> On Sep 24, 2020, at 6:21 PM, Camiel Vanderhoeven via cctalk > <cctalk@classiccmp.org> wrote: > > > > > >> On Sep 24, 2020, at 5:23 PM, Grant Taylor via cctalk <cctalk@classiccmp.org> >> wrote: >> >> On 9/24/20 1:12 AM, Camiel Vanderhoeven via cctalk wrote: >>> I used an Ancot Ultra2080/Lite SCSI-bus Analyzer. This is a device that >>> connects to a SCSI bus and has a serial port. Over the serial port, you can >>> monitor the signals on the SCSI bus, and use it as a SCSI protocol >>> analyzer. There’s also the possibility to construct and send a SCSI >>> command. Rather than connect a serial terminal to the serial port, I >>> connected a PC, then wrote a C program to control the Ancot. I had the >>> Ancot send commands to the disk to read a sector at a time, and recorded >>> the data sent in response to a file to create a disk image. Slow as hell >>> (each byte on the disk requires sending two hex characters and a space over >>> a slow serial line), but it works. I had to make several passes over the >>> disk, because occasionally the data received from the disk turned out to be >>> data from a different sector than the one I was trying to read. By reading >>> the disk multiple times, I could get rid of these mis-read sectors. >> >> Very NICE hack Camiel. I like it! >> >> I am a little surprised by getting different data for the same sectors. I >> find that mildly concerning. Did you do something like read each byte >> multiple times to find a majority sample? 2/3, 3/5, 4/6, etc? Do you think >> the different data was an artifact of the old drive? Or was it a tickle of >> a bug elsewhere in the chain? > > > The drive was what seems to be a pre-production Fujitsu model (with a serial > number of 37), 17MB in size; I’m pretty sure the disk itself had issues. On > the first pass of reading the disk, about 8% of all sectors wouldn’t read, > returning an “Unrecoverable Read Error”; when I did a second pass reading > just these failed sectors, I found that I could read about 2/3rds of them. I > repeated this 20 times or so, and then I had almost everything. There were 43 > sectors near the end of the disk that I never managed to read, but these were > beyond the filesystems on the disk. That gave me a disk that would boot, but > I has some corrupt files and a corrupt directory. Looking at these, I found > that the corrupt bits were sector-sized, and were identical copies of sectors > elsewhere on the disk. > > In the end, what I ended up doing was read the entire disk 2 more times (so > two more times the 20 iterations of reading the missing sectors until I had a > complete image), and compared the resulting three disk images > sector-by-sector. For some 99% of the sectors, all three copies were in > agreement; where they were not, there were always two images in agreement, so > I knew which one to pick. > > I’ve documented some of this here: > > https://www.vaxbarn.com/index.php/42-repair/577-convex-c1-xp-power-on > <https://www.vaxbarn.com/index.php/42-repair/577-convex-c1-xp-power-on> > <https://www.vaxbarn.com/index.php/42-repair/577-convex-c1-xp-power-on > <https://www.vaxbarn.com/index.php/42-repair/577-convex-c1-xp-power-on>> > > https://www.vaxbarn.com/index.php/42-repair/579-getting-a-convex-c1-to-pass-all-diagnostics > > <https://www.vaxbarn.com/index.php/42-repair/579-getting-a-convex-c1-to-pass-all-diagnostics> > > <https://www.vaxbarn.com/index.php/42-repair/579-getting-a-convex-c1-to-pass-all-diagnostics%5C > > <https://www.vaxbarn.com/index.php/42-repair/579-getting-a-convex-c1-to-pass-all-diagnostics%5C>>
Oh, I was lucky when I spotted it; the first corrupt file I encountered was part of the microcode to be loaded onto the vector processor. Those files have checksums, and the microcode loader complained. I wrote a small utility to extract the files from the disk image (it uses a variant of the UFS file system), and had a look at the corrupt file; one of the blocks in the file looked like part of a directory, which was easy to spot, and that led me to the conclusion that it was valid data, but read from the wrong place. Camiel
