397 	struct indir indirs[UFS_NIADDR + 2];
1398 	ufs2_daddr_t last, *bap;
1399 	struct buf *bp;
1400 
1401 	if (blkno == 0) {
1402 		if (expungetype == BLK_NOCOPY)
1403 			return (0);
1404 		panic("indiracct_ufs2: missing indir");
1405 	}
1406 	if ((error = ufs_getlbns(cancelvp, rlbn, indirs, &num)) != 0)
1407 		return (error);
1408 	if (lbn != indirs[num - 1 - level].in_lbn || num < 2)
1409 		panic("indiracct_ufs2: botched params");
1410 	/*
1411 	 * We have to expand bread here since it will deadlock looking
1412 	 * up the block number for any blocks that are not in the cache.
1413 	 */
1414 	bp = getblk(cancelvp, lbn, fs->fs_bsize, 0, 0, 0);
1415 	bp->b_blkno = fsbtodb(fs, blkno);
1416 	if ((bp->b_flags & B_CACHE) == 0 &&
1417 	    (error = readblock(cancelvp, bp, fragstoblks(fs, blkno)))) {
1418 		brelse(bp);
1419 		return (error);
1420 	}
1421 	/*
1422 	 * Account for the block pointers in this indirect block.
1423 	 */
1424 	last = howmany(remblks, blksperindir);
1425 	if (last > NINDIR(fs))
1426 		last = NINDIR(fs);
1427 	bap = malloc(fs->fs_bsize, M_DEVBUF, M_WAITOK);
1428 	bcopy(bp->b_data, (caddr_t)bap, fs->fs_bsize);
1429 	bqrelse(bp);
1430 	error = (*acctfunc)(snapvp, &bap[0], &bap[last], fs,
1431 	    level == 0 ? rlbn : -1, expungetype);
1432 	if (error || level == 0)
1433 		goto out;
1434 	/*
1435 	 * Account for the block pointers in each of the indirect blocks
1436 	 * in the levels below us.
1437 	 */
1438 	subblksperindir = blksperindir / NINDIR(fs);
1439 	for (lbn++, level--, i = 0; i < last; i++) {
1440 		error = indiracct_ufs2(snapvp, cancelvp, level, bap[i], lbn,
1441 		    rlbn, remblks, subblksperindir, fs, acctfunc, expungetype);
1442 		if (error)
1443 			goto out;
1444 		rlbn += blksperindir;
1445 		lbn -= blksperindir;
1446 		remblks -= blksperindir;
1447 	}
1448 out:
1449 	free(bap, M_DEVBUF);
1450 	return (error);
1451 }
1452 
1453 /*
1454  * Do both snap accounting and map accounting.
1455  */
1456 static int
1457 fullacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype)
1458 	struct vnode *vp;
1459 	ufs2_daddr_t *oldblkp, *lastblkp;
1460 	struct fs *fs;
1461 	ufs_lbn_t lblkno;
1462 	int exptype;	/* BLK_SNAP or BLK_NOCOPY */
1463 {
1464 	int error;
1465 
1466 	if ((error = snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype)))
1467 		return (error);
1468 	return (mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, exptype));
1469 }
1470 
1471 /*
1472  * Identify a set of blocks allocated in a snapshot inode.
1473  */
1474 static int
1475 snapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1476 	struct vnode *vp;
1477 	ufs2_daddr_t *oldblkp, *lastblkp;
1478 	struct fs *fs;
1479 	ufs_lbn_t lblkno;
1480 	int expungetype;	/* BLK_SNAP or BLK_NOCOPY */
1481 {
1482 	struct inode *ip = VTOI(vp);
1483 	ufs2_daddr_t blkno, *blkp;
1484 	ufs_lbn_t lbn;
1485 	struct buf *ibp;
1486 	int error;
1487 
1488 	for ( ; oldblkp < lastblkp; oldblkp++) {
1489 		blkno = *oldblkp;
1490 		if (blkno == 0 || blkno == BLK_NOCOPY || blkno == BLK_SNAP)
1491 			continue;
1492 		lbn = fragstoblks(fs, blkno);
1493 		if (lbn < UFS_NDADDR) {
1494 			blkp = &ip->i_din2->di_db[lbn];
1495 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1496 		} else {
1497 			error = ffs_balloc_ufs2(vp, lblktosize(fs, (off_t)lbn),
1498 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1499 			if (error)
1500 				return (error);
1501 			blkp = &((ufs2_daddr_t *)(ibp->b_data))
1502 			    [(lbn - UFS_NDADDR) % NINDIR(fs)];
1503 		}
1504 		/*
1505 		 * If we are expunging a snapshot vnode and we
1506 		 * find a block marked BLK_NOCOPY, then it is
1507 		 * one that has been allocated to this snapshot after
1508 		 * we took our current snapshot and can be ignored.
1509 		 */
1510 		if (expungetype == BLK_SNAP && *blkp == BLK_NOCOPY) {
1511 			if (lbn >= UFS_NDADDR)
1512 				brelse(ibp);
1513 		} else {
1514 			if (*blkp != 0)
1515 				panic("snapacct_ufs2: bad block");
1516 			*blkp = expungetype;
1517 			if (lbn >= UFS_NDADDR)
1518 				bdwrite(ibp);
1519 		}
1520 	}
1521 	return (0);
1522 }
1523 
1524 /*
1525  * Account for a set of blocks allocated in a snapshot inode.
1526  */
1527 static int
1528 mapacct_ufs2(vp, oldblkp, lastblkp, fs, lblkno, expungetype)
1529 	struct vnode *vp;
1530 	ufs2_daddr_t *oldblkp, *lastblkp;
1531 	struct fs *fs;
1532 	ufs_lbn_t lblkno;
1533 	int expungetype;
1534 {
1535 	ufs2_daddr_t blkno;
1536 	struct inode *ip;
1537 	ino_t inum;
1538 	int acctit;
1539 
1540 	ip = VTOI(vp);
1541 	inum = ip->i_number;
1542 	if (lblkno == -1)
1543 		acctit = 0;
1544 	else
1545 		acctit = 1;
1546 	for ( ; oldblkp < lastblkp; oldblkp++, lblkno++) {
1547 		blkno = *oldblkp;
1548 		if (blkno == 0 || blkno == BLK_NOCOPY)
1549 			continue;
1550 		if (acctit && expungetype == BLK_SNAP && blkno != BLK_SNAP)
1551 			*ip->i_snapblklist++ = lblkno;
1552 		if (blkno == BLK_SNAP)
1553 			blkno = blkstofrags(fs, lblkno);
1554 		ffs_blkfree(ITOUMP(ip), fs, vp, blkno, fs->fs_bsize, inum,
1555 		    vp->v_type, NULL, SINGLETON_KEY);
1556 	}
1557 	return (0);
1558 }
1559 
1560 /*
1561  * Decrement extra reference on snapshot when last name is removed.
