rebbarb/docs/bba_compat_todo.md

# BBA compatibility TODO

Gaps between this implementation and what real GC software expects, found by
reading the three drivers that actually drive a BBA:

- **Swiss (in-game patches)** — `cube/patches/bba/bba.c` in
  [emukidid/swiss-gc](https://github.com/emukidid/swiss-gc). Note that
  directory holds one file per adapter (`bba.c` is the real DOL-015;
  `enc28j60.c`, `w5500.c`, `w6100.c`, `w6300.c` are other hardware) — check
  which file you are in.
- **libogc2** — `lwip/arch/gc/netif/gcif.c` in
  [extremscorner/libogc2](https://github.com/extremscorner/libogc2). This is
  what Swiss's own menu networking uses, via `if_configex()`.
- **stock devkitPro libogc** — same shapes as libogc2; read out of
  `/opt/devkitpro/libogc/lib/cube/libbba.a` (`gcif.o`) with `objdump`.

⚠️ **Games use the official Nintendo SDK driver, which none of us can read.**
Everything below is derived from the three open drivers plus the register map
they document. A game may exercise a mechanism no open driver touches, so
"no open driver does this" is weaker evidence than "the hardware has no such
register".

The register map referenced throughout is the one in libogc2's `gcif.c`
(`BBA_*` defines), which is the official one.

---

## 1. No hardware RX filtering — **it hung Swiss**

**Status: ✅ DONE 2026-10-10** — `S0_MR` now sets the MACRAW **MF** bit, so the
W5100 delivers own-MAC (`SHAR`) + broadcast only, matching what the real BBA
does. `NCRB.PR` (bit 0) clears MF if the GC asks for promiscuous, sampled when
socket 0 is OPENed (i.e. at init, so a later `PR` change takes effect on the
next NCRA reset).

**How it presented, which is worth knowing because it does not look like a
filtering problem.** Swiss hung forever on "Initialising Network". libogc2's
`bba_interrupt()` is

```c
ir = bba_in8(BBA_IR); imr = bba_in8(BBA_IMR); status = ir & imr;
while (status) { bba_out8(BBA_IR, status); ... start_rx ...; status = ir & imr; }
```

— it only exits when `IR & IMR` reaches zero. Unfiltered, every frame on the
segment set `RI`, which re-asserted faster than the loop could clear it, so it
**never exited**. Not slow: livelocked. Unplugging the ethernet cable let Swiss
get all the way through (then fail DHCP, there being no server), which is what
identified it — a one-minute experiment after a lot of theorising.

**Why filtering against `SHAR` was safe after all.** The earlier worry was that
MF filters on `SHAR`, which we write only during init, so a driver that WRITES
PAR0–5 would leave it stale and MF would drop everything. libogc2 **reads** the
MAC out of PAR0–5 rather than writing it, so for this path `SHAR` (written from
`reg.par` in the same init) cannot go stale. Broadcast still passes MF, and a
DHCP OFFER unicast to our MAC passes too.

⚠️ **Still to do for a driver that writes PAR0–5 after init:** refresh `SHAR`
on a PAR write and re-open the socket, otherwise MF filters on the old MAC.
Nothing we have seen does this, but the official SDK driver is unread.

⚠️ **The W5500 back-end was NOT changed** — it is the alternate path and
unused. If it is ever revived it needs the same treatment (`Sn_MR` MF bit).

Tests: `w5100_parallel_master` **T1** (the init sequence expects
`S0_MR = MACRAW|MF`) and **T1b** (`promisc` high re-inits with MF clear).
Mutation-tested: never setting MF fails both, with the message naming the
Swiss hang.

### Original analysis, kept for the reasoning

**Was: OPEN. Needs a decision, not just code.**

The real BBA filters received frames. libogc2 configures
`NCRB = AB|CA|2_PACKETS_PER_INT` = `0x52`, i.e. `PR=0` (not promiscuous),
`PM=0` (no multicast), `AB=1` (accept broadcast) — so the device delivers
**own-MAC + broadcast only**.

We run the W5100 in MACRAW with the **MF (MAC filter) bit clear**
(`_S0_MR_MACRAW = 0x04`, `w5100_parallel_master.py`), so **every frame on the
segment enters the SPRAM ring.**

Observed consequence, already seen on hardware: the test app's ring filled to
page 15 from what looked like "ambient broadcast traffic" but was in fact all
unicast traffic on the segment too. Ring pressure makes RBFI drop frames we
DO want, and a game's lean driver may not expect foreign unicast at all.

