Carver prompt: carve RCBLO and close the ENTER-DIRECTORY -> SCSI gap¶

Full path: SINTRAN/Devices/SCSI/CARVER-PROMPT-RCBLO.md

Hand the text below (everything from "MISSION" down) to the carving session. It is self-contained.

MISSION (one line)¶

Carve and analyse RCBLO and the resident disk-driver overlay that turns a filesystem "read page 0" request into a call to the SCSI driver entry SCLLD - and find why, during @ENTER-DIRECTORY,,DISC-SCSI-1,0, that call is never made.

Use the sintran-carving skill. Running system = L-VSX-500, symbol set = L07.

1. GROUND TRUTH - VERIFIED THIS SESSION. DO NOT RE-LITIGATE ANY OF IT.¶

Every layer below the gap has been checked and is CORRECT. The bug is NOT in any of them.

1.1 What happens on the wire (from the device trace)¶

@ENTER-DIRECTORY,,DISC-SCSI-1,0 produces exactly three CDBs, then silence: 1. INQUIRY (0x12) - completes. 2. READ CAPACITY (0x25) - completes, returns blockSize=1024, lastLBA=129311. 3. READ_6 (0x08) of LBA 129311 - completes, 1024 bytes DMA'd to ND-100 memory. 4. Target disconnects. Driver idles correctly. 5. NOTHING FURTHER. Page 0 / block 0 is NEVER read. Console prints APPROACHING END OF ACCOUNTING FILE (a known non-fatal, coincidental warning) and returns to @.

Critical nuance: this is not a wrong-LBA bug. No read is issued AT ALL. Whatever aborts, aborts before any further transfer is requested.

1.2 The SCSI controller emulation is CORRECT (verified)¶

The final bare WCONT=0x0005 is SCINT's unconditional exit re-arm (IP-P2-SCSI-DRIV.NPL line 187: 5\/SCCCW; T:=HDEV+WCONT; *IOXT then GO SCWTI, with SCCCW=0). It is not a request for work. Silence after it is correct behaviour.
Do NOT "fix" ExecuteGo() to raise an interrupt. That would be a bug.

1.3 The SCSI driver (`IP-P2-SCSI-DRIV.NPL`) is EXONERATED (verified)¶

On the disconnect it takes A=4 -> DCTHR -> TEROP (sets T=0 = success) -> GO FAR BUSFP -> SELEC. SELEC finds the arbitration queue SCWAQ empty, writes BUSFL=0 and a single WCONT=5, returns. Control falls into EXDRI -> A+2=:P = the normal/success return.
Trace signatures confirming this: exactly ONE WCONT=5 (the GO FAR skips line 187), and NO WDESI/WNCOM/transfer-counter writes (a non-empty queue would have emitted them).
The driver never initiates work. Work enters ONLY via SCLLD -> INITO -> SCWAQ. All SCLLD callers are in IP-P2-SCSI-DISK.NPL lines 116, 445, 1140, 1191, 1364, 1383, 1412, 1413.
The driver never tests BSY/REQ/ACK (SCREQ is defined at line 32 and used ZERO times), never reads RSTAU bits 7/15 (transceiver type), never tests bit 6. Observed RSTAU values (0x0208/0x3208/0x5208) all have bit 4 (Error) and bit 5 (Bus reset) CLEAR, so the error pre-filter at lines 124-131 was skipped entirely. "The bus was in a bad state" is RULED OUT.

1.4 The DMA is byte-exact (verified from the trace's own logging)¶

totalWritten=1  MAR=0x04C600  first16=08     <- byte 0 -> HIGH byte, no MAR increment
totalWritten=2  MAR=0x04C601  first16=00     <- byte 1 -> LOW byte, MAR increments
totalWritten=3  MAR=0x04C601  first16=54
totalWritten=4  MAR=0x04C602  first16=D9

ND-100 memory at 0x04C600 = 0x0800, 0x04C601 = 0x54D9, 0x04C602 = 0x8000 - identical to the disk image. Big-endian, correct. totalWritten reaching exactly 1024 proves the counter was reset before the data phase. Byte-order bugs are RULED OUT.

1.5 The disk is GOOD (verified by reading the image directly)¶

Image: E:\Dev\Ronny\RetroFS\demo\test-images\ndfs\tor-disk.img (132,415,488 bytes = 129312 x 1024). LBA 129311 first 32 bytes: 08 00 54 D9 80 00 00 00 00 00 00 00 00 01 DC D8 C0 00 00 00 00 01 F9 1F ... - word0 = 0x0800 -> NPART = 8 - XOR of all 512 words = 0 - w10:w11 = 0x0001F91F = 129311 (the record states the disk's own last block - self-consistent, and proves the image stores words BIG-ENDIAN)