1562  * It will not be freed until the last open reference goes away.
1563  */
1564 void
1565 ffs_snapgone(ip)
1566 	struct inode *ip;
1567 {
1568 	struct inode *xp;
1569 	struct fs *fs;
1570 	int snaploc;
1571 	struct snapdata *sn;
1572 	struct ufsmount *ump;
1573 
1574 	/*
1575 	 * Find snapshot in incore list.
1576 	 */
1577 	xp = NULL;
1578 	sn = ITODEVVP(ip)->v_rdev->si_snapdata;
1579 	if (sn != NULL)
1580 		TAILQ_FOREACH(xp, &sn->sn_head, i_nextsnap)
1581 			if (xp == ip)
1582 				break;
1583 	if (xp != NULL)
1584 		vrele(ITOV(ip));
1585 	else if (snapdebug)
1586 		printf("ffs_snapgone: lost snapshot vnode %ju\n",
1587 		    (uintmax_t)ip->i_number);
1588 	/*
1589 	 * Delete snapshot inode from superblock. Keep list dense.
1590 	 */
1591 	ump = ITOUMP(ip);
1592 	fs = ump->um_fs;
1593 	UFS_LOCK(ump);
1594 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++)
1595 		if (fs->fs_snapinum[snaploc] == ip->i_number)
1596 			break;
1597 	if (snaploc < FSMAXSNAP) {
1598 		for (snaploc++; snaploc < FSMAXSNAP; snaploc++) {
1599 			if (fs->fs_snapinum[snaploc] == 0)
1600 				break;
1601 			fs->fs_snapinum[snaploc - 1] = fs->fs_snapinum[snaploc];
1602 		}
1603 		fs->fs_snapinum[snaploc - 1] = 0;
1604 	}
1605 	UFS_UNLOCK(ump);
1606 }
1607 
1608 /*
1609  * Prepare a snapshot file for being removed.
1610  */
1611 void
1612 ffs_snapremove(vp)
1613 	struct vnode *vp;
1614 {
1615 	struct inode *ip;
1616 	struct vnode *devvp;
1617 	struct buf *ibp;
1618 	struct fs *fs;
1619 	ufs2_daddr_t numblks, blkno, dblk;
1620 	int error, i, last, loc;
1621 	struct snapdata *sn;
1622 
1623 	ip = VTOI(vp);
1624 	fs = ITOFS(ip);
1625 	devvp = ITODEVVP(ip);
1626 	/*
1627 	 * If active, delete from incore list (this snapshot may
1628 	 * already have been in the process of being deleted, so
1629 	 * would not have been active).
1630 	 *
1631 	 * Clear copy-on-write flag if last snapshot.
1632 	 */
1633 	VI_LOCK(devvp);
1634 	if (ip->i_nextsnap.tqe_prev != 0) {
1635 		sn = devvp->v_rdev->si_snapdata;
1636 		TAILQ_REMOVE(&sn->sn_head, ip, i_nextsnap);
1637 		ip->i_nextsnap.tqe_prev = 0;
1638 		VI_UNLOCK(devvp);
1639 		lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
1640 		for (i = 0; i < sn->sn_lock.lk_recurse; i++)
1641 			lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
1642 		KASSERT(vp->v_vnlock == &sn->sn_lock,
1643 			("ffs_snapremove: lost lock mutation"));
1644 		vp->v_vnlock = &vp->v_lock;
1645 		VI_LOCK(devvp);
1646 		while (sn->sn_lock.lk_recurse > 0)
1647 			lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
1648 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
1649 		try_free_snapdata(devvp);
1650 	} else
1651 		VI_UNLOCK(devvp);
1652 	/*
1653 	 * Clear all BLK_NOCOPY fields. Pass any block claims to other
1654 	 * snapshots that want them (see ffs_snapblkfree below).
1655 	 */
1656 	for (blkno = 1; blkno < UFS_NDADDR; blkno++) {
1657 		dblk = DIP(ip, i_db[blkno]);
1658 		if (dblk == 0)
1659 			continue;
1660 		if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1661 			DIP_SET(ip, i_db[blkno], 0);
1662 		else if ((dblk == blkstofrags(fs, blkno) &&
1663 		     ffs_snapblkfree(fs, ITODEVVP(ip), dblk, fs->fs_bsize,
1664 		     ip->i_number, vp->v_type, NULL))) {
1665 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) -
1666 			    btodb(fs->fs_bsize));
1667 			DIP_SET(ip, i_db[blkno], 0);
1668 		}
1669 	}
1670 	numblks = howmany(ip->i_size, fs->fs_bsize);
1671 	for (blkno = UFS_NDADDR; blkno < numblks; blkno += NINDIR(fs)) {
1672 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)blkno),
1673 		    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1674 		if (error)
1675 			continue;
1676 		if (fs->fs_size - blkno > NINDIR(fs))
1677 			last = NINDIR(fs);
1678 		else
1679 			last = fs->fs_size - blkno;
1680 		for (loc = 0; loc < last; loc++) {
1681 			if (I_IS_UFS1(ip)) {
1682 				dblk = ((ufs1_daddr_t *)(ibp->b_data))[loc];
1683 				if (dblk == 0)
1684 					continue;
1685 				if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1686 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1687 				else if ((dblk == blkstofrags(fs, blkno) &&
1688 				     ffs_snapblkfree(fs, ITODEVVP(ip), dblk,
1689 				     fs->fs_bsize, ip->i_number, vp->v_type,
1690 				     NULL))) {
1691 					ip->i_din1->di_blocks -=
1692 					    btodb(fs->fs_bsize);
1693 					((ufs1_daddr_t *)(ibp->b_data))[loc]= 0;
1694 				}
1695 				continue;
1696 			}
1697 			dblk = ((ufs2_daddr_t *)(ibp->b_data))[loc];
1698 			if (dblk == 0)
1699 				continue;
1700 			if (dblk == BLK_NOCOPY || dblk == BLK_SNAP)
1701 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1702 			else if ((dblk == blkstofrags(fs, blkno) &&
1703 			     ffs_snapblkfree(fs, ITODEVVP(ip), dblk,
1704 			     fs->fs_bsize, ip->i_number, vp->v_type, NULL))) {
1705 				ip->i_din2->di_blocks -= btodb(fs->fs_bsize);
1706 				((ufs2_daddr_t *)(ibp->b_data))[loc] = 0;
1707 			}
1708 		}
1709 		bawrite(ibp);
1710 	}
1711 	/*
1712 	 * Clear snapshot flag and drop reference.
1713 	 */
1714 	ip->i_flags &= ~SF_SNAPSHOT;
1715 	DIP_SET(ip, i_flags, ip->i_flags);
1716 	ip->i_flag |= IN_CHANGE | IN_UPDATE;
1717 	/*
1718 	 * The dirtied indirects must be written out before
1719 	 * softdep_setup_freeblocks() is called.  Otherwise indir_trunc()
1720 	 * may find indirect pointers using the magic BLK_* values.