⚠️ **Setting MF is not a one-line fix.** MF filters against the W5100's
`SHAR`, and we write `SHAR` only during the init sequence. libogc2 *reads* the
MAC out of PAR0–5 and never writes it, so SHAR never diverges there — but a
game that WRITES PAR0–5 would leave SHAR stale and MF would then drop
everything. Doing it properly needs:

1. decode `NCRB` (see item 3) so `PR`/`PM` can steer the filter;
2. refresh `SHAR` when PAR0–5 changes, not only at init;
3. re-open socket 0 when the filter mode changes (`Sn_MR` is sampled at OPEN).

**The decision:** filter correctly (matches real hardware, fixes the ring
pressure) at the cost of that machinery, or stay a permissive superset where
nothing is ever lost but the ring fills with other hosts' traffic.

## 2. Ring geometry is hardcoded — silent corruption risk

**Status: ✅ DONE 2026-10-10.** `RXFrameAssembler` takes the ring bounds from
`bp`/`rhbp`, wired from the GC's `BP` (0x0A/0B) and `RHBP` (0x1A/1B) registers
in `BBATop`. The wrap and the free-space calculation both use them, so a
driver's smaller ring is respected instead of ignored.

**Garbage is rejected, not obeyed.** The bounds are used only when
`bp >= 1` (page 0 is the TX buffer), `rhbp <= 0x0F` and `rhbp > bp`;
otherwise the hardcoded pages 1–15 stand and `geom_honoured` goes low.
Without that check, `BP = 0` or an `RHBP` below `BP` gives a zero or negative
ring size and the assembler writes outside it.

⚠️ **`geom_honoured` is NOT yet observable from outside the FPGA.** It is
declared on `RXFrameAssembler` and asserted correctly (T7/T7b check it), but
`BBATop` does not wire it to anything — no status register, no LED, no UART
line. So a rejected geometry is currently silent to the GC and to the bench.
Wiring it to the `StatusPanel` or the UART console is a small loose end worth
closing; until then "rejected rather than obeyed" is the guarantee, not
"reported".

⚠️ **A ring must be LARGER than `_MAX_FRAME_PAGES` (6) or nothing is ever
stored.** A frame's length is unknown until EOF, so the assembler reserves the
worst case (1518 + 4 bytes = 6 pages) before accepting one. With a 4-page ring
`free` can never reach 6, so RBFI drops every frame — correct, conservative,
and useless. This was found by writing the test with `RHBP = 4` first, where it
passed **vacuously** because nothing was stored at all. If a real driver ever
configures a ring that small, the reservation has to become length-aware
(accept a frame whose actual length fits), which is a bigger change than this
one.

Tests: `rx_frame_assembler` **T7** (an 8-page ring is honoured, frames land
only in pages 1–8) and **T7b** (all three garbage cases rejected, and the
fallback layout still stores). Mutation-tested: wrapping at the hardcoded page
15 writes pages 9–11 and fails T7; removing the sanity check fails all four
T7b assertions.

⚠️ **This change cost a 0/8 timing failure before it was tuned, and the shape
of that is worth remembering.** The first version computed the sanitised
bounds combinationally, which put

```
bp/rhbp -> comparators -> Mux -> subtract -> pg_total -> add -> raw
        -> compare -> free -> compare -> have_room
```

in ONE combinational path where it had previously been constant-folded. `clk`
fmax collapsed **29 → 17 MHz and all eight seeds failed** the 24 MHz target.
Two things were needed:

1. **Register the bounds** (`pg_first`/`pg_last`/`pg_total`, and `rrp_s`).
   They change only when the GC writes `BP`/`RHBP`/`RRP`, so a cycle of
   staleness costs nothing, and `free` already reaches the FSM through the
   registered `have_room`. That alone got it to ~20 MHz — still failing.
2. **Narrow them to 4 bits.** Ring pages are 1–15. An 8-bit
   variable-vs-variable comparison is a carry chain, where
   variable-vs-CONSTANT folds into a couple of LUTs — which is exactly what
   was lost by making the bounds dynamic. `geom_ok` guarantees
   `bp < rhbp <= 0x0F`, which is what makes the narrow slice safe.

Result: 26.9–28.5 MHz on the two sampled seeds, against 28.0–29.2 for the
build without this change. **Always sweep after touching the assembler's
ring arithmetic** — and note the register-space decode (item 4) was measured
separately and cost nothing, so attributing a regression needs the A/B.

For the record, libogc2 writes `TLBP = 0x0000`, `BP = 0x0001`,
`RHBP = 0x000F` — exactly the old hardcoded layout, which is why this never
showed up in practice.

