mem: NUMA-style page-migration acceptance example

[syifan]

What

A multi-CPU page-migration acceptance test at mem/acceptancetests/pagemigration, plus a latent datamover bug fix it depends on.

NUMA model

Several memaccessagents sit on a shared hierarchy whose physical address space is split across two memory devices. The L2 routes each physical address to the owning device's controller, so a remote access just works over the interconnect — migration is a transparent performance optimization, not a correctness requirement (the test passes with -migrate=false). Each page reserves a home slot on both devices at the same in-device offset, so migrating is a PAddr/DeviceID flip with no frame allocator.

Migration controller

A periodic (round-robin) controller relocates a page between devices, driving the existing memcontrolprotocol control ports:

drain ROBs → pause AT/L1$/L2$/L1TLB/L2TLB/MMU → flush write-back L2
  → data-mover copy (src→dst) → repoint page table → invalidate caches+TLBs → resume

The ordering is the correctness argument: drain quiesces in-flight writes, flush makes memory authoritative before the copy, invalidate drops stale mappings/lines after the repoint. The agents' value-checks are the oracle — any non-transparent migration yields a read mismatch.

Validation

• Passes serial + parallel engines, 2/4/8 agents, various sizes and intervals (up to ~80 migrations/run).
• Negative control: disabling the flush step makes the oracle panic (Mismatch when read), confirming the test actually catches migration corruption.
• Registered in mem/acceptance_test.py (baseline-no-migration + migration matrix, incl. 4-agent and parallel).

datamover fix (included)

readFromSrc compared the absolute read address against the relative buffer window, so a move whose SrcAddress was at/beyond BufferSize issued zero reads and hung. Every existing data-mover test used SrcAddress=0, so it was latent. Fixed to compare in transaction-relative space + a regression test for a non-zero source address. (Happy to split this into its own PR if you'd prefer — the example just needs it to land first.)

Notes / follow-ups

• The controller currently ticks each cycle during the idle countdown and stops by inspecting agent state. A cleaner approach (a "background event" engine property so the engine terminates when only heartbeat events remain, plus future-scheduling instead of per-cycle ticks) is planned as a separate core-engine change; this PR is the working baseline.

go build ./..., golangci-lint, and the data-mover tests pass.

🤖 Generated with Claude Code