1721 	 */
1722 	if (DOINGSOFTDEP(vp))
1723 		ffs_syncvnode(vp, MNT_WAIT, 0);
1724 #ifdef QUOTA
1725 	/*
1726 	 * Reenable disk quotas for ex-snapshot file.
1727 	 */
1728 	if (!getinoquota(ip))
1729 		(void) chkdq(ip, DIP(ip, i_blocks), KERNCRED, FORCE);
1730 #endif
1731 }
1732 
1733 /*
1734  * Notification that a block is being freed. Return zero if the free
1735  * should be allowed to proceed. Return non-zero if the snapshot file
1736  * wants to claim the block. The block will be claimed if it is an
1737  * uncopied part of one of the snapshots. It will be freed if it is
1738  * either a BLK_NOCOPY or has already been copied in all of the snapshots.
1739  * If a fragment is being freed, then all snapshots that care about
1740  * it must make a copy since a snapshot file can only claim full sized
1741  * blocks. Note that if more than one snapshot file maps the block,
1742  * we can pick one at random to claim it. Since none of the snapshots
1743  * can change, we are assurred that they will all see the same unmodified
1744  * image. When deleting a snapshot file (see ffs_snapremove above), we
1745  * must push any of these claimed blocks to one of the other snapshots
1746  * that maps it. These claimed blocks are easily identified as they will
1747  * have a block number equal to their logical block number within the
1748  * snapshot. A copied block can never have this property because they
1749  * must always have been allocated from a BLK_NOCOPY location.
1750  */
1751 int
1752 ffs_snapblkfree(fs, devvp, bno, size, inum, vtype, wkhd)
1753 	struct fs *fs;
1754 	struct vnode *devvp;
1755 	ufs2_daddr_t bno;
1756 	long size;
1757 	ino_t inum;
1758 	enum vtype vtype;
1759 	struct workhead *wkhd;
1760 {
1761 	struct buf *ibp, *cbp, *savedcbp = NULL;
1762 	struct thread *td = curthread;
1763 	struct inode *ip;
1764 	struct vnode *vp = NULL;
1765 	ufs_lbn_t lbn;
1766 	ufs2_daddr_t blkno;
1767 	int indiroff = 0, error = 0, claimedblk = 0;
1768 	struct snapdata *sn;
1769 
1770 	lbn = fragstoblks(fs, bno);
1771 retry:
1772 	VI_LOCK(devvp);
1773 	sn = devvp->v_rdev->si_snapdata;
1774 	if (sn == NULL) {
1775 		VI_UNLOCK(devvp);
1776 		return (0);
1777 	}
1778 	if (lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
1779 	    VI_MTX(devvp)) != 0)
1780 		goto retry;
1781 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
1782 		vp = ITOV(ip);
1783 		if (DOINGSOFTDEP(vp))
1784 			softdep_prealloc(vp, MNT_WAIT);
1785 		/*
1786 		 * Lookup block being written.
1787 		 */
1788 		if (lbn < UFS_NDADDR) {
1789 			blkno = DIP(ip, i_db[lbn]);
1790 		} else {
1791 			td->td_pflags |= TDP_COWINPROGRESS;
1792 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1793 			    fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
1794 			td->td_pflags &= ~TDP_COWINPROGRESS;
1795 			if (error)
1796 				break;
1797 			indiroff = (lbn - UFS_NDADDR) % NINDIR(fs);
1798 			if (I_IS_UFS1(ip))
1799 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
1800 			else
1801 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
1802 		}
1803 		/*
1804 		 * Check to see if block needs to be copied.
1805 		 */
1806 		if (blkno == 0) {
1807 			/*
1808 			 * A block that we map is being freed. If it has not
1809 			 * been claimed yet, we will claim or copy it (below).
1810 			 */
1811 			claimedblk = 1;
1812 		} else if (blkno == BLK_SNAP) {
1813 			/*
1814 			 * No previous snapshot claimed the block,
1815 			 * so it will be freed and become a BLK_NOCOPY
1816 			 * (don't care) for us.
1817 			 */
1818 			if (claimedblk)
1819 				panic("snapblkfree: inconsistent block type");
1820 			if (lbn < UFS_NDADDR) {
1821 				DIP_SET(ip, i_db[lbn], BLK_NOCOPY);
1822 				ip->i_flag |= IN_CHANGE | IN_UPDATE;
1823 			} else if (I_IS_UFS1(ip)) {
1824 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] =
1825 				    BLK_NOCOPY;
1826 				bdwrite(ibp);
1827 			} else {
1828 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] =
1829 				    BLK_NOCOPY;
1830 				bdwrite(ibp);
1831 			}
1832 			continue;
1833 		} else /* BLK_NOCOPY or default */ {
1834 			/*
1835 			 * If the snapshot has already copied the block
1836 			 * (default), or does not care about the block,
1837 			 * it is not needed.
1838 			 */
1839 			if (lbn >= UFS_NDADDR)
1840 				bqrelse(ibp);
1841 			continue;
1842 		}
1843 		/*
1844 		 * If this is a full size block, we will just grab it
1845 		 * and assign it to the snapshot inode. Otherwise we
1846 		 * will proceed to copy it. See explanation for this
1847 		 * routine as to why only a single snapshot needs to
1848 		 * claim this block.
1849 		 */
1850 		if (size == fs->fs_bsize) {
1851 #ifdef DEBUG
1852 			if (snapdebug)
1853 				printf("%s %ju lbn %jd from inum %ju\n",
1854 				    "Grabonremove: snapino",
1855 				    (uintmax_t)ip->i_number,
1856 				    (intmax_t)lbn, (uintmax_t)inum);
1857 #endif
1858 			/*
1859 			 * If journaling is tracking this write we must add
1860 			 * the work to the inode or indirect being written.
1861 			 */
1862 			if (wkhd != NULL) {
1863 				if (lbn < UFS_NDADDR)
1864 					softdep_inode_append(ip,
1865 					    curthread->td_ucred, wkhd);
1866 				else
1867 					softdep_buf_append(ibp, wkhd);
1868 			}
1869 			if (lbn < UFS_NDADDR) {
1870 				DIP_SET(ip, i_db[lbn], bno);
1871 			} else if (I_IS_UFS1(ip)) {
1872 				((ufs1_daddr_t *)(ibp->b_data))[indiroff] = bno;
1873 				bdwrite(ibp);
1874 			} else {
1875 				((ufs2_daddr_t *)(ibp->b_data))[indiroff] = bno;
1876 				bdwrite(ibp);
1877 			}
1878 			DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + btodb(size));
1879 			ip->i_flag |= IN_CHANGE | IN_UPDATE;
1880 			lockmgr(vp->v_vnlock, LK_RELEASE, NULL);
1881 			return (1);
1882 		}
1883 		if (lbn >= UFS_NDADDR)
1884 			bqrelse(ibp);
1885 		/*
1886 		 * Allocate the block into which to do the copy. Note that this
1887 		 * allocation will never require any additional allocations for
1888 		 * the snapshot inode.