## Throughput: measured 850 KB/s, and it is NOT the adapter

**Measured 2026-10-10**, Swiss dumping a disc over the adapter, captured on
the PC and analysed with `scripts/analyse-pcap.py`:

```
GC -> PC  18395 frames, 19.83 MB, 850 KB/s (6.97 Mbit/s, 25.8% of the EXI ceiling)
          807 frames/s, mean frame 1078 B, retransmissions 0
PC -> GC   9200 frames, mean 54 B (pure ACKs)
GC advertised window 2920 B;  gaps >5 ms total 56 ms (0.2% of 22.78 s)
```

**The flow is window × RTT bound and nothing else:**

| derived | value |
|---|---|
| data frames per ACK | 2.0 |
| payload per round trip | 2048 B |
| implied RTT | 2.48 ms |
| 2048 B / 2.48 ms | **826 KB/s** ≈ the measured 850 |
| payload per frame | **1024 B** (wire 1078 − 54 of headers) |
| window ÷ segment | 2920 / 1024 → **2 segments in flight** |

So the sender can have 2048 bytes outstanding and each round trip costs
2.48 ms. **Zero retransmissions** means the 15-page ring never overflowed and
we drop nothing; **0.2% stall time** means nothing is waiting on the disc.
Our W5100 bus engine runs at roughly 17% utilisation at this rate. The adapter
is not the constraint.

The levers, in order of effect, and **none of the first two are ours**:

1. **A larger TCP send buffer on the GC** (lwIP `TCP_SND_BUF`). Throughput
   scales with it almost linearly while the flow is window-bound.
2. **Full-MSS segments.** Swiss writes 1024 bytes at a time, so a 2920 B
   window carries 2048 B of payload. At 1460 B segments the same window and
   RTT would give ~**1150 KB/s, a 1.43× gain** for free.
3. **A shorter round trip** — throughput scales 1/RTT. The 2.48 ms is GC-side
   interrupt dispatch and lwIP, plus EXI lock contention: libogc2's
   `bba_event_handler` bails out if `EXI_Lock` fails and defers until the TX
   path unlocks, so RX handling queues behind TX on the single channel.

⚠️ **RX interrupt coalescing would make this WORSE, and that reverses an
earlier recommendation in this file.** `NCRB.RXINTC` asks for 2 packets per
interrupt and `RXINTT` adds a timer; both were listed as throughput levers.
For a **window-bound** flow they are not: delaying the ACK's interrupt
lengthens the round trip, and throughput is bytes-per-round-trip ÷ RTT. The
measurement is what showed this — the guess was backwards.

### ⚠️ The 850 KB/s figure does NOT bound ROM streaming — different stack

The dump was Swiss's **menu** path: libogc2 + lwIP + TCP, 1024-byte
application writes, a 2920-byte window. **ROM streaming uses a completely
different code path** — Swiss's in-game patch driver and its own stack in
`cube/patches/bba/tcpip.c`, which has **no TCP in it at all**:

- **UDP + FSP** (File Service Protocol), with ICMP and ARP. No TCP, no lwIP,
  no congestion control, no window.
- Reads are sized to the path MTU, not 1024:
  `_fsp.data_length = MIN(length, env->pmtu - (fsp->data - eth->data))`
  — about 1446 B of payload per datagram at a 1500 MTU.
- `QUEUE_SIZE 2` requests outstanding.
- `udp->checksum = 0x0000` — no payload checksum, which is real CPU saved
  versus TCP.
- It runs in the game's context via the patch driver, which uses the leaner
  4-byte command-space form (see the command-space note in CLAUDE.md).

So the TCP window × RTT limit is not the operative one there, and the measured
850 KB/s says nothing about it. What the dump DID establish, and which carries
over, is that the adapter has headroom: **zero retransmissions**, ~17% W5100
bus utilisation, 26% of the EXI ceiling.

**Not yet measured.** `analyse-pcap.py` now counts UDP payload and picks the
bulk direction automatically (streaming flows PC → GC, the opposite of a
dump), so `--bulk to_gc` on a streaming capture gives the real number. Until
that is run, "can it stream?" is open — but it is open on the GC's side, not
on the adapter's.

The hard ceiling remains **3375 KB/s** (27 MHz × 1 bit/clock), and the GC's
own disc delivers roughly 2000–3100 KB/s, so there is no headroom to spare if
a game wants full disc rate — but games read in bursts, not continuously at
peak.

## 3. `NCRB` (0x01) is not decoded at all

**Status: ✅ DECODED 2026-10-10** — stored, readable, and exported as
`BBARegisterFile.ncrb`. Nothing consumes the bits yet; that is item 1.
Bits, from libogc2:

| bit | name | meaning |
|---|---|---|
| 0 | `PR` | promiscuous mode |
| 1 | `CA` | capture effect mode |
| 2 | `PM` | pass multicast |
| 3 | `PB` | pass bad frame |
| 4 | `AB` | accept broadcast |
| 6–7 | `RXINTC` | RX interrupt counter: 1 / 2 / … packets per interrupt |

We interrupt per frame where libogc2 asks for every 2 packets, which is benign
(more interrupts than requested). `PR`/`PM`/`AB` feed item 1.