1.6 The disk layer's function 42 SUCCEEDS (verified by walking the NPL with real values)¶

Source: SINTRAN/NPL-SOURCE/NPL/IP-P2-SCSI-DISK.NPL - The last-block read is NOT a geometry probe. Line 364 says % ADDRESS OF CONTROLL RECORD. SINTRAN reads the LAST 1024-byte page because that is where the control record lives. - Lines 366-367 (X.MEMAD; T:=1000; D+T; A:=A+C / AD=:X.MEMAD) advance MEMA1:MEMA2 by 0o1000 (512) words, so the validation at line 391 reads exactly where the DMA landed (0x04C600). Correct. - Line 397 IF D><0 OR L<=2 OR L>"NCOPA" THEN T:=NOCRC; GO FAR RETEX - with D=0, L=8, NCOPA=0o12=10: 0><0 false, 8<=2 false, 8>10 false. PASSES. - Lines 400-405 then copy NPART*6 = 48 words of partition table to CMADR and set UHLIM = 0x0001DCD8 = 121,560 (data-area size, from word 4 of the partition table). Line 408: T:=0 = success. - The INQUI arithmetic (lines 1222-1290) is clean for blockSize=1024: line 1253 IF A><0 GO FAR RSZER % TO BIG does not fire (A=0, only fires if blockSize >= 65536); the power-of-2 loop exits with A=1 so line 1280 does not fire. All three RSZER exits are passed.

1.7 MYTHS - do not repeat these, they are FALSE¶

ECAPD does not exist in any SCSI source (zero occurrences). It is a TERMINAL driver routine (MP-P2-TERM-DRIV.NPL:144: ECAPD: A=:D; T:=HDEV+DST; *IOXT % READ STATUS). Several docs claim an "ECAPD capacity check" rejected a capacity value. There is no such check. Ignore it entirely.
Reporting a different capacity cannot fix this. The real capacity (UHLIM) comes from INSIDE the on-disk control record (line 402), never from READ CAPACITY. Capacity only selects WHICH block is read as the control record. 122071 and raw 129311 both failed - as this model predicts.
The ND-3201's Z80 firmware (45900E.bin) is NOT in the SCSI path. The board is two independent halves: Z80 + AM9517 + FD1797 = FLOPPY; the NCR 5386 is hardware-decoded onto the ND-100 IOX bus and driven by SINTRAN. Verified by an exhaustive I/O port sweep: the Z80 never touches the NCR. Do not reverse-engineer that ROM for this bug. Details: SINTRAN/Devices/SCSI/nd-scsi-3201.md.

2. THE GAP - THIS IS WHAT YOU ARE CARVING¶

Existing carve: SINTRAN/Filesystem/code-logic/enter-directory.md (501 lines) covers ENTER-DIRECTORY end-to-end: command dispatch -> unit reserve + directory datafield -> the page-0 read (RXDIR -> RCBLO) -> master-block parse/validation (CHDSI) -> success/error exits.

But it declares its own boundary OPEN. Quoting that document: - line 268: RCBLO - the "Octobus / SCSI driver transfer primitive" - is a "resident overlay, not part of [the carved] segment". - line 290: | Driver function code / DMA count / SCSI opcode+LBA | set inside the resident driver | **OPEN - uncarved boundary** |

So: we have the filesystem down to the moment it calls the driver, and we have the SCSI driver from SCLLD onward. The code that joins them is uncarved. That is where the missing page-0 request dies.

Known anchors (from `enter-directory.md`, verify them - do not trust blindly)¶

RCBLO at 35766B (octal).
The device-transfer dispatch sites: 036135, 036142, 036167, 036174 - all JPL I ,B 10 (036142 is annotated "alternate read/format entry").
RXDIR - reads via the page cache (R3BUF), section 4.1 of that document.
Cross-reference document (NOT yet read - read it first): SINTRAN/Filesystem/DEBUG-scsi-enter-directory.md

3. THE LEADING HYPOTHESIS (UNVERIFIED - test it, do not assume it)¶

A page/block unit mismatch of exactly 2.

enter-directory.md (lines 286, 292): a SINTRAN filesystem page = 1024 WORDS = 2048 BYTES.
The SCSI logical block is 1024 BYTES.
Therefore page 0 = TWO SCSI blocks (LBA 0 and LBA 1), not one.
But INQUI computes its page<->block shift instructions against a 1024-BYTE reference: line 1267 X:=12 % RECORD SIZE 1024 BYTES (0o12 = 10 = log2(1024 bytes)). With blockSize=1024 the loop yields shift = 10 - 10 = 0, so SUSI1/SUSI2/SUSI3 come out as SAD 0 = NO SHIFT.
If some layer expects those shift instructions to convert filesystem pages (2048 bytes) to device blocks (1024 bytes), a no-shift converter is wrong by a factor of 2.

Whether that mismatch is benign (a layer above simply issues 2 blocks per page) or fatal is exactly what the carve must determine. Do not assert it either way without evidence.