1889 		 */
1890 		td->td_pflags |= TDP_COWINPROGRESS;
1891 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
1892 		    fs->fs_bsize, KERNCRED, 0, &cbp);
1893 		td->td_pflags &= ~TDP_COWINPROGRESS;
1894 		if (error)
1895 			break;
1896 #ifdef DEBUG
1897 		if (snapdebug)
1898 			printf("%s%ju lbn %jd %s %ju size %ld to blkno %jd\n",
1899 			    "Copyonremove: snapino ", (uintmax_t)ip->i_number,
1900 			    (intmax_t)lbn, "for inum", (uintmax_t)inum, size,
1901 			    (intmax_t)cbp->b_blkno);
1902 #endif
1903 		/*
1904 		 * If we have already read the old block contents, then
1905 		 * simply copy them to the new block. Note that we need
1906 		 * to synchronously write snapshots that have not been
1907 		 * unlinked, and hence will be visible after a crash,
1908 		 * to ensure their integrity. At a minimum we ensure the
1909 		 * integrity of the filesystem metadata, but use the
1910 		 * dopersistence sysctl-setable flag to decide on the
1911 		 * persistence needed for file content data.
1912 		 */
1913 		if (savedcbp != NULL) {
1914 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
1915 			bawrite(cbp);
1916 			if ((vtype == VDIR || dopersistence) &&
1917 			    ip->i_effnlink > 0)
1918 				(void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT);
1919 			continue;
1920 		}
1921 		/*
1922 		 * Otherwise, read the old block contents into the buffer.
1923 		 */
1924 		if ((error = readblock(vp, cbp, lbn)) != 0) {
1925 			bzero(cbp->b_data, fs->fs_bsize);
1926 			bawrite(cbp);
1927 			if ((vtype == VDIR || dopersistence) &&
1928 			    ip->i_effnlink > 0)
1929 				(void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT);
1930 			break;
1931 		}
1932 		savedcbp = cbp;
1933 	}
1934 	/*
1935 	 * Note that we need to synchronously write snapshots that
1936 	 * have not been unlinked, and hence will be visible after
1937 	 * a crash, to ensure their integrity. At a minimum we
1938 	 * ensure the integrity of the filesystem metadata, but
1939 	 * use the dopersistence sysctl-setable flag to decide on
1940 	 * the persistence needed for file content data.
1941 	 */
1942 	if (savedcbp) {
1943 		vp = savedcbp->b_vp;
1944 		bawrite(savedcbp);
1945 		if ((vtype == VDIR || dopersistence) &&
1946 		    VTOI(vp)->i_effnlink > 0)
1947 			(void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT);
1948 	}
1949 	/*
1950 	 * If we have been unable to allocate a block in which to do
1951 	 * the copy, then return non-zero so that the fragment will
1952 	 * not be freed. Although space will be lost, the snapshot
1953 	 * will stay consistent.
1954 	 */
1955 	if (error != 0 && wkhd != NULL)
1956 		softdep_freework(wkhd);
1957 	lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
1958 	return (error);
1959 }
1960 
1961 /*
1962  * Associate snapshot files when mounting.
1963  */
1964 void
1965 ffs_snapshot_mount(mp)
1966 	struct mount *mp;
1967 {
1968 	struct ufsmount *ump = VFSTOUFS(mp);
1969 	struct vnode *devvp = ump->um_devvp;
1970 	struct fs *fs = ump->um_fs;
1971 	struct thread *td = curthread;
1972 	struct snapdata *sn;
1973 	struct vnode *vp;
1974 	struct vnode *lastvp;
1975 	struct inode *ip;
1976 	struct uio auio;
1977 	struct iovec aiov;
1978 	void *snapblklist;
1979 	char *reason;
1980 	daddr_t snaplistsize;
1981 	int error, snaploc, loc;
1982 
1983 	/*
1984 	 * XXX The following needs to be set before ffs_truncate or
1985 	 * VOP_READ can be called.
1986 	 */
1987 	mp->mnt_stat.f_iosize = fs->fs_bsize;
1988 	/*
1989 	 * Process each snapshot listed in the superblock.
1990 	 */
1991 	vp = NULL;
1992 	lastvp = NULL;
1993 	sn = NULL;
1994 	for (snaploc = 0; snaploc < FSMAXSNAP; snaploc++) {
1995 		if (fs->fs_snapinum[snaploc] == 0)
1996 			break;
1997 		if ((error = ffs_vget(mp, fs->fs_snapinum[snaploc],
1998 		    LK_EXCLUSIVE, &vp)) != 0){
1999 			printf("ffs_snapshot_mount: vget failed %d\n", error);
2000 			continue;
2001 		}
2002 		ip = VTOI(vp);
2003 		if (!IS_SNAPSHOT(ip) || ip->i_size ==
2004 		    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag))) {
2005 			if (!IS_SNAPSHOT(ip)) {
2006 				reason = "non-snapshot";
2007 			} else {
2008 				reason = "old format snapshot";
2009 				(void)ffs_truncate(vp, (off_t)0, 0, NOCRED);
2010 				(void)ffs_syncvnode(vp, MNT_WAIT, 0);
2011 			}
2012 			printf("ffs_snapshot_mount: %s inode %d\n",
2013 			    reason, fs->fs_snapinum[snaploc]);
2014 			vput(vp);
2015 			vp = NULL;
2016 			for (loc = snaploc + 1; loc < FSMAXSNAP; loc++) {
2017 				if (fs->fs_snapinum[loc] == 0)
2018 					break;
2019 				fs->fs_snapinum[loc - 1] = fs->fs_snapinum[loc];
2020 			}
2021 			fs->fs_snapinum[loc - 1] = 0;
2022 			snaploc--;
2023 			continue;
2024 		}
2025 		/*
2026 		 * Acquire a lock on the snapdata structure, creating it if
2027 		 * necessary.
2028 		 */
2029 		sn = ffs_snapdata_acquire(devvp);
2030 		/*
2031 		 * Change vnode to use shared snapshot lock instead of the
2032 		 * original private lock.
2033 		 */
2034 		vp->v_vnlock = &sn->sn_lock;
2035 		lockmgr(&vp->v_lock, LK_RELEASE, NULL);
2036 		/*
2037 		 * Link it onto the active snapshot list.
2038 		 */
2039 		VI_LOCK(devvp);
2040 		if (ip->i_nextsnap.tqe_prev != 0)
2041 			panic("ffs_snapshot_mount: %ju already on list",
2042 			    (uintmax_t)ip->i_number);
2043 		else
2044 			TAILQ_INSERT_TAIL(&sn->sn_head, ip, i_nextsnap);
2045 		vp->v_vflag |= VV_SYSTEM;
2046 		VI_UNLOCK(devvp);
2047 		VOP_UNLOCK(vp, 0);
2048 		lastvp = vp;
2049 	}
2050 	vp = lastvp;
2051 	/*
2052 	 * No usable snapshots found.
