vms-648: cache the ACP backing-device handle so PRODUCT INSTALL stops timing out (R1 e2e)

[baron-3dl]

Summary

The R1 release-install e2e (vms-37f) went red 6 passed / 10 failed on main. The symptom read as non-persistence (SCSNODE, SYSTEM credential, USERS.* home dirs all failing to land on the target), but the real cause is upstream of any write: container 1 hangs at Configuring OVMX ... OpenVMS Operating System and never reaches %PCSI-I-DONE within the harness's 90s per-step timeout. AUTHORIZE / SYSGEN / the USERS tree never run — the "non-persistence" is a downstream consequence of PCSI being killed mid-install.

This is NOT a code regression in the window

Built the bootable image at the known-good baseline 7811bf30 and the known-bad tip 9c507bc6 and ran the install under --cpus=1 (the slow single-CPU TCG profile a GitHub runner presents):

image	install under --cpus=1
baseline 7811bf30	FATAL, 7/21 — hung at Configuring
bad-tip 9c507bc6	FATAL, 7/21 — hung at Configuring
this fix	21/21, full install in 72s

No commit in 7811bf30..9c507bc6 touches the ACP write hot path (the two vms_lock.c DLM changes only alter cross-node xn-guarded code; #846/#870/#871 are Alpha-tooling test-only; the OS kit is byte-identical 4.4M at both ends). The CI "baseline green / bad-tip red" was runner-speed variance across a razor-thin margin, not a product change. Confirmed against the actual failing CI job (run 33218016964, job 99005938554) and the green baseline job (run 33190974044, job 98928619663).

Root cause

exec_blockdev_read_block/write_block (exec_kbackend_linux.h) open AND close the backing block device (bdev_file_open_by_dev + fput) and alloc a page for every 512-byte LBN. PRODUCT INSTALL writes the ~4.4M OS kit as thousands of blocks — thousands of full bdev open/close cycles — which under single-CPU TCG exceeds 90s. The NetBSD twin already caches its backing vnode (vms_blockdev_netbsd.c); this is the missing Linux equivalent.

Fix

A module-global backing-device handle cache (exec_bdev_get_cached / exec_bdev_cache_release_all in vms_module.c): each device opened once (READ|WRITE, non-exclusive), keyed by dev_t, reused for every block; released together at module exit (vms_exit), the OS-lifetime analogue of NetBSD's release-at-detach. Slots are only added, never evicted mid-life, so a handed-back block_device stays valid across the bio (which runs outside the cache lock) — no use-after-free. A full table or failed open returns NULL and the primitives fall back to the original open-per-call path, so correctness never depends on the cache. Only the OPEN is amortized; every block is still a real synchronous bio to a real device (INV-6 / Rule 8). The test is unchanged — not weakened.

Proof (real /dev/vms, qemu-system-x86_64, TCG, --cpus=1)

Full 16-check R1 gate, all containers --cpus=1, this fix: RESULTS: 16 passed, 0 failed — SCSNODE OVMXR1 persists, SYSTEM logs in with the install-set password, PRODUCT SHOW + DIRECTORY resolve, USERS home dirs resolve, negctl boundary intact.

NetBSD unaffected (backend already caches; new symbols are Linux-only). Standalone vms.ko builds+modposts on 6.8 (pre-6.9 branch); in-tree bootable module builds+modposts on 6.12 (post-6.9 branch).

🤖 Generated with Claude Code