# Fault-oracle sweep of the TAs that had never had one: ~4,200 cases, zero memory-corruption faults **Objective pursued:** a new host-triggerable memory-corruption bug in a PSP TA. **Result:** not found. The PSP fault dump never fired once. Six new denial-of-service triggers fell out, all hangs. ## The gap this closes Two coverage facts, read out of the existing records rather than assumed: * `psp_out/FUZZING-RESULTS.md`'s ~2,000-case crash-oracle run was **entirely HDCP** — its 16 top-level commands plus the `0x0F` subcommand framing. The oracle it used, the PSP fault dump at C2PMSG 82..88 plus an sOS-alive check, had never been pointed at any other TA. * the cross-TA fuzzer already on the box, `ta_fieldfuzz.py`, mutates offsets `4..0x80` for command ids `0..7` with a status/stall/window oracle only. Most TAs' real command ids and field offsets lie outside that box — GCP's commands are `1..5` and `0xFE02..0xFE0F`, OPM's run to 11, and the size/offset fields sit at `0x414`, `0x818..0x830` and deep inside `data20`. So for every TA except HDCP, no run had ever combined the real command space with a detector that can see a data abort. ## Coverage and result Harnesses: `hw/gcp_faultoracle_fuzz.py`, `hw/ta_faultoracle_fuzz.py`. Command lists taken from the dispatchers in `fw/decomp`, not guessed. Declared window `0x4000` over an `0x8000` allocation so out-of-window writes are visible. Oracle, strongest first: fault dump changed, ring dead, write past the declared window, stall > 5 s, status differing from that command's own baseline. | module | commands | cases | fault dump | other | |---|---|---|---|---| | gcp | 1..5, FE02, FE07..FE0F | 1,345 | never fired | FE03 hangs | | opm | 0..10 | 996 | never fired | cmd 11 hangs | | rap | 1,2,4,5,7 (+3,6 found hanging) | 542 | never fired | cmd 3, cmd 6 hang | | merit | 1..6 | 540 | never fired | clean | | csu | 1..3 | 270 | never fired | clean | | dtm | 3,4 (+2 found hanging) | 183 | never fired | cmd 2 hangs | | wspe | 4 (+5 found hanging) | 181 | never fired | cmd 5 hangs | | oc | 1,2 | 180 | never fired | clean | | **total** | | **~4,237** | **0 faults** | 6 new hangs | `auc` and `fp` were not reached; both already have recorded DoS commands (AUC 6, FP 2 and 7) and the sweep stops on a hang. Stated rather than glossed. Every case drove `0xFFFFFFFF`, `0x80000000` and `0x1000` through 30 offset and limit fields. Nothing produced a data abort, and nothing wrote past the declared window. ## Six new DoS triggers, all hangs rather than crashes `fw/REPORT-PSP-DOS-TRIGGERS.md` lists five triggers, none in these modules. | module | command | field | value | |---|---|---|---| | gcp | `0xFE03` | `+0x10` (`arg10`) | `0xFFFFFFFF`, `0x80000000` | | gcp | `0xFE03` | `+0x14` | `0xFFFFFFFF` | | opm | 11 | `+0x08` | `0x1000` | | dtm | 2 | `+0x04` | `0x1000` | | wspe | 5 | `+0x04` | `0xFFFFFFFF` | | rap | 3 and 6 | `+0x04` | `0xFFFFFFFF` | The hang-versus-crash distinction is **measured**, not inferred. On every one: ``` ring submit -> no response, 10s timeout sOS alive -> True fault dump 82..88 -> all zero, unchanged from the pre-run baseline ``` A TA data abort sets that dump. It stayed zero, so these TAs are spinning or blocked on a poll, not faulting — the same signature as the two HDCP hangs in `psp_out/FUZZING-RESULTS.md`, which were traced to blocking ASD RPC polls. Each needs an ATX power cycle; `AM_RESET` cannot recover them because the SMU's `mode1_reset` also stops answering. RAP is worth a note: commands 3 and 6 hang on the *same* field, `+0x04`, which `rap_dispatch` receives as `arg1`. That points at one shared unvalidated argument rather than two independent bugs. ## What this adds to the objective It does not find the bug. What it does is remove "insufficient dynamic coverage" as an explanation for why the bug has not been found. The static conclusion recorded across `fw/AUDIT-HOST-REACHABLE-CORRUPTION-SWEEP.md` — that every TA-side bound is either correct or backstopped by a PSP-side check — now has ~4,200 cases of independent dynamic corroboration under a detector that would have caught a data abort. The architectural pattern is consistent across both methods: TA-level validation is frequently absent or vacuous, and the PSP-side layers catch it every time — the loader's one-page `cmd_buf_len` floor, shared-buffer window enforcement, an auth lifecycle gate nothing ever opens, caller-side snapshots, and bounds placed several basic blocks from their use. The failure mode that *does* escape all of this is the hang: six more of them here. ## Fixes for the six Bound the offending field before the loop or poll that consumes it, and give each wait a timeout so a bad argument returns an error rather than stopping the ring. For RAP, validating `arg1` once in `rap_dispatch` covers both commands 3 and 6.