2053 	 */
2054 	if (sn == NULL || vp == NULL)
2055 		return;
2056 	/*
2057 	 * Allocate the space for the block hints list. We always want to
2058 	 * use the list from the newest snapshot.
2059 	 */
2060 	auio.uio_iov = &aiov;
2061 	auio.uio_iovcnt = 1;
2062 	aiov.iov_base = (void *)&snaplistsize;
2063 	aiov.iov_len = sizeof(snaplistsize);
2064 	auio.uio_resid = aiov.iov_len;
2065 	auio.uio_offset =
2066 	    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag));
2067 	auio.uio_segflg = UIO_SYSSPACE;
2068 	auio.uio_rw = UIO_READ;
2069 	auio.uio_td = td;
2070 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2071 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
2072 		printf("ffs_snapshot_mount: read_1 failed %d\n", error);
2073 		VOP_UNLOCK(vp, 0);
2074 		return;
2075 	}
2076 	snapblklist = malloc(snaplistsize * sizeof(daddr_t),
2077 	    M_UFSMNT, M_WAITOK);
2078 	auio.uio_iovcnt = 1;
2079 	aiov.iov_base = snapblklist;
2080 	aiov.iov_len = snaplistsize * sizeof (daddr_t);
2081 	auio.uio_resid = aiov.iov_len;
2082 	auio.uio_offset -= sizeof(snaplistsize);
2083 	if ((error = VOP_READ(vp, &auio, IO_UNIT, td->td_ucred)) != 0) {
2084 		printf("ffs_snapshot_mount: read_2 failed %d\n", error);
2085 		VOP_UNLOCK(vp, 0);
2086 		free(snapblklist, M_UFSMNT);
2087 		return;
2088 	}
2089 	VOP_UNLOCK(vp, 0);
2090 	VI_LOCK(devvp);
2091 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_mount");
2092 	sn->sn_listsize = snaplistsize;
2093 	sn->sn_blklist = (daddr_t *)snapblklist;
2094 	devvp->v_vflag |= VV_COPYONWRITE;
2095 	VI_UNLOCK(devvp);
2096 }
2097 
2098 /*
2099  * Disassociate snapshot files when unmounting.
2100  */
2101 void
2102 ffs_snapshot_unmount(mp)
2103 	struct mount *mp;
2104 {
2105 	struct vnode *devvp = VFSTOUFS(mp)->um_devvp;
2106 	struct snapdata *sn;
2107 	struct inode *xp;
2108 	struct vnode *vp;
2109 
2110 	VI_LOCK(devvp);
2111 	sn = devvp->v_rdev->si_snapdata;
2112 	while (sn != NULL && (xp = TAILQ_FIRST(&sn->sn_head)) != NULL) {
2113 		vp = ITOV(xp);
2114 		TAILQ_REMOVE(&sn->sn_head, xp, i_nextsnap);
2115 		xp->i_nextsnap.tqe_prev = 0;
2116 		lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE,
2117 		    VI_MTX(devvp));
2118 		lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);
2119 		KASSERT(vp->v_vnlock == &sn->sn_lock,
2120 		("ffs_snapshot_unmount: lost lock mutation"));
2121 		vp->v_vnlock = &vp->v_lock;
2122 		lockmgr(&vp->v_lock, LK_RELEASE, NULL);
2123 		lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
2124 		if (xp->i_effnlink > 0)
2125 			vrele(vp);
2126 		VI_LOCK(devvp);
2127 		sn = devvp->v_rdev->si_snapdata;
2128 	}
2129 	try_free_snapdata(devvp);
2130 	ASSERT_VOP_LOCKED(devvp, "ffs_snapshot_unmount");
2131 }
2132 
2133 /*
2134  * Check the buffer block to be belong to device buffer that shall be
2135  * locked after snaplk. devvp shall be locked on entry, and will be
2136  * leaved locked upon exit.
2137  */
2138 static int
2139 ffs_bp_snapblk(devvp, bp)
2140 	struct vnode *devvp;
2141 	struct buf *bp;
2142 {
2143 	struct snapdata *sn;
2144 	struct fs *fs;
2145 	ufs2_daddr_t lbn, *snapblklist;
2146 	int lower, upper, mid;
2147 
2148 	ASSERT_VI_LOCKED(devvp, "ffs_bp_snapblk");
2149 	KASSERT(devvp->v_type == VCHR, ("Not a device %p", devvp));
2150 	sn = devvp->v_rdev->si_snapdata;
2151 	if (sn == NULL || TAILQ_FIRST(&sn->sn_head) == NULL)
2152 		return (0);
2153 	fs = ITOFS(TAILQ_FIRST(&sn->sn_head));
2154 	lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno));
2155 	snapblklist = sn->sn_blklist;
2156 	upper = sn->sn_listsize - 1;
2157 	lower = 1;
2158 	while (lower <= upper) {
2159 		mid = (lower + upper) / 2;
2160 		if (snapblklist[mid] == lbn)
2161 			break;
2162 		if (snapblklist[mid] < lbn)
2163 			lower = mid + 1;
2164 		else
2165 			upper = mid - 1;
2166 	}
2167 	if (lower <= upper)
2168 		return (1);
2169 	return (0);
2170 }
2171 
2172 void
2173 ffs_bdflush(bo, bp)
2174 	struct bufobj *bo;
2175 	struct buf *bp;
2176 {
2177 	struct thread *td;
2178 	struct vnode *vp, *devvp;
2179 	struct buf *nbp;
2180 	int bp_bdskip;
2181 
2182 	if (bo->bo_dirty.bv_cnt <= dirtybufthresh)
2183 		return;
2184 
2185 	td = curthread;
2186 	vp = bp->b_vp;
2187 	devvp = bo2vnode(bo);
2188 	KASSERT(vp == devvp, ("devvp != vp %p %p", bo, bp));
2189 
2190 	VI_LOCK(devvp);
2191 	bp_bdskip = ffs_bp_snapblk(devvp, bp);
2192 	if (bp_bdskip)
2193 		bdwriteskip++;
2194 	VI_UNLOCK(devvp);
2195 	if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10 && !bp_bdskip) {
2196 		(void) VOP_FSYNC(vp, MNT_NOWAIT, td);
2197 		altbufferflushes++;
2198 	} else {
2199 		BO_LOCK(bo);
2200 		/*
2201 		 * Try to find a buffer to flush.
2202 		 */
2203 		TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) {
2204 			if ((nbp->b_vflags & BV_BKGRDINPROG) ||
2205 			    BUF_LOCK(nbp,
2206 				     LK_EXCLUSIVE | LK_NOWAIT, NULL))
2207 				continue;
2208 			if (bp == nbp)
2209 				panic("bdwrite: found ourselves");
2210 			BO_UNLOCK(bo);
2211 			/*
2212 			 * Don't countdeps with the bo lock
2213 			 * held.
