vms-d69 (P4-VFS V2b): extract ODS-2 allocator + dir scanner + header codec to vmsfs-core over the block/inode seam

[baron-3dl]

What

Builds seam #2 — the block/inode seam designed in src/kernel/vmsfs/README-backend.md §3 (rd vms-d69, epic vms-8e8), completing the ODS-2 core extraction. The block-coupled ODS-2 logic that still lived in vmsfs_blkdev.c — the storage/FID/cluster allocator, the directory-block scanner, and file-header read/write — moves into src/kernel-core/vmsfs/, rewritten against a substrate-neutral vmsfs_bio vocabulary. After this, src/kernel/vmsfs/ is the thin VFS + vmsfs_bio backend; all ODS-2 logic is in the core.

The seam as built (vs README §3)

• vmsfs_bio (block half, §3a): struct vmsfs_volume (geometry + bitmap + opaque host) embedded in vmsfs_sb_info; opaque struct vmsfs_bh; ops vmsfs_bget/vmsfs_bdata/vmsfs_bdirty[_sync]/vmsfs_bput, vmsfs_bit_find_zero/set/clear/test, vmsfs_le*_to_cpu/vmsfs_cpu_to_le*, vmsfs_now_seconds — declared in the core vmsfs_bio.h, realized as trivial forwarders in vmsfs_backend_linux.h (sb_bread/mark_buffer_dirty/find_next_zero_bit/le*_to_cpu/ktime_get_real_seconds). Mirrors the exec-core exec_kbackend.h pattern.
• vmsfs_fh_info (inode half, §3b): POD + pure vmsfs_fh_decode/vmsfs_fh_encode; the backend keeps only the thin POD <-> struct inode copy.
• Two design refinements (noted in README §3): vmsfs_fh_decode returns an enum vmsfs_fh_status (so the backend reproduces the original's three distinct diagnostics/errnos: EIO magic, EIO checksum, ENOENT not-in-use) rather than a bare int; and flush/rename use explicit partial-update writers (vmsfs_fh_write_meta/vmsfs_fh_write_rename) rather than a decode-on-write, preserving every untouched on-disk field.

New core files (0 <linux/>): vmsfs_bio.h, vmsfs_alloc.c, vmsfs_dirscan.c, vmsfs_header.c.

Behaviour: ZERO change (Rule 6/7)

Behaviour-preserving restructure; guarantee is BEHAVIOUR, proven by the QEMU oracle. test_kmod_vmsfs_blkdev + every vmsfs suite is byte-identical baseline vs after:

suite	baseline	after
test_kmod_vmsfs	19/0	19/0
test_kmod_vmsfs_blkdev	51/0	51/0
test_kmod_vmsfs_exepath	28/0	28/0
test_kmod_vmsfs_mountvis	7/0	7/0
test_kmod_vmsfs_readdir	8/0	8/0
test_kmod_vmsfs_rename	26/0	26/0
test_kmod_vmsfs_sysgroup	4/0	4/0
FINAL	70 suites / 0 failed	70 suites / 0 failed

make -C src/kernel/vmsfs compiles clean before and after.

Invariants

• src/kernel-core/vmsfs/** has 0 <linux/> includes (grep-verified).
• src/kernel/vmsfs/** contains no ODS-2 allocator/dir/header algorithm — only VFS + the vmsfs_bio impl. (find_next_zero_bit/set_bit/… gone from the backend; sb_bread/mark_buffer_dirty remain only as VFS block I/O around core calls.)
• No facility_defects.sh anchor repoint needed: its coverage glob is kernel-core/*.c (non-recursive), so kernel-core/vmsfs/*.c is unscanned — consistent with the existing vmsfs core files; the kernel/vmsfs scope-out is unaffected (vmsfs_blkdev.c stays there with vmsfs_blkdev_rename intact).

This completes the ODS-2 core extraction. V4 (NetBSD vnode backend, rd vms-308) can now be written 1:1 against the seam (README §4 checklist updated).

🤖 Generated with Claude Code