Other candidates worth keeping open (do not fixate on the one above): - Something between function 42's success return and RXDIR/RCBLO aborts on the data function 42 produced (UHLIM=121560, NPART=8, the 48-word partition table at CMADR, SUTYP flags). - 5SCIN ("initialization finished") handling - function 42 forces a re-INQUIRY by clearing it (IP-P2-SCSI-DISK.NPL lines 337-340); check the state machine does not deadlock or re-enter. - RXDIR's page cache (R3BUF) deciding it already has page 0, or failing to allocate a buffer.

4. DELIVERABLES¶

Produce a golden-path analysis folder in the carver tree, matching the existing mon-analysis convention (<Name>.ASM, <Name>.pseudo.c, README.md):

tools/sintran-segment-carver/versions/L-VSX-500/re/<...>\RCBLO\

Content required:

RCBLO.ASM - the carved, commented disassembly of RCBLO and the dispatch sites 036135/036142/036167/036174, plus whatever resident overlay they land in. Byte-verified.
RCBLO.pseudo.c - readable pseudo-code of the logic.
README.md covering:
- The call chain, closed end to end: ENTER-DIRECTORY -> RXDIR -> RCBLO -> ??? -> SCLLD. Name every routine and address. This is the whole point - the two existing analyses must MEET.
- What RCBLO does with: the ABFUN function code, the DMA word count, the memory address, the block/page number, and the SUSI1/SUSI2/SUSI3 shift instructions. Show the actual conversion arithmetic from "filesystem page N" to "SCSI LBA".
- THE ANSWER: the exact instruction/decision that prevents SCLLD being called for page 0 after function 42 returns success. If you cannot find it, say so explicitly and state what would settle it.
- Verification of the factor-2 hypothesis in section 3: confirmed, refuted, or undetermined.
- A VERIFIED / INFERRED / OPEN summary table (follow enter-directory.md section 8's style).
Update SINTRAN/Filesystem/code-logic/enter-directory.md: close or narrow the OPEN boundary at line 290 and cross-link the new folder.

5. CARVING TRAPS (learned the hard way - heed these)¶

The MON dispatch table is MCTAB@005620B, NOT GOTAB. Byte-verified.
Validate any table against known slots before trusting it.
Pick the overlay by sibling coherence, not by a single entry test.
STD-I is NOT a universal entry test. Do not use it as one.
56 segments have never been disassembled - the routine may live in one of them.
RCBLO is a resident overlay. Resolve which overlay is actually mapped at the time ENTER-DIRECTORY runs; do not assume the first plausible one.
The NPL sources are a DIFFERENT REVISION from the running L binary. Treat NPL logic as reliable and NPL addresses as indicative only. The carved binary is ground truth.

6. CONSTRAINTS¶

NEVER ASSUME. If you do not know, say "I don't know." Label every claim VERIFIED (you read the bytes) vs INFERRED (you reasoned). Four confident-but-wrong conclusions have already died this session - each one was killed by looking at actual data. Prefer "undetermined + here is the one check that settles it" over a confident narrative.
Cite addresses (octal, as SINTRAN writes them) and quote the actual instructions.
Give full absolute paths for every file mentioned or created.
ASCII only in .ASM / .md / anything fed to period tools. No Unicode, no em dashes, no arrows - use - and ->.
Never mention any AI assistant or tool in files, comments, or commits.
Keep as many comments as possible in the carved output; never delete an existing correct comment.

7. KEY FILE PATHS¶

What	Path
Carver tree (L-VSX-500)	`tools/sintran-segment-carver/versions/L-VSX-500/`
ENTER-DIRECTORY carve (has the OPEN boundary)	`SINTRAN/Filesystem/code-logic/enter-directory.md`
SCSI-side cross-ref (read first)	`SINTRAN/Filesystem/DEBUG-scsi-enter-directory.md`
SCSI disk layer NPL (function 42, control record)	`SINTRAN/NPL-SOURCE/NPL/IP-P2-SCSI-DISK.NPL`
SCSI driver NPL (SCLLD/SCINT/SELEC)	`SINTRAN/NPL-SOURCE/NPL/IP-P2-SCSI-DRIV.NPL`
Board reference (corrected)	`SINTRAN/Devices/SCSI/nd-scsi-3201.md`
Investigation handoff	`SINTRAN/Devices/SCSI/SCSI-CONTROLLER-PWSH-HANDOFF.md`
Device trace	`C:\Users\ronny\AppData\Local\trace\file-trace.txt`
Disk image under test	`E:\Dev\Ronny\RetroFS\demo\test-images\ndfs\tor-disk.img`
L07 symbols	`SINTRAN/NPL-SOURCE/SYMBOLS/L07/FILSYS-SYMBOLS.SYMB.TXT`
Corrected mount doc (last-block = control-record connect; bug = block 0 not read)	`SINTRAN/Filesystem/code-logic/scsi-mount-geometry.md`