2214 			 */
2215 			if (buf_countdeps(nbp, 0)) {
2216 				BO_LOCK(bo);
2217 				BUF_UNLOCK(nbp);
2218 				continue;
2219 			}
2220 			if (bp_bdskip) {
2221 				VI_LOCK(devvp);
2222 				if (!ffs_bp_snapblk(vp, nbp)) {
2223 					VI_UNLOCK(devvp);
2224 					BO_LOCK(bo);
2225 					BUF_UNLOCK(nbp);
2226 					continue;
2227 				}
2228 				VI_UNLOCK(devvp);
2229 			}
2230 			if (nbp->b_flags & B_CLUSTEROK) {
2231 				vfs_bio_awrite(nbp);
2232 			} else {
2233 				bremfree(nbp);
2234 				bawrite(nbp);
2235 			}
2236 			dirtybufferflushes++;
2237 			break;
2238 		}
2239 		if (nbp == NULL)
2240 			BO_UNLOCK(bo);
2241 	}
2242 }
2243 
2244 /*
2245  * Check for need to copy block that is about to be written,
2246  * copying the block if necessary.
2247  */
2248 int
2249 ffs_copyonwrite(devvp, bp)
2250 	struct vnode *devvp;
2251 	struct buf *bp;
2252 {
2253 	struct snapdata *sn;
2254 	struct buf *ibp, *cbp, *savedcbp = NULL;
2255 	struct thread *td = curthread;
2256 	struct fs *fs;
2257 	struct inode *ip;
2258 	struct vnode *vp = NULL;
2259 	ufs2_daddr_t lbn, blkno, *snapblklist;
2260 	int lower, upper, mid, indiroff, error = 0;
2261 	int launched_async_io, prev_norunningbuf;
2262 	long saved_runningbufspace;
2263 
2264 	if (devvp != bp->b_vp && IS_SNAPSHOT(VTOI(bp->b_vp)))
2265 		return (0);		/* Update on a snapshot file */
2266 	if (td->td_pflags & TDP_COWINPROGRESS)
2267 		panic("ffs_copyonwrite: recursive call");
2268 	/*
2269 	 * First check to see if it is in the preallocated list.
2270 	 * By doing this check we avoid several potential deadlocks.
2271 	 */
2272 	VI_LOCK(devvp);
2273 	sn = devvp->v_rdev->si_snapdata;
2274 	if (sn == NULL ||
2275 	    TAILQ_EMPTY(&sn->sn_head)) {
2276 		VI_UNLOCK(devvp);
2277 		return (0);		/* No snapshot */
2278 	}
2279 	ip = TAILQ_FIRST(&sn->sn_head);
2280 	fs = ITOFS(ip);
2281 	lbn = fragstoblks(fs, dbtofsb(fs, bp->b_blkno));
2282 	snapblklist = sn->sn_blklist;
2283 	upper = sn->sn_listsize - 1;
2284 	lower = 1;
2285 	while (lower <= upper) {
2286 		mid = (lower + upper) / 2;
2287 		if (snapblklist[mid] == lbn)
2288 			break;
2289 		if (snapblklist[mid] < lbn)
2290 			lower = mid + 1;
2291 		else
2292 			upper = mid - 1;
2293 	}
2294 	if (lower <= upper) {
2295 		VI_UNLOCK(devvp);
2296 		return (0);
2297 	}
2298 	launched_async_io = 0;
2299 	prev_norunningbuf = td->td_pflags & TDP_NORUNNINGBUF;
2300 	/*
2301 	 * Since I/O on bp isn't yet in progress and it may be blocked
2302 	 * for a long time waiting on snaplk, back it out of
2303 	 * runningbufspace, possibly waking other threads waiting for space.
2304 	 */
2305 	saved_runningbufspace = bp->b_runningbufspace;
2306 	if (saved_runningbufspace != 0)
2307 		runningbufwakeup(bp);
2308 	/*
2309 	 * Not in the precomputed list, so check the snapshots.
2310 	 */
2311 	while (lockmgr(&sn->sn_lock, LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
2312 	    VI_MTX(devvp)) != 0) {
2313 		VI_LOCK(devvp);
2314 		sn = devvp->v_rdev->si_snapdata;
2315 		if (sn == NULL ||
2316 		    TAILQ_EMPTY(&sn->sn_head)) {
2317 			VI_UNLOCK(devvp);
2318 			if (saved_runningbufspace != 0) {
2319 				bp->b_runningbufspace = saved_runningbufspace;
2320 				atomic_add_long(&runningbufspace,
2321 					       bp->b_runningbufspace);
2322 			}
2323 			return (0);		/* Snapshot gone */
2324 		}
2325 	}
2326 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
2327 		vp = ITOV(ip);
2328 		if (DOINGSOFTDEP(vp))
2329 			softdep_prealloc(vp, MNT_WAIT);
2330 		/*
2331 		 * We ensure that everything of our own that needs to be
2332 		 * copied will be done at the time that ffs_snapshot is
2333 		 * called. Thus we can skip the check here which can
2334 		 * deadlock in doing the lookup in UFS_BALLOC.
2335 		 */
2336 		if (bp->b_vp == vp)
2337 			continue;
2338 		/*
2339 		 * Check to see if block needs to be copied. We do not have
2340 		 * to hold the snapshot lock while doing this lookup as it
2341 		 * will never require any additional allocations for the
2342 		 * snapshot inode.
2343 		 */
2344 		if (lbn < UFS_NDADDR) {
2345 			blkno = DIP(ip, i_db[lbn]);
2346 		} else {
2347 			td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2348 			error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2349 			   fs->fs_bsize, KERNCRED, BA_METAONLY, &ibp);
2350 			td->td_pflags &= ~TDP_COWINPROGRESS;
2351 			if (error)
2352 				break;
2353 			indiroff = (lbn - UFS_NDADDR) % NINDIR(fs);
2354 			if (I_IS_UFS1(ip))
2355 				blkno=((ufs1_daddr_t *)(ibp->b_data))[indiroff];
2356 			else
2357 				blkno=((ufs2_daddr_t *)(ibp->b_data))[indiroff];
2358 			bqrelse(ibp);
2359 		}
2360 #ifdef INVARIANTS
2361 		if (blkno == BLK_SNAP && bp->b_lblkno >= 0)
2362 			panic("ffs_copyonwrite: bad copy block");
2363 #endif
2364 		if (blkno != 0)
2365 			continue;
2366 		/*
2367 		 * Allocate the block into which to do the copy. Since
2368 		 * multiple processes may all try to copy the same block,
2369 		 * we have to recheck our need to do a copy if we sleep
2370 		 * waiting for the lock.
2371 		 *
2372 		 * Because all snapshots on a filesystem share a single
2373 		 * lock, we ensure that we will never be in competition
2374 		 * with another process to allocate a block.
