DLM rung-2: real cross-node $ENQ GRANT (vms-e8f1 / vms-17c, H4)

[baron-3dl]

H4 — the crown of the 0.6 DLM ladder: the first rung that MODIFIES THE EXECUTIVE

Makes vms_lock_dlm_xnode_dispatch (src/kernel-core/vms_lock.c) actually GRANT an inbound cross-node $ENQ for a resource mastered on this node, instead of the rung-1 honest SS$_UNSUPPORTED (2296) stub. Grounded on docs/research-alpha-dlm-wire.md §7 (rung-2 GRANT semantics from public $ENQ/$LCKDEF/IDSM/Cluster Systems sources). Clean-room Rule 8 respected — no VSI/HPE disassembly.

Executive (kernel-core; rides the 3-way gate {x86_64, VAX ILP32, Alpha LP64})

• The grant: the ENQ case marshals the decoded request into vms_enq_core() with owner_csid = req->req_csid, so the master's lock record is held for the REMOTE requester's cluster identity, not the local delivery daemon. Compatible → SS$_NORMAL.
• Scope fence (INV-6, everything past the foundation fails honestly): LCK_M_NOQUEUE fences out cross-node contention (incompatible → SS$_NOTQUEUED, no wait queue); no SYNC (never blocks the delivery thread); no VALBLK/AST (LVB replication + blocking-AST are later rungs). DEQ/GRANT/BLKAST still decline with SS$_UNSUPPORTED.
• vms_lock_entry gains req_csid (the CSID a lock is held for; 0 = local), stamped in vms_enq_core. GET_RESMASTER reports remote_holder_csid so a test can PROVE the master genuinely holds a lock for a peer. Both fields reuse existing struct pad — no ABI size change, static_asserts unchanged. Mirrored into the NetBSD headers.

Transport (SCSD)

On a granted ENQ, scsd_dlm_dispatch_to_executive reads B's own lock DB back (GET_RESMASTER, same registered fd) and prints SCSD-I-DLMHELD — a READ of real executive state. DLMRX/DLMGRANT/DLMDONE wording is now status-aware. H0 --dlm-selftest switches ENQ→DEQ so it stays a side-effect-free "executive reached" probe (still 2296) — H0 harness untouched.

Proof — H4 harness (dlm-harness-h4 CI job)

Two real-/dev/vms QEMU nodes join, node A issues a cross-node $ENQ for a B-mastered resource. Ran locally under KVM — PASSED:

node B: SCSD-I-DLMRX  ... executive status=0x00000001 (GRANTED)
node B: SCSD-I-DLMHELD, master DB: resnam='RESONE' found=1 is_local_master=1 n_granted=1 held_for_csid=1601 master_csid=1
node A: SCSD-I-DLMDONE ... GRANT status=0x00000001 ... cross-node lock GRANTED by the master (SS$_NORMAL)

DLM HARNESS H4 PASSED: cross-node $ENQ GRANTED rc=SS$_NORMAL, B holds lock for A's CSID
  B_DISPATCH_STATUS=0x00000001  held_for_csid=1601 (A_CSID=1601)  n_granted=1  is_local_master=1

held_for_csid=1601 is node A's real SCSSYSTEMID (OVMXA) — B's resource DB genuinely holds the lock for A's cluster identity, not a fabricated SS$_NORMAL.

• test_kmod_dlm_xnode retargeted to the foundation grant + a contention scope-fence (SS$_NOTQUEUED) proof.
• H3's verdict updated to the grant (SS$_NORMAL) so it stays green (re-ran locally — PASSED).

Alpha caveat

Alpha vms.ko is P0-broken on main (rd vms-d38b / #835, unrelated). This change adds no new Alpha break; x86_64 grant proof above is the authoritative leg. Conductor release-gates Alpha convergence once the P0 lands.

🤖 Generated with Claude Code