Skip to content

Floppy stack review (12-JUL-2026) - open items only

Status (13-JUL-2026, re-checked against the code 28-SEP-2026)

Settled against the ND-11.021.01 manual (spec: floppy-3112-register-spec-ND-11.021.md) and FIXED in ND_FLOPPY_DMA.v, testbench-verified; the detail is in git history:

  • C1 sector-count mode (command-block word 4), C3 stale tail on a partial write, M1 real octal error codes, M2 error code in bits 9-14, M3 IOX +4 = hardware status word (the format word is Status Word 2 at CB+7). Root cause of M2/M3: two distinct status words (hardware section 3.7 vs memory section 3.4) had been merged into one.
  • M11 (+0 read constant): no manual basis for 1 or 0x0F; left as 1 on purpose (do not invent).
  • The == 1 BINT harness bug quoted in the NEVER-READY section below is fixed (NDBus.cpp, see BUG-tape400-sd-level12-storm.md).

Still open:

  • C2 (below): a watchdog exists (DISK_TIMEOUT) but defaults to 0 = off.
  • M4 and M5 (below): checked in code 28-SEP-2026 - s_test_mode is latched and never used, and disk_media_fmt is one input for all drives.
  • M6-M10, the MINOR list, the stub table and the TESTGAP list were not re-checked item by item after the fixes above; treat them as UNVERIFIED.
  • Need a backend-input contract, not bugs: Status Word 2 extra bits (5.25"/96tpi/sector-track, section 3.5.2.2), a write-protect input, the streamer port.
  • Emulator-side fixes for the other repositories: HANDOFF-floppy-pio-c-and-csharp-fixes.md (the nd100x DMA fixes landed as nd100x commit efed8ae).

The review (12-JUL-2026; line numbers are from that date)

Reviewed RTL: ND-BUS-DEVICES/FLOPPY-DMA/circuit/ND_FLOPPY_DMA.v (octal 1560), ND-BUS-DEVICES/FLOPPY/circuit/ND_FLOPPY_PIO.v, SD-FAT/circuit/nd_storage_floppy_adapter.v. References: the nd100x repository, src/devices/floppy/ deviceFloppyDMA.{c,h} + deviceFloppyPIO.{c,h} (the stated port source); the RetroCore repository, Emulated.HW/ND/CPU/NDBUS/ NDBusFloppyDMA.cs + NDBusFloppyPIO.cs (richer register-level reference); Verilog/simDevices/NDBus.cpp + NDDevices.cpp; docs/nd100x-device-semantics.md; FLOPPY-DMA/NEVER-READY-ANALYSIS.md. Every finding verified against both sides.

CRITICAL (would break SINTRAN/boot or corrupt data)

C1. FIXED (sector-count mode) - see the status above.

C2. Selecting a drive with no adapter instance HANGS the controller forever - no error, no timeout, only device clear recovers. Adapter nd_storage_floppy_adapter.v:209 answers only its own DRIVE (silence otherwise, by design); controller E_DISK_RD/:420 and E_DISK_WR/:499 wait on disk_done with no timeout. Drive 2/3 = one bad command-word bit = hang. Reference: DRIVE_NOT_READY + status writeback + RFT + interrupt for unattached units (NDBusFloppyDMA.cs ~720; deviceFloppyDMA.c:411-418). The sim harness MASKS this: NDBus.cpp process_verilog_floppy() ignores FDISK_DRIVE and serves every unit from FLOPPY.IMG - passes runSim, hangs silicon.

C3. FIXED (partial-write stale tail) - see the status above.