2375 		 */
2376 		td->td_pflags |= TDP_COWINPROGRESS | TDP_NORUNNINGBUF;
2377 		error = UFS_BALLOC(vp, lblktosize(fs, (off_t)lbn),
2378 		    fs->fs_bsize, KERNCRED, 0, &cbp);
2379 		td->td_pflags &= ~TDP_COWINPROGRESS;
2380 		if (error)
2381 			break;
2382 #ifdef DEBUG
2383 		if (snapdebug) {
2384 			printf("Copyonwrite: snapino %ju lbn %jd for ",
2385 			    (uintmax_t)ip->i_number, (intmax_t)lbn);
2386 			if (bp->b_vp == devvp)
2387 				printf("fs metadata");
2388 			else
2389 				printf("inum %ju",
2390 				    (uintmax_t)VTOI(bp->b_vp)->i_number);
2391 			printf(" lblkno %jd to blkno %jd\n",
2392 			    (intmax_t)bp->b_lblkno, (intmax_t)cbp->b_blkno);
2393 		}
2394 #endif
2395 		/*
2396 		 * If we have already read the old block contents, then
2397 		 * simply copy them to the new block. Note that we need
2398 		 * to synchronously write snapshots that have not been
2399 		 * unlinked, and hence will be visible after a crash,
2400 		 * to ensure their integrity. At a minimum we ensure the
2401 		 * integrity of the filesystem metadata, but use the
2402 		 * dopersistence sysctl-setable flag to decide on the
2403 		 * persistence needed for file content data.
2404 		 */
2405 		if (savedcbp != NULL) {
2406 			bcopy(savedcbp->b_data, cbp->b_data, fs->fs_bsize);
2407 			bawrite(cbp);
2408 			if ((devvp == bp->b_vp || bp->b_vp->v_type == VDIR ||
2409 			    dopersistence) && ip->i_effnlink > 0)
2410 				(void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT);
2411 			else
2412 				launched_async_io = 1;
2413 			continue;
2414 		}
2415 		/*
2416 		 * Otherwise, read the old block contents into the buffer.
2417 		 */
2418 		if ((error = readblock(vp, cbp, lbn)) != 0) {
2419 			bzero(cbp->b_data, fs->fs_bsize);
2420 			bawrite(cbp);
2421 			if ((devvp == bp->b_vp || bp->b_vp->v_type == VDIR ||
2422 			    dopersistence) && ip->i_effnlink > 0)
2423 				(void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT);
2424 			else
2425 				launched_async_io = 1;
2426 			break;
2427 		}
2428 		savedcbp = cbp;
2429 	}
2430 	/*
2431 	 * Note that we need to synchronously write snapshots that
2432 	 * have not been unlinked, and hence will be visible after
2433 	 * a crash, to ensure their integrity. At a minimum we
2434 	 * ensure the integrity of the filesystem metadata, but
2435 	 * use the dopersistence sysctl-setable flag to decide on
2436 	 * the persistence needed for file content data.
2437 	 */
2438 	if (savedcbp) {
2439 		vp = savedcbp->b_vp;
2440 		bawrite(savedcbp);
2441 		if ((devvp == bp->b_vp || bp->b_vp->v_type == VDIR ||
2442 		    dopersistence) && VTOI(vp)->i_effnlink > 0)
2443 			(void) ffs_syncvnode(vp, MNT_WAIT, NO_INO_UPDT);
2444 		else
2445 			launched_async_io = 1;
2446 	}
2447 	lockmgr(vp->v_vnlock, LK_RELEASE, NULL);
2448 	td->td_pflags = (td->td_pflags & ~TDP_NORUNNINGBUF) |
2449 		prev_norunningbuf;
2450 	if (launched_async_io && (td->td_pflags & TDP_NORUNNINGBUF) == 0)
2451 		waitrunningbufspace();
2452 	/*
2453 	 * I/O on bp will now be started, so count it in runningbufspace.
2454 	 */
2455 	if (saved_runningbufspace != 0) {
2456 		bp->b_runningbufspace = saved_runningbufspace;
2457 		atomic_add_long(&runningbufspace, bp->b_runningbufspace);
2458 	}
2459 	return (error);
2460 }
2461 
2462 /*
2463  * sync snapshots to force freework records waiting on snapshots to claim
2464  * blocks to free.
2465  */
2466 void
2467 ffs_sync_snap(mp, waitfor)
2468 	struct mount *mp;
2469 	int waitfor;
2470 {
2471 	struct snapdata *sn;
2472 	struct vnode *devvp;
2473 	struct vnode *vp;
2474 	struct inode *ip;
2475 
2476 	devvp = VFSTOUFS(mp)->um_devvp;
2477 	if ((devvp->v_vflag & VV_COPYONWRITE) == 0)
2478 		return;
2479 	for (;;) {
2480 		VI_LOCK(devvp);
2481 		sn = devvp->v_rdev->si_snapdata;
2482 		if (sn == NULL) {
2483 			VI_UNLOCK(devvp);
2484 			return;
2485 		}
2486 		if (lockmgr(&sn->sn_lock,
2487 		    LK_INTERLOCK | LK_EXCLUSIVE | LK_SLEEPFAIL,
2488 		    VI_MTX(devvp)) == 0)
2489 			break;
2490 	}
2491 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
2492 		vp = ITOV(ip);
2493 		ffs_syncvnode(vp, waitfor, NO_INO_UPDT);
2494 	}
2495 	lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
2496 }
2497 
2498 /*
2499  * Read the specified block into the given buffer.
2500  * Much of this boiler-plate comes from bwrite().
2501  */
2502 static int
2503 readblock(vp, bp, lbn)
2504 	struct vnode *vp;
2505 	struct buf *bp;
2506 	ufs2_daddr_t lbn;
2507 {
2508 	struct inode *ip;
2509 	struct bio *bip;
2510 	struct fs *fs;
2511 
2512 	ip = VTOI(vp);
2513 	fs = ITOFS(ip);
2514 
2515 	bip = g_alloc_bio();
2516 	bip->bio_cmd = BIO_READ;
2517 	bip->bio_offset = dbtob(fsbtodb(fs, blkstofrags(fs, lbn)));
2518 	bip->bio_data = bp->b_data;
2519 	bip->bio_length = bp->b_bcount;
2520 	bip->bio_done = NULL;
2521 
2522 	g_io_request(bip, ITODEVVP(ip)->v_bufobj.bo_private);
2523 	bp->b_error = biowait(bip, "snaprdb");
2524 	g_destroy_bio(bip);
2525 	return (bp->b_error);
2526 }
2527 
2528 #endif
2529 
2530 /*
2531  * Process file deletes that were deferred by ufs_inactive() due to
2532  * the file system being suspended. Transfer IN_LAZYACCESS into
2533  * IN_MODIFIED for vnodes that were accessed during suspension.