⚠️ **Do NOT implement `RXINTC`/`RXINTT` coalescing as a throughput
optimisation.** Measurement showed the bulk flow is window × RTT bound (see
"Throughput" above), so delaying an interrupt lengthens the round trip and
*lowers* throughput. Implement it for fidelity if a driver depends on the
behaviour, not for speed.

`bba_register_file` **T15** asserts the export is `0x52` with `PR=0`, `PM=0`,
`AB=1` — i.e. the real BBA would deliver own-MAC + broadcast only — which is
the fact item 1 has to act on.

## 4. Registers that read `0x00` but have non-zero hardware defaults

**Status: ✅ DONE 2026-10-10** for every register libogc2 writes. They now
store writes and read back, with the documented hardware defaults at reset.
`bba_register_file` **T15** checks the defaults, replays libogc2's twelve
init writes and reads every one back; mutation-tested (dropping `NWAYC`'s
`0x84` default fails it, as does dropping `NCRB` writes).

Still reading `0x00`, deliberately: `LTPS` (0x04), `LRPS` (0x05), `TWP`
(0x0E/0F), `TRP` (0x12/13) and `TXFIFOCNT` (0x3E/3F). These are **status**
registers, not configuration — a driver reads them to learn what the device
did, so storing what was written to them would be a lie rather than a default.
Drivers read `LRPS`/`LTPS` only on error bits, and `0x00` reads as "no error",
which is the truthful answer for a device that does not detect those errors.
`TWP`/`TRP`/`TXFIFOCNT` would need the TX engine to actually report its page
pointers and FIFO depth; worth doing if a driver is ever seen to poll them.

The original problem for reference — a driver doing read-modify-write, or
sanity-checking a default, saw zero. libogc2 *does* read `NWAYC` (0x30) before
writing it.

| reg | name | hardware default |
|---|---|---|
| 0x01 | `NCRB` | — |
| 0x04 | `LTPS` | last transmitted packet status, RO |
| 0x05 | `LRPS` | last received packet status, RO |
| 0x0A/0B | `BP` | boundary page pointer |
| 0x0C/0D | `TLBP` | TX low boundary page |
| 0x0E/0F | `TWP` | TX buffer write page |
| 0x12/13 | `TRP` | TX buffer read page |
| 0x14/15 | `RXINTT` | RX interrupt timer |
| 0x1A/1B | `RHBP` | RX high boundary page |
| 0x30 | `NWAYC` | **0x84** |
| 0x32 | `GCA` | 0x00 |
| 0x3D | `MISC` | **0x3C** |
| 0x3E/3F | `TXFIFOCNT` | TX FIFO counter, RO |
| 0x50 | `MISC2` | 0x00 |
| 0x58–0x5F | SI regs | libogc2 writes `80 00 03 83 32 FE 1F 1F` |
| 0x60 | `SI_ACTRL2` | libogc2 writes `0x00` then `0x74` |

## 5. Two of our registers do not exist on real hardware

**Status: KNOWN, low risk, do not "fix" blindly.**

`0x34`/`0x35` (`TWD_LO`/`TWD_HI`, our TX length latch) and `0x3A` (`HIPR`,
returning `0x01`) are **not in the official register map** — `0x3A` is
unassigned there and `MISC` is `0x3D`. They are our own inventions, and it is
now confirmed that **no driver writes TWD** (the TX length is counted at
TXDATA instead — see CLAUDE.md). They are harmless unless a game writes those
addresses for some other purpose and trips our TX-length latch.

Removing them would break `gc_bba_test`, which does write TWD, so keep them
but do not add more inventions.

## 6. Lower risk / no action

- **Challenge/response** (command regs 0x08/0x09/0x0B) is **device-initiated**
  — gated on command status bit `0x10`, which we never raise, so the path
  cannot be entered. Do NOT implement `bba_calc_response` emulation.
- **No EEPROM**: the MAC is a fixed reset value in PAR0–5. Since libogc2 reads
  the MAC from there rather than writing it, that value is what the console
  uses. Games keying anything off the MAC see ours.
- **`RXINTT`** (RX interrupt timer) is interrupt coalescing; ignoring it means
  we interrupt earlier than asked, which is safe.
- **Command register `0x0F`** may be a reset strobe rather than a register:
  `bba_cmd_in8_slow(0x0F)` holds CS for 200 µs between two 10 ms delays and
  discards the result. Implemented as storage; the NCRA reset follows
  immediately in that sequence so ignoring any side effect costs nothing
  today.