MAJOR

M1-M3. FIXED (error codes, their bit position, IOX +4) - see the status above. M4. execute+testMode / execute+streamer run a REAL command (only iox_wdata[8] checked, ND_FLOPPY_DMA.v:349-358); reference routes to ExecuteTest / streamer (no CB fetch). FLOPPY-STREAM/TPE test programs trigger stray DMA command fetches from stale pointers. s_test_mode is latched (:302) and never read - dead register. M5. READ FORMAT succeeds on an unmounted drive (no not-ready path; disk_media_fmt defaulted 0xF = healthy 1.2MB descriptor, NDBus.cpp:346) and media format is ONE GLOBAL input, not per-drive (ND_FLOPPY_DMA.v:124) - drive 1 reports drive 0's geometry. M6. Failed-read memory side effects match NEITHER reference: C zero-fills the full count and advances to the end; C# aborts before data DMA; Verilog aborts mid-transfer with partial last-address in w8/w9. Retry-from-checkpoint drivers see three different worlds. M7. w10/w11 writeback: C/C# write wordsTransfered (=N on success); Verilog writes 0/s_words_left (=0). Verilog matches the FIELD NAME (remaining), C matches the emulator that actually boots SINTRAN. Deliberate decision needed, currently a silent choice. M8. PIO: sector 0 not clamped to 1 (deviceFloppyPIO.c:354-355 vs ND_FLOPPY_PIO.v:309) - 1-based backend math underflows to the previous track's last sector, clean status. M9. PIO autoload (b2) entirely ignored - 1560& on the PIO flavor does nothing (stale buffer, ready status). Documented as not ported, but it is a functional hole; should at least error visibly. M10. PIO same-write control-bit ordering bugs: b5 clear-buffer + command latches the STALE pre-clear pointer (ND_FLOPPY_PIO.v:262-265 vs :280); b4 device-clear + command uses the OLD drive select (C deselects first -> not-ready). Drivers combine WCWD bits in one IOX routinely. M11. +0 read constant: left as 1 on purpose - see the status above.

MINOR

m1. Reset RFT: Verilog 1, C reference 0 (assignment commented out). m2. Device clear clears the error code; C preserves it for post-mortem. m3. Format 2 sector size 128 in Verilog vs 123 in both references (their OCR typo, almost certainly) - fmt-2 images byte-incompatible. m4. First CB+6 writeback carries BUSY=1 (C: busy never set). m5. CB fetch 6 words vs C's 7 (no effect). m6. DELAY_TICKS: PIO 3000 vs DMA 300, both claiming C's IODELAY_FLOPPY (=300); units differ from C's IO-delay queue anyway. m7. PIO status bits that can never assert: timeout, deleted, crc_err, overrun, write_prot (no WP input anywhere; NDBus.cpp read-only fopen fallback silently drops writes = success reported). m8. PIO pre-drive-select error class differs (driveNotReady vs C's sectorMissing). m9. "BRAM" buffers are 2-3-port ASYNC-read arrays in both controllers - cannot map to block RAM; ~16 Kbit LUTRAM each on the Tang. The adapter's s_blkbuf does it right (registered read). m10. Boot-mode +0 read not gated by RFT: double read skips a boot word; read at bootptr==512 serves stale content. No reference exists (both emulators stub autoload) - empirically validated only. m11. Adapter tail-block rule makes the last partial block of an odd-sized image silently unwritable (fine for real floppy sizes, trap for hand-truncated images). m12. Harness media-format default: sizes matching neither floppy size report the 1.2MB descriptor; C reports 0. m13. DMA tb line 486 self-sabotage: control word octal 420 = clear + EXECUTE (stray command from stale CB while tb rewrites memory). Works by timing luck.

