1
0
mirror of https://github.com/hsoft/collapseos.git synced 2024-11-05 23:40:56 +11:00
collapseos/recipes/rc2014/eeprom/README.md
Virgil Dupras 0044011f06 Move AT28 and TRS80 drivers to higher blocknos
I'm reserving the 100-200 range for non-bootstrapping needs.
2020-04-29 21:04:10 -04:00

3.9 KiB

Writing to a AT28 from Collapse OS

Goal

Write in an AT28 EEPROM from within Collapse OS so that you can have it update itself.

Gathering parts

  • A RC2014 Classic
  • stage2.bin from the base recipe
  • An extra AT28C64B
  • 1x 40106 inverter gates
  • Proto board, RC2014 header pins, wires, IC sockets, etc.

Building the EEPROM holder

The AT28 is SRAM compatible so you could use a RAM module for it. However, there is only one RAM module with the Classic version of the RC2014 and we need it to run Collapse OS.

You could probably use the 64K RAM module for this purpose, but I don't have one and I haven't tried it. For this recipe, I built my own module which is the same as the regular ROM module but with WR wired and geared for address range 0x2000-0x3fff.

If you're tempted by the idea of hacking your existing RC2014 ROM module by wiring WR and write directly to the range 0x0000-0x1fff while running it, be aware that it's not that easy. I was also tempted by this idea, tried it, but on bootup, it seems that some random WR triggers happen and it corrupts the EEPROM contents. Theoretically, we could go around that by putting the AT28 in write protection mode, but I preferred building my own module.

I don't think you need a schematic. It's really simple.

Assembling stage 3

Stage 2 gives you a full interpreter, but it's missing the "Addressed devices" module and the AT28 driver. We'll need to assemble a stage 3.

When you'll have a system with function disk block system, you'll be able to directly LOAD them, but for this recipe, we can't assume you have, so what you'll have to do is to manually paste the code from the appropriate blocks.

Addressed devices are at B140. To know what you have to paste, open the loader block (B142) and see what blocks it loads. For each of the blocks, copy/paste the code in your interpreter.

Do the same thing with the AT28 driver (B480)

If you're doing the real thing and not using the emulator, pasting so much code at once might freeze up the RC2014, so it is recommended that you use /tools/exec that let the other side enough time to breathe.

After your pasting, you'll have a compiled dict of that code in memory. You'll need to relocate it in the same way you did for stage 2, but instead of using RLCORE, which is a convenience word hardcoded for stage 1, we'll parametrize RLDICT, the word doing the real work.

RLDICT takes 2 arguments, target and offset. target is the first word of your relocated dict. In our case, it's going to be ' ADEVMEM+. offset is the offset we'll apply to every eligible word references in our dict. In our case, that offset is the offset of the beginning of the ADEVMEM+ entry (that is, ' ADEVMEM+ WORD( minus the offset of the last word (which should be a hook word) in the ROM binary.

That offset can be conveniently fetched from code because it is the value of the LATEST constant in stable ABI, which is at offset 0x08. Therefore, our offset value is:

' ADEVMEM+ WORD( 0x08 @ -

You can now run RLDICT and proceed with concatenation (and manual adjustments of course) as you did with stage 2. Don't forget to adjust run.fs so that it runs ADEV$.

Writing contents to the AT28

The driver provides AT28! which can be plugged in adev's A!*.

First, upload your binary to some place in memory, for example a000. To do so, run this from your modern computer:

./upload <tty device> a000 <filename>

Then, activate AT28! with ' AT28! A!* ! and then run 0xa000 0x2000 <size-of-bin> AMOVE. AT28! checks every myte for integrity, so it there's no error, you should be fine. Your content is now on the EEPROM!

Why not upload content directly to 0x2000 after having activated AT28!? Technically, you could. It was my first idea too. However, at the time of this writing, I always get weird mismatch errors about halfway through. Maybe that the ACIA interrupt does something wrong...