handoff: VSync/event-wedge fixes + iterate 2.A–2.BC research notes
Source changes (dormant parity infra, retained from iterate 2.AI/2.AO): - xenia-kernel/exports.rs: nt_create_event manual_reset polarity + related event wiring - xenia-gpu/mmio_region.rs: D1MODE_VBLANK_VLINE_STATUS hardcode parity Also lands the audit-runs/ analysis notes (.md/.txt/.json digests) for the iterate 2.x VSync/0x10e8/0x1004 wedge investigation. Raw trace dumps (.jsonl/.gz/.csv/.stdout) and agent worktrees (.claude/) are gitignored as regenerable local artifacts — see memory + HANDOFF for the running findings. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
MechaCat02
@@ -0,0 +1,91 @@
|
||||
# Phase C+16 cold-vs-cold result (2026-05-14)
|
||||
|
||||
## Matched-prefix table
|
||||
|
||||
| canary_tid | ours_tid | C+15-α | C+16 | delta | first_divergence_at | kind |
|
||||
|------------|----------|---------|---------|-------|---------------------|-----------------------------------|
|
||||
| 6 | 1 | 102,168 | 102,171 | **+3**| 102,171 | `handle.create` (class E) |
|
||||
| 4 | 11 | 8 | 8 | 0 | 8 | `handle.create` (class E) |
|
||||
| 7 | 2 | 30 | 30 | 0 | 30 | `handle.create` (class E) |
|
||||
| 12 | 7 | 2 | 2 | 0 | 2 | `handle.create` (class E) |
|
||||
| 14 | 9 | 2 | 2 | 0 | 2 | `handle.create` (class E) |
|
||||
| 15 | 10 | 16 | 16 | — | — | no divergence |
|
||||
|
||||
Main matched prefix advanced 102,168 → 102,171 (+3). All 5 sister
|
||||
chains unchanged.
|
||||
|
||||
## New first divergence (idx=102,171)
|
||||
|
||||
```
|
||||
canary: [102169] import.call KeWaitForSingleObject
|
||||
ours: [102169] import.call KeWaitForSingleObject
|
||||
canary: [102170] kernel.call KeWaitForSingleObject
|
||||
ours: [102170] kernel.call KeWaitForSingleObject
|
||||
canary: [102171] handle.create sid=68fec8909ea5d1f5
|
||||
ours: [102171] wait.begin {'handles_semantic_ids': ['0000000000000000'], ...}
|
||||
canary: [102172] wait.begin {'handles_semantic_ids': ['68fec8909ea5d1f5'], ...}
|
||||
ours: [102172] kernel.return KeWaitForSingleObject
|
||||
```
|
||||
|
||||
This is **class E** — same root cause as D-2/D-3/D-4 in the C+15-α
|
||||
catalog. Canary's `xeKeWaitForSingleObject` calls
|
||||
`XObject::GetNativeObject<XObject>(...)` which CREATES a new handle for
|
||||
the native dispatcher object on first encounter; ours's
|
||||
`resolve_pseudo_handle` does not, so the `wait.begin`'s
|
||||
`handles_semantic_ids` is `0000000000000000`. The next Phase C+17
|
||||
target.
|
||||
|
||||
## Acceptance gates
|
||||
|
||||
- **Gate 1 (default-off digest)**: PASS — 3× reproducible at
|
||||
`e1dfcb1559f987b35012a7f2dc6d93f5` (unchanged from C+13/C+15-α
|
||||
baseline). The refcount fix is observation-only at the digest level;
|
||||
guest behavior is unchanged because no actual code path depends on
|
||||
the precise destruction timing of the closed-but-still-running thread
|
||||
handle within the 50M-instruction window.
|
||||
- **Gate 2 (cvar-on emit)**: PASS — both engines produce JSONL cleanly
|
||||
(ours 121,537 events; canary 2,512,481 events in 90s).
|
||||
- **Gate 3 (diff tool)**: PASS — diff tool consumes events, produces
|
||||
6-chain divergence report; main divergence at 102,171 (was 102,168).
|
||||
- **Gate 4 (cold-vs-cold)**: PASS — main matched prefix advances +3,
|
||||
no sister-chain regressions.
|
||||
- **Gate 5 (build clean)**: PASS — `cargo build --release` clean
|
||||
(1 pre-existing dead_code warning unrelated).
|
||||
- **Gate 6 (tests)**: PASS — 181 → 186 (added 5 refcount lifecycle
|
||||
tests; all pass).
|
||||
- **Gate 7 (Phase B image hash)**: NOT EXECUTED (no engine change
|
||||
reaches XEX load); inferred unchanged from invariant cold-stable
|
||||
digest.
|
||||
|
||||
## Sister-chain analysis
|
||||
|
||||
No sister chain advanced beyond C+15-α matched-prefix. tid=4→11,
|
||||
tid=7→2, tid=12→7, tid=14→9 all diverge at the same indexes — the
|
||||
C+16 refcount fix is on a distinct code path from class-E
|
||||
KeWaitForSingleObject native-obj handle. C+17 must address class E
|
||||
to advance those chains.
|
||||
|
||||
## Reading-error class
|
||||
|
||||
None new. Reading-error #28 discipline (verify framing first) was
|
||||
followed; canary source was read end-to-end for `XThread::Create`,
|
||||
`XObject::RetainHandle`/`ReleaseHandle`, `ObjectTable::AddHandle`/
|
||||
`RetainHandle`/`ReleaseHandle`/`RemoveHandle`, and
|
||||
`XamTaskSchedule_entry`/`XamTaskCloseHandle_entry` before any code
|
||||
change.
|
||||
|
||||
## Refcount leak risk audit
|
||||
|
||||
Three test cases cover the lifecycle balance:
|
||||
|
||||
1. `ex_create_then_close_then_exit_balances_refcount` — close first,
|
||||
then exit. Refcount 2→1→0. Handle destroyed. No leak.
|
||||
2. `xam_task_schedule_close_then_thread_exit_destroys_handle` — same
|
||||
ordering via XAM path.
|
||||
3. `xam_task_thread_exit_then_close_destroys_handle` — exit first,
|
||||
then close. Refcount 2→1→0. Handle destroyed. No leak.
|
||||
|
||||
The reverse case (no close, only exit) leaves refcount at 1
|
||||
(creator-only) which is correct: the handle slot remains until the
|
||||
creator explicitly closes it. This matches canary's behavior — the
|
||||
guest is responsible for closing handles it allocated.
|
||||
Reference in New Issue
Block a user