vms-c60: ACP answers IO$_READVBLK / IO$_WRITEVBLK (virtual-block I/O, implicit extend, epic vms-208)

[baron-3dl]

Fifth rung of the Files-11 ODS-2 ACP-in-the-executive (epic vms-208). Builds on the IO$_ACCESS window rung #633 (vms-204), which is already merged to main — this branch was rebased onto origin/main (additive, no conflicts).

The model

The ACP now services virtual-block transfers on an ACCESSED file channel, mapping {VBN, byte-offset, length} through the channel's window (built by IO$_ACCESS) to LBN block I/O over a real /dev/vms.

• IO$_READVBLK (ioctl 0x6D): read through the window, clamped at EOF. Byte-exact vs the userspace codec. Read starting past EOF → SS$_ENDOFFILE.
• IO$_WRITEVBLK (ioctl 0x6E): write through the window to the mapped LBNs. A write whose end lies past the file's highest allocated VBN triggers an implicit EXTEND — allocate the shortfall from BITMAP.SYS, append an FM2 retrieval pointer to the file's FH2, grow HIBLK/EOF, reseal the header checksum, extend the channel window. Read-only channel → SS$_NOPRIV; no free run → SS$_DEVICEFULL; bad offset → SS$_BADPARAM.

(0x6F left reserved for IO$_ACPCONTROL / wildcard $SEARCH.)

Byte-exact proof — real /dev/vms (QEMU kernel-executive harness), 20/20 PASS

tests/qemu/test_syssvc_acp_rw.c, over the real-VAX ODS-2 fixture on DKA0::

• IO$_READVBLK returns all 17218 valid bytes of [OVMXDIR]HELLO.TXT byte-exact vs the committed codec golden.
• IO$_WRITEVBLK in place round-trips byte-exact, and persists across DEACCESS + re-ACCESS (window rebuilt from the on-disk FH2 → it hit the platter, INV-6).
• IO$_WRITEVBLK past EOF: extended=1, HIBLK 35, EOF 35; after re-ACCESS the FH2 reports the grown size and VBN 35 reads back byte-exact — a real BITMAP.SYS block was allocated and the retrieval map grew, on disk.
• Fail-honest: read-past-EOF → SS$_ENDOFFILE; write-to-read-only-channel → SS$_NOPRIV; bad offset → SS$_BADPARAM.

The full harness ran green (87 suites passed, 0 failed) on this branch's predecessor source; the only delta to the final commit is an ODS2_BLOCK_SIZE→ACP_BLOCK_SIZE rename (same value, 512, made the codec-free path independent of the gated header) plus a comment fix — CI re-verifies the final source.

Clean-room (Rule 8)

Every on-disk format operation (FM2 map append, RECATTR EOF/HIBLK, storage-bitmap bit alloc, FH2 checksum) is a pure codec helper in the new src/vmsfs/ods2/ods2_edit.c — the write-side twins of the reader's pure parsers. vmsfs_acp.c only sequences the raw block I/O around them, exactly as #633's IO$_ACCESS sequences the pure parse helpers. exec_kbackend gains a write twin exec_blockdev_write_block (Linux bio; NetBSD contract-only stub).

Dual-build (codec-free overlay must still link — #623 class)

All codec use is gated behind OVMX_ODS2_KERNEL. Verified BOTH:

• out-of-tree make -C src/kernel (with codec) → vms.ko builds + modposts clean;
• codec-free single-TU compile of vmsfs_acp.o (overlay mode, no codec) → clean, and nm shows zero ods2_ references (no dangling symbol). ods2_edit.o stays out of the bootable overlay, like ods2_reader.o.

Cascade compliance

• New executive kif symbols vms_kif_acp_readvb/_writevb appended to src/vmslink/libvmssys_shr.vec (append-only); symvec_freeze_gate + kif_caller_census green (OVMX-UNWIRED declared).
• Negctl anchor acp-writevb-extend-alloc-offbyone in facility_defects.sh (a real FM2 map-append LBN off-by-one the suite's re-ACCESS read-back catches); FLOOR-NO-BUMP; coverage + selftest pass.
• Full Debug ctest green (191/193; env_identity_census_negctl is the documented ~180s flake vms-3f9 under concurrent build load — passes in isolation).

Concurrency note: the extend's allocate/RMW span is single-writer for now (stated, not faked); the per-volume DLM synchronization lock (vms-233, a later rung) is the VMS-authentic serialization and supersedes this.

🤖 Generated with Claude Code