A serious bug in 4.2 bsd restore

[dlw at ucbopal.CC.Berkeley.ARPA]

Index:	/usr/src/etc/restore	4.2	Fix

Description:
	There is an extremely serious bug in the 4.2bsd restore.
	You may NOT be able to restore a multi-level dump, even
	though the tape is readable and everything is there.

	The problem is actually in ``dump''. However, the kludge
	given below will allow (most) existing incremental dumps
	to be read. I will post a fix to dump asap.

	The problem is that the number of inodes in a filesystem
	is not recorded directly in the dump tape. restore infers
	the number from the size of the bit maps at the beginning
	of the tape. But the bit maps are truncated to the smallest
	possible size by dump before they are recorded. Thus, if
	the last inode USED is 3216, the map will be (roughly)
	that size, instead of the ``real'' number in the filesystem.

	Unlike any previous dump/restore, this version saves the
	inode and symbol table between incremental restores.
	If you do a level 0 on a new filesystem, few inodes
	are actually ``used''. Thus the map will be small. Several
	days later you do a level 3 dump. Lots of inodes are used.
	The maps are much bigger. NOW, assume you have a head crash.
	The level 0 restore will have made a small map and symtable
	because of what is on the level 0 tape. When you go to
	put on the level 3, BINGO! restore will complain of "corrupted
	directories", "expected inode M, got N", etc. This is because
	the inodes it finds on the tape are outside the level 0 map, etc.

	The ``real'' fix is to have dump always record the full map.
	(It isn't THAT big.) Then restore would ``work''. However,
	we (and you) have racks of dump tapes already. The fix
	below causes restore to always use a large map. This
	``works'' as long as MAXINO is larger than required for your
	biggest filesystem. See the output of newfs for the number
	you require (which is ipg * ncg).

Repeat-By:
	See above.

Fix:
	Below is a diff of tape.c. The line numbers are slightly
	off because we have a few other changes in there. While
	I was debugging this, I was frustrated by the fact that verbose
	& debug output comes out randomly out of sync on stdout and stderr
	due to buffering. I modified restore.h to cause everything to
	come out on stdout. It was simpler that way.
-------
diff restore.h.old restore.h
-------
6a7,9
> 
> #undef	stderr
> #define	stderr	stdout
-------
diff tape.c.old tape.c
-------
14a15,16
> #define	MAXINO	65535	/* KLUDGE! */
> 
148,150c150,152
< 	maxino = (spcl.c_count * TP_BSIZE * NBBY) + 1;
< 	dprintf(stdout, "maxino = %d\n", maxino);
< 	map = calloc((unsigned)1, (unsigned)howmany(maxino, NBBY));
---
> 	maxino = (spcl.c_count * TP_BSIZE * NBBY) + 1;	/* WRONG! */
> 	dprintf(stdout, "ino map size = %d\n", maxino);
> 	map = calloc((unsigned)1, (unsigned)howmany(maxino>MAXINO?maxino:MAXINO, NBBY));
160c162
< 	map = calloc((unsigned)1, (unsigned)howmany(maxino, NBBY));
---
> 	map = calloc((unsigned)1, (unsigned)howmany(maxino>MAXINO?maxino:MAXINO, NBBY));
165a168,169
> 	if (MAXINO > maxino)
> 		maxino = MAXINO;	/* Kludge, since we don't really know */

****** context of the 2 lines above ******
	dumpmap = map;
	curfile.action = USING;
	getfile(xtrmap, xtrmapskip);
	if (MAXINO > maxino)
		maxino = MAXINO;	/* Kludge, since we don't really know */
}
******************************************