[Unix-jun72] I recovered 100% of the s1 src code fragments
doug at remarque.org
Mon May 12 14:05:01 AEST 2008
Brad Parker ...brad at heeltoe.com... wrote:
> I'm curious, how did you do it? I took a look at it and gave up after
> about 15 minutes. It looked like a jigsaw puzzle where all the pieces
> were the same shape... :-)
Well, first I stared at it all and wished for some nice powerful
tools that would do N-way approximate diffs with special AI features,
then I started thinking about what sorts of such tools I could
actually create and how long that would take and what sorts of
limitations they'd have, and then I gave up on nice solutions and just
used brute force elbow grease for quite an extended period of time. :-)
Seriously, mostly it was just a matter of being very determined
and wading into it; I consider this to be a pretty important
historical matter. The rest of it is just details:
Warren's frags were consecutive 512 byte blocks containing
string(1) strings or something similar. The underlying file system
had each file represented as a sequence of full 512 byte blocks
plus a trailing partially-filled 512 byte block at the end,
possibly scattered all over the disk, but (think reasons for
defragmentation) only sometimes, and with the physically adjacent
sequences of blocks interleaved with blocks on the free list.
Naturally the trailing portion of the final block of a file,
and all of the free blocks on the disk, were *not* zeroed out
back then. Doing so was a moderately expensive operation introduced
very, very late for most non-DOD-contract operating systems
reluctantly when it was proven that extreme security issues
resulted otherwise (although this was widely known in some
quarters quite early, the cost/benefit analysis was very
different in early days than it usually is today).
Therefore, any sequence of string(1)-type blocks was reasonably
likely to indeed include multiple blocks of a single file, in
original order (with luck), followed possibly by one or more free list
blocks that happened to have string(1)-contents, and/or followed
by a final partial block beginning with a string(1)-sequence,
but with those trailing contents being accidental left-overs
from previous versions of various source files, which upon
editing, left scattered free blocks with string(1)-type contents.
BTW the largest appears to have been frag32, constituted of 45
sequential 512 byte blocks of ascii chars, which turned out to
be "ar.s" with some or all of "dc1.s" appended.
(I probably should have used as starting point all of the 512-byte
blocks of the s1 disk image, rather than the frags Warren found,
but that didn't occur to me until I was largely done. I did generate
a dir full of files containing those 512 byte blocks, with which I've
done some prodding and poking, but haven't done much yet.)
So -- not entirely trusting comments at starts of files, I searched for
matches between data/code labels between the frags and the v5 sources;
that initially helped verify when and if starting comments about
filenames were true, and later helped figure out which files
middle-of-file-frags came from.
Diff was quite a mixed blessing, as usual for this kind of thing,
because it gave me, in effect, lots of false positives/negatives
due to its longest-sequence goals versus my goal of longest-sequence
Part of the confusion is that many of the disk blocks had an
invisible termination mark (lost with the missing inode info),
and the rest of the 512 bytes of the block were filled with,
confusingly enough, more assembly code from one source or another.
Sometimes the garbage at the end was a dupe of the same asm at
the start of the block, sometimes it came from a different .s
I tried to carefully check to ensure that all of the trailing garbage
was in fact a dupe/garbage, and that none such constituted new frags
with irreplaceable information that should be incorporated into some file.
I was in general very concerned about not wasting any possibly useful
bits, since there's a shortage of them.
So there was just a huge amount of eyeballing and searching and
cross referencing, and good old vi work for cutting off garbage
and piecing together related frags, with repeated comparisons
to v5 sources for sanity checking etc.
I kept (imperfect) notes about my conclusions per frag; that definitely
Elbow grease + obsessive compulsive/A.D.D. hyperfocus for
many hours over many days. :-)
Based on the experience, I could probably write a diff-variant
tuned to support this kind of work, but it would still simply
support a human expert, not replace them.
Thanks for asking.
Professional Wild-eyed Visionary Member, Crusaders for a Better Tomorrow
More information about the TUHS