2534  */
2535 void
2536 process_deferred_inactive(struct mount *mp)
2537 {
2538 	struct vnode *vp, *mvp;
2539 	struct inode *ip;
2540 	struct thread *td;
2541 	int error;
2542 
2543 	td = curthread;
2544 	(void) vn_start_secondary_write(NULL, &mp, V_WAIT);
2545  loop:
2546 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2547 		/*
2548 		 * IN_LAZYACCESS is checked here without holding any
2549 		 * vnode lock, but this flag is set only while holding
2550 		 * vnode interlock.
2551 		 */
2552 		if (vp->v_type == VNON ||
2553 		    ((VTOI(vp)->i_flag & IN_LAZYACCESS) == 0 &&
2554 		    ((vp->v_iflag & VI_OWEINACT) == 0 || vp->v_usecount > 0))) {
2555 			VI_UNLOCK(vp);
2556 			continue;
2557 		}
2558 		vholdl(vp);
2559 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2560 		if (error != 0) {
2561 			vdrop(vp);
2562 			if (error == ENOENT)
2563 				continue;	/* vnode recycled */
2564 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2565 			goto loop;
2566 		}
2567 		ip = VTOI(vp);
2568 		if ((ip->i_flag & IN_LAZYACCESS) != 0) {
2569 			ip->i_flag &= ~IN_LAZYACCESS;
2570 			ip->i_flag |= IN_MODIFIED;
2571 		}
2572 		VI_LOCK(vp);
2573 		if ((vp->v_iflag & VI_OWEINACT) == 0 || vp->v_usecount > 0) {
2574 			VI_UNLOCK(vp);
2575 			VOP_UNLOCK(vp, 0);
2576 			vdrop(vp);
2577 			continue;
2578 		}
2579 		vinactive(vp, td);
2580 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2581 			 ("process_deferred_inactive: got VI_OWEINACT"));
2582 		VI_UNLOCK(vp);
2583 		VOP_UNLOCK(vp, 0);
2584 		vdrop(vp);
2585 	}
2586 	vn_finished_secondary_write(mp);
2587 }
2588 
2589 #ifndef NO_FFS_SNAPSHOT
2590 
2591 static struct snapdata *
2592 ffs_snapdata_alloc(void)
2593 {
2594 	struct snapdata *sn;
2595 
2596 	/*
2597 	 * Fetch a snapdata from the free list if there is one available.
2598 	 */
2599 	mtx_lock(&snapfree_lock);
2600 	sn = LIST_FIRST(&snapfree);
2601 	if (sn != NULL)
2602 		LIST_REMOVE(sn, sn_link);
2603 	mtx_unlock(&snapfree_lock);
2604 	if (sn != NULL)
2605 		return (sn);
2606 	/*
2607  	 * If there were no free snapdatas allocate one.
2608 	 */
2609 	sn = malloc(sizeof *sn, M_UFSMNT, M_WAITOK | M_ZERO);
2610 	TAILQ_INIT(&sn->sn_head);
2611 	lockinit(&sn->sn_lock, PVFS, "snaplk", VLKTIMEOUT,
2612 	    LK_CANRECURSE | LK_NOSHARE);
2613 	return (sn);
2614 }
2615 
2616 /*
2617  * The snapdata is never freed because we can not be certain that
2618  * there are no threads sleeping on the snap lock.  Persisting
2619  * them permanently avoids costly synchronization in ffs_lock().
2620  */
2621 static void
2622 ffs_snapdata_free(struct snapdata *sn)
2623 {
2624 	mtx_lock(&snapfree_lock);
2625 	LIST_INSERT_HEAD(&snapfree, sn, sn_link);
2626 	mtx_unlock(&snapfree_lock);
2627 }
2628 
2629 /* Try to free snapdata associated with devvp */
2630 static void
2631 try_free_snapdata(struct vnode *devvp)
2632 {
2633 	struct snapdata *sn;
2634 	ufs2_daddr_t *snapblklist;
2635 
2636 	ASSERT_VI_LOCKED(devvp, "try_free_snapdata");
2637 	sn = devvp->v_rdev->si_snapdata;
2638 
2639 	if (sn == NULL || TAILQ_FIRST(&sn->sn_head) != NULL ||
2640 	    (devvp->v_vflag & VV_COPYONWRITE) == 0) {
2641 		VI_UNLOCK(devvp);
2642 		return;
2643 	}
2644 
2645 	devvp->v_rdev->si_snapdata = NULL;
2646 	devvp->v_vflag &= ~VV_COPYONWRITE;
2647 	lockmgr(&sn->sn_lock, LK_DRAIN|LK_INTERLOCK, VI_MTX(devvp));
2648 	snapblklist = sn->sn_blklist;
2649 	sn->sn_blklist = NULL;
2650 	sn->sn_listsize = 0;
2651 	lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
2652 	if (snapblklist != NULL)
2653 		free(snapblklist, M_UFSMNT);
2654 	ffs_snapdata_free(sn);
2655 }
2656 
2657 static struct snapdata *
2658 ffs_snapdata_acquire(struct vnode *devvp)
2659 {
2660 	struct snapdata *nsn, *sn;
2661 	int error;
2662 
2663 	/*
2664 	 * Allocate a free snapdata.  This is done before acquiring the
2665 	 * devvp lock to avoid allocation while the devvp interlock is
2666 	 * held.
2667 	 */
2668 	nsn = ffs_snapdata_alloc();
2669 
2670 	for (;;) {
2671 		VI_LOCK(devvp);
2672 		sn = devvp->v_rdev->si_snapdata;
2673 		if (sn == NULL) {
2674 			/*
2675 			 * This is the first snapshot on this
2676 			 * filesystem and we use our pre-allocated
2677 			 * snapdata.  Publish sn with the sn_lock
2678 			 * owned by us, to avoid the race.
2679 			 */
2680 			error = lockmgr(&nsn->sn_lock, LK_EXCLUSIVE |
2681 			    LK_NOWAIT, NULL);
2682 			if (error != 0)
2683 				panic("leaked sn, lockmgr error %d", error);
2684 			sn = devvp->v_rdev->si_snapdata = nsn;
2685 			VI_UNLOCK(devvp);
2686 			nsn = NULL;
2687 			break;
2688 		}
2689 
2690 		/*
2691 		 * There is a snapshots which already exists on this
2692 		 * filesystem, grab a reference to the common lock.
2693 		 */
2694 		error = lockmgr(&sn->sn_lock, LK_INTERLOCK |
2695 		    LK_EXCLUSIVE | LK_SLEEPFAIL, VI_MTX(devvp));
2696 		if (error == 0)
2697 			break;
2698 	}
2699 
2700 	/*
2701 	 * Free any unused snapdata.
2702 	 */
2703 	if (nsn != NULL)
2704 		ffs_snapdata_free(nsn);
2705 
2706 	return (sn);
2707 }
2708 
2709 #endif
2710