Skip to content

HANDOFF — floppy PIO (3027) controller: C and C# defects

For: whoever owns the nd100x C emulator and/or the RetroCore C# emulator. Subject: the OLDER PIO (programmed-I/O) floppy controller "3027" — device octal 1560/1570, ident 021/022, interrupt level 11. This is a different device from the 3112 DMA controller (whose nd100x fixes landed as nd100x commit efed8ae; spec in floppy-3112-register-spec-ND-11.021.md).

Files: - C: the nd100x repository, src/devices/floppy/deviceFloppyPIO.c + .h - C#: the RetroCore repository, Emulated.HW/ND/CPU/NDBUS/NDBusFloppyPIO.cs

Manual: there is no dedicated 3027 / ND-11.015 / ND-11.012 PIO manual in NDInsight (confirmed). The authoritative register spec used here is §B.4 "ND-100 Floppy Disk Programming Specification" in Reference-Manuals/ND-06.015.02 ND-100 Functional Description.md in the NDInsight repository (lines ~9747-9990). Line numbers below are approximate — the implementer should confirm each against current source before editing.

Rule: no guessing. Every defect is grounded in a manual line or a cross-model divergence. Items that cannot be verified (autoload bit 2) are labelled UNVERIFIABLE, not "fixed."

What is already correct (do NOT touch): IOX register map (+0..+7), RSR1/RSR2 core bit layouts, WCWD command bits, WDAD/WSCT decode, sector geometry (fmt0/1=128B/26sec, fmt2=256B/15sec, fmt3=512B/8sec), and the READ_DATA / READ_ID / SEEK / RECALIBRATE paths — all faithful to §B.4 and matching between the two models.

C-side defects (deviceFloppyPIO.c) — real, unique to C

C-PIO-1 [write-broken + wrong-error] Image opened read-only; write-protect guard is dead code

  • Constructor opens the image fopen(floppyName, "r") (≈ :229) — read-only.
  • ExecuteGo forces RSR2.bits.writeProtect = false (≈ :351), so the write-protect guards in WRITE / WRITE_DELETED / FORMAT_TRACK (≈ :390/435/480) can never fire.
  • Result: every write actually fails at the C library level, surfacing as drive-not-ready (inferred), and the manual's real write-protect = RSR2 bit 9 (manual ≈9868-9870) is never reported.
  • RetroCore C# does this correctly (honors real read-only via IsDeviceReadOnly, reports WriteProtect), so this is both a divergence-from-C# and a manual-violation.
  • Fix: open the image read/write ("rb+"), track the real read-only state, and set RSR2.writeProtect (bit 9) when a write is attempted on RO media — instead of forcing it false. Only fall back to drive-not-ready for genuine seek/IO failures.

C-PIO-2 [data-corruption] FORMAT_TRACK start position off-by-one

  • FORMAT_TRACK computes position = (1 * bytes_pr_sector) + track*bytes*sectors (≈ :396-397) — it starts one sector into the track.
  • Effect: it overruns the end of the track and leaves sector 1 unformatted.
  • C# uses per-sector (s-1)*bytes starting at 0 (≈ :825), which is the correct track layout.
  • Fix: start at the track base (position = track*bytes_pr_sector*sectors) and format sectors 1..sectors_pr_track from offset 0.

C#-side defects (NDBusFloppyPIO.cs) — real, unique to C

CS-PIO-1 [breaks command in debug builds] Debug.Assert(false) landmines

  • Debug.Assert(false) sits in FORMAT_TRACK (≈ :820, :845) and CONTROL_RESET (≈ :1030). In any debug build these commands abort — FORMAT_TRACK is effectively unimplemented.
  • C runs both commands (however imperfectly — see C-PIO-2).
  • Fix: either implement FORMAT_TRACK / CONTROL_RESET properly or remove the asserts and let them complete; do not leave an assert that aborts a legal driver command.

