[dm-devel] Re: Help tracking down problem --- endless loop in __find_get_block_slow

[Thomas S. Iversen]

> OK, so we're looking for the buffer_head for block 101 and the first
> buffer_head which is attached to the page represents block 100.  So the
> next buffer_head _should_ represent block 101.  Please print it out:

Not quite the same, but simelar:

Feb 23 14:50:24 localhost kernel: __find_get_block_slow() failed. block=102,
b_blocknr=128, next=129
Feb 23 14:50:24 localhost kernel: b_state=0x00000013, b_size=2048
Feb 23 14:50:24 localhost kernel: device blocksize: 2048
Feb 23 14:50:24 localhost kernel: ------------[ cut here ]------------

> Could be UFS.  But what does "transparent block encryption and sector
> shuffling" mean?  How is the sector shuffling implemented?

GDBE is a block level encrypter. It encrypts the actual sectors
transparently via the GEOM API (corresponds to the devicemapper api in linux).

GBDE assigns a key for each block, thereby introducing keysectors.
Furthermore the sectors are remapped so that one can not guess where e.g.
metadata is located on the physical disk. It is a rather simple remap:

Taken from GDBE:

/*
 * Convert an unencrypted side offset to offsets on the encrypted side.
 *
 * Security objective:  Make it harder to identify what sectors contain what
 * on a "cold" disk image.
 *
 * We do this by adding the "keyoffset" from the lock to the physical sector
 * number modulus the available number of sectors.  Since all physical sectors
 * presumably look the same cold, this will do.
 *
 * As part of the mapping we have to skip the lock sectors which we know
 * the physical address off.  We also truncate the work packet, respecting
 * zone boundaries and lock sectors, so that we end up with a sequence of
 * sectors which are physically contiguous.
 *
 * Shuffling things further is an option, but the incremental frustration is
 * not currently deemed worth the run-time performance hit resulting from the
 * increased number of disk arm movements it would incur.
 *
 * This function offers nothing but a trivial diversion for an attacker able
 * to do "the cleaning lady attack" in its current static mapping form.
 */"
  
Further reading:

	http://phk.freebsd.dk/pubs/bsdcon-03.gbde.paper.pdf

Regards Thomas