STUB INVENTORY

# Where Reference behavior Verilog behavior Consequence
S1 DMA :153-159 w4 WC/SC + count high byte fetched, never decoded C1
S2 DMA :404-414 fns 0x02 FIND_EOF, 0x05 WRITE_EOF, 0x21 FORMAT, 0x23/24 R/W DELETED, 0x2C COPY, 0x2D FORMAT TRACK, 0x2E CHECK, 0x38 IDENTIFY no-op, clean status (C same) FORMAT/CHECK fake success
S3 DMA :302 testMode -> ExecuteTest latched, never read M4
S4 DMA ctl b5 streamer path ignored, floppy go runs M4
S5 DMA :141-150 deletedRecord/retry status bits hardwired 0 FIND_EOF impossible
S6 DMA :377-423 error codes oct 20/41/42/43... invented 1/2 M1
S7 DMA :200 +0: C TODO=1, C# hints 0x0F constant 1 M11
S8 DMA :47-347 autoload = TODO stub in BOTH refs fully invented boot server unverified-by-reference spec; only empirical (runSim TPE)
S9 DMA :124 per-unit media info single global input M5
S10 PIO ctl b2 autoload ignored M9
S11 PIO b6/RSR1 b8 timeout enable/status never set (C same) dead diagnostics
S12 PIO :104-114 writeProtect/crcError/overrun/deleted only cleared, no WP input RO media "writes" succeed
S13 PIO absent C deleted-sector shadow (WRITEDEL/READID) not ported (documented) diagnostics fail
S14 PIO absent RSR1 "magic" bits 9-11 not ported (documented) FLOPPY-FU-1986F fails
S15 PIO backend - no storage adapter exists (DMA-only, owner decision) PIO tb-only, unusable on hw
S16 adapter :139 - s_direct unreachable (wordcount <= 512) dead code
S17 NDBus.cpp:369-447 drive passed to callbacks FDISK_DRIVE ignored C2/M5 masked in sim

TESTGAP (feeds the comprehensive-testbench task)

nd_floppy_dma_tb.v never exercises: sector-count mode / w4 at all; partial-sector WRITE; zero wordcount; disk_err_in (read/write/boot); dma_err (fetch/data/writeback); any drive != 0; data ops in fmt 1/2; READ FORMAT 8-inch / no-media; testMode/streamer; clear-during-transfer; boot +0 double-read / mid-chunk error / boot-after-command; 24-bit DMA addresses; CB+8 value; error codes in CB+6. Backend answers all drives, never errors. Line 486 stray-execute race to fix.

nd_floppy_pio_tb.v: FORMAT TRACK / WRITE DELETED / READ ID / CONTROL RESET untested (4 of 8 commands); test regs +6/+7; b4/b5/b2 control bits; sector 0; end-of-track auto-increment stop; pointer wrap; formats 0/1/2; track clamps; deselect b11; same-write bit combos; backend error pulses.

nd_storage_floppy_adapter_tb.v (strongest): missing wordcount=0; s_span>1024 guard; TWO instances OR-wired with alternating drives; composite request for a drive with NO instance (the C2 hang has no test and no owner); open_start racing an op; woff+wc==1024 boundary; fmt 1/2 sector sizes.

System: runSim golden backend must become drive-aware or C2/M5 class stays invisible; no SINTRAN @ENTER-DIRECTORY check for the +0 constant; no BFDIS driver-sequence replay (SINTRAN J p.411); no error-code decode check against the oct-20/41 table.

NEVER-READY-ANALYSIS.md verification (parallel workstream)

Their F1/F2/F3 fixes are implemented and tb-covered (E_FINAL CB+6 rewrite, real s_status2, +0=1, autoload priority). Open ends: (a) control-latch-while-busy only partial - autoload/execute during a transfer still silently dropped (E_IDLE gates :315,:349) where C always acts; (b) the inert harness interrupt check NDBus.cpp:238-241 ((bits & 1<<11) == 1, never true) is still in the tree. Pushback: the doc treats nd100x C as known-good, but C# contradicts it on +4 (M3), error bit position (M2) and the +0 constant (M11) - "matches the C model" is not "matches the 3112"; resolve against ND-11.021/ND-11.015.

Judgement

Happy path (whole-sector word-count transfers on drive 0, 1560& boot, READ FORMAT on a mounted 1.2MB image, completion interrupt, IDENT 021) is faithful and well-tested. Off it: C1/C2 each break SINTRAN outright, C3 corrupts data, error reporting (M1-M3) is unvalidated invention, PIO has ordering bugs (M10) and no boot story (M9). The nd_storage adapter is the soundest component; its hole is architectural (silent non-answer for unserved drives), not internal.