FireWire S800 for the Power Mac G4 MDD (FW800) under Mac OS 9

Restores all three FireWire ports on a Power Mac G4 MDD FW800 (PowerMac3,6) running Mac OS 9.2.2: the two FW400 ports at S400, and the FW800 port at S800, at the same time. One patched extension, FireWire Enabler. Nothing else is modified.

So the gain over stock is that the two FW400 ports become usable. The gain over the previously released fix is keeping S800 on the FW800 port instead of trading it away for the FW400 ports. Measurements in docs/QUICKBENCH-S800-vs-S400.md.

Mac OS 9's FireWire family builds the IEEE 1394 speed map from the sp field of each node's self-ID packet. sp is two bits, and the value 3 is reserved in 1394a. 1394b uses it to mean "S400 or better, ask the PHY". It does not mean S800.

Apple's own IOFireWireController::buildTopology says exactly that:

speedCode = (id0 & kFWSelfID0SP) >> kFWSelfID0SPPhase;
if( speedCode == kFWSpeedReserved )
 speedCode = kFWSpeed800MBit | kFWSpeedUnknownMask; // we don't know how fast it is

and then verifies by transfer: on a failed config-ROM read it calls setNodeSpeed(node, local, current - 1) and retries, down to S100. TI's TSB81BA3 errata recommends the same try-and-see approach.

Mac OS 9's family has no step-down. It stores sp verbatim and treats 3 as S800. So two 1394b nodes joined by a legacy (1394a data/strobe) cable both claim S800 on a hop that physically carries S400, the transmit is never received, and the device simply never appears. That is the dead-FW400-ports bug.

OS 9 cannot try-and-see without a rewrite, but the FWIM owns the PHY, and the PHY knows. So the FWIM (FireWire Enabler, the OHCIFWIM container) is hooked immediately before it hands the self-IDs up to FWProcessSelfIDs. For each of its own ports it reads page 0 registers 8, 9 and 11 (Connected, Negotiated_speed, Beta_mode), maps each node to the port it arrives on, and clamps that node's self-ID sp to what its own hop can actually carry. The family's existing min-propagation then produces a correct map, with nothing else in the system altered.

The local node's own sp is deliberately left at S800. That is what allows one speed map to give S800 to a beta neighbour and S400 to a legacy one simultaneously.

Fail-safe throughout. PHY register 2 is a liveness oracle, and an unreadable PHY, an unexpected port count, or any topology that is not a plain star centred on this Mac all fall back to a conservative S400 rather than risking the original symptom.

Measured on the target machine, Mac OS 9.2.2, one LaCie FW800 drive and a clamshell iBook (FireWire) in Target Disk Mode.

Run 6 is the one that matters most. Both devices are 1394b and both report sp = 3, so it is the configuration stock cannot handle at all, and the only one where a wrong port-to-node mapping fails visibly rather than silently. Log in logs/FWFixCheck_G5_TDM_v006.log.

Throughput, measured on the same drive and controller with only one variable changed at a time. Patched against stock on the beta port is +1.5% read, which is to say nothing: stock was already S800 there. S800 against S400 as hop speeds is +10.6% read and +7.5% write, which is what the earlier global clamp gave away and this patch keeps. A legacy device merely being present on the bus costs the beta device a further 8.9%. Details and the full tables in docs/QUICKBENCH-S800-vs-S400.md.

FireWire Enabler in the System Folder's Extensions folder is the only file that changes. Keep the original.

Expand artifacts/FireWire S800 Enabler.bin. It produces a file already named FireWire Enabler, so it drops straight in. It reads 2.8.8 in Extensions Manager against a stock 2.8.7, which is how you tell at a glance which one is installed.

FWFixCheck is a read-only diagnostic. It scans the System heap for the patch's counter block and reports what it decided: per-port state, beta flag and negotiated speed, and each node's own sp against the ceiling applied to it.

• Only a plain star centred on this Mac gets per-connection speeds. Behind a hub or a daisy chain, everything is capped at S400. This is correct rather than conservative: the PHY can only measure the hops it terminates, a legacy hop deeper in the tree is invisible in the self-ID stream, and assuming S800 there would reproduce the original defect. See the warning in tools/patch-firewire-enabler.py.
• Sleep and wake are broken on this machine independently of FireWire, confirmed with no FireWire device ever attached. Not addressed here.

python3 tools/patch-firewire-enabler.py "FireWire Enabler" out/
python3 tools/wrap-macbinary.py "out/ascend/FireWire Enabler" "FireWire S800 Enabler.bin"

Give it your machine's own FireWire Enabler, transferred with its resource fork intact. The patcher refuses to touch a file whose vers is not 2.8.7, so it cannot silently patch a build it was not derived from.

What it does: grows the OHCIFWIM container's code section, appends the hook, redirects the two bl instructions that call FWProcessSelfIDs, updates the container length in the cfrg resource, and stamps vers (1) to 2.8.8. Two bytes change inside the original code, everything else is appended, and the data and loader sections come through byte-identical.

Reverse engineering, the patch, the diagnostics and the documentation were produced with substantial help from Claude (Anthropic), working from Apple's published IOFireWireFamily source, the TI TSB81BA3 datasheet, and disassembly of the shipping OS 9 binaries. All hardware testing was run on a real Power Mac G4 MDD FW800.