CS-PIO cosmetics (low priority, note only)

  • Deleted-sector map DeletedSector[sector, track] is unbounded and index-swapped vs C's bounds-checked [track][sector-1] (≈ :1060). Works in-range; would throw if indices exceeded 99.
  • IDENT clears InterruptEnabled unconditionally (≈ :1105) even when this device did not raise the level. C acts only if it owns the pending interrupt.
  • Reset does not clear selectedDrive / RSR2 (≈ :420-425); C resets selectedDrive=-1 and interruptBits=0.
  • Dead duplicate SetBit911 (≈ :734-756).
  • Drive-select decoded as 2 bits ((value>>8)&0b11, ≈ :647) vs the manual's 3-bit unit field (b8-10); harmless for units 0-2 only.

Shared deviations from §B.4 (both models agree with each other, both wrong vs the spec)

SH-PIO-1 [wrong-status] CONTROL_RESET (control bit 15) leaves RFT=0 and generates busy

  • ExecuteGo clears ReadyForTransfer on entry (C ≈ :346, C# ≈ :770). A lone control-reset word (0x8000) passes the 0xFF00 command mask, so it enters dispatch, transiently sets busy, and CONTROL_RESET then clears only busy — queuing no completion, leaving RFT=0 with no interrupt.
  • Manual ≈9840: bit 15 (control reset) must NOT generate device-busy and gives no interrupt. Fix: route control-reset outside the busy/command path; do not clear RFT for it.

SH-PIO-2 [wrong-status] Device-clear + command in one WCWD forces drive-not-ready

  • WCWD processing order is device-clear(b4) → clear-buffer(b5) → command dispatch (C ≈ :169-203, C# ≈ :561-629). A word combining device-clear and a command deselects the drive (selectedDrive=-1) then runs the command → guaranteed drive-not-ready (C ≈ :184/368, C# ≈ :591/792). Drivers do combine bits in one IOX. Fix: decide ordering deliberately (a real controller latches the command against the drive selected at that write).

SH-PIO-3 [wrong-status] Sector auto-increment steps one past the last sector

  • Both use if (sector <= sectors_pr_track) sector++ (C ≈ :284, C# ≈ :1077). Manual ≈9983: auto-increment is "not valid past the last sector." At the last sector this increments one beyond range. Fix: use < (stop at the last sector).

SH-PIO-4 [UNVERIFIABLE] Autoload on control bit 2

  • Both use control bit 2 = Activate Autoload (C ≈ .h:82, C# ≈ :182), and both implement the 388-byte floppy_boot blob one byte per 16-bit word (byte-per-word matches the FLOPPY-FU-1986F "Expected 000261"=0xB1 evidence).
  • §B.4 line ≈9833 lists bit 2 as "Not used" — but the real autoload spec is in the unavailable ND-11.015 3027 manual. The two models AGREE, so leave as-is; mark as unverifiable until the 3027 manual surfaces. Do not "correct" it to the §B.4 "not used" reading.

SH-PIO-5 [cosmetic] "Magic" RSR1 bits 9-11 and never-asserted diagnostics

  • Both inject RSR1 bits 9/10/11 from bufferPointer (C ≈ :104-110, C# ≈ :461-463) — not in §B.4, both label it a FLOPPY-FU-1986F test hack. Identical in both; leave (removing it breaks that test program).
  • Diagnostic status bits the manual defines but neither model ever raises: timeout (RSR1 b8), CRC/overrun (RSR2), and — for C — write-protect (dead per C-PIO-1). Note only.

Priority summary

Sev Model Item
write-broken C C-PIO-1 read-only image + dead WP guard
data-corruption C C-PIO-2 FORMAT_TRACK off-by-one
breaks-in-debug C# CS-PIO-1 Debug.Assert(false) in FORMAT_TRACK / CONTROL_RESET
wrong-status both SH-PIO-1 control-reset RFT/busy
wrong-status both SH-PIO-2 device-clear+command ordering
wrong-status both SH-PIO-3 auto-increment boundary
unverifiable both SH-PIO-4 autoload bit 2 (needs ND-11.015)
cosmetic both/C# magic bits, deleted-map, IDENT, reset, dead code

How to validate

  • The FLOPPY-FU-1986F floppy function test program is the real acceptance gate (the C model already tunes the magic RSR1 bits to it). Run it against each model before/after.
  • Cross-check register/command behavior against §B.4 in ND-06.015.02 ND-100 Functional Description.md (lines ~9747-9990).
  • If/when the ND-11.015 3027 manual is located, re-verify SH-PIO-4 (autoload bit 2) and the diagnostic status bits against it.