Files

T

CydandClaude Fable 5 b01db52d43 BT410 5.3.117: the pod talks -- the entire NetNub networked layer reconstructed, and the last runtime stub falls

Twenty-two Fail sites in L4NET.CPP -- the whole console/mesh transport --
replaced with decomp-verified bodies (part_007.c @004501d4..@00452378),
every error string extracted byte-for-byte from the shipped exe, every
magic number resolved to a macro from the AUTHENTIC 1995 NETNUB.HPP that
survives in the archive (NETCOM_VERSION 11; 0x63c == MAX_SEND_DATA_SIZE).

The real-mode bridge came free: the 1995 netshare.OBJ has been linking
into mungal4.lib all along via the prebuilt fallback -- this file only had
to define the six globals it imports and call it. NetNub::SendCommand is
the gateway (setRMBuff, DPMI int 31h/0300h on the TSR's vector, getRMBuff);
BC4.52's <dos.h> really has int386x -- the binary's own "int386x():ERROR="
text named the RTL call.

The topology, decomp-proven: console listens on 1501, the pod mesh on
1502; hosts before ours in the egg get an OPEN, ours becomes the local
host by address match, hosts after ours get a LISTEN -- and CheckBuffers
heals the ordering race by SWAPPING sockets when a peer lands on the wrong
listen. The console's listen is a wildcard that learns its address; the
console connection SURVIVES mission teardown (Shutdown zeroes both
counters anyway -- reproduced faithfully). Send discipline: direct while
the retry buffer is empty, everything queues behind a backlog in per-host
FIFO, one round-robin drain per idle slot; only a dumped-to-avoid-blocking
send is retried, and RemovePacket is empty in the binary because delivery
already compacted the pad buffer. Internal connect/disconnect events ride
the GetNetworkClientPointer(0)->ReceiveNetworkPacket self-dispatch (the
decomp's FUN_0041d7f0 resolves to the ClientID-0-is-me idiom).

Ghidra had dropped register and trailing arguments at half a dozen call
sites (GetRemoteHost's HostID among them) -- the surviving HPP signatures
arbitrated each.

Stubs: 8 across 8 files -- ALL RUNTIME STUBS ARE NOW CLEAR; what remains
is the authoring-side CreateStreamedSubsystem family, which missions never
call. Single-user regression clean on the rig; the networked path awaits
its console + NetNub-TSR live session (two-agreeing-runs).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

2026-08-04 10:05:13 -05:00

5.0 KiB

Raw Blame History

L4NET.CPP — the NetNub networked layer (reconstructed 2026-08-04, 5.3.117)

The pod's entire console/mesh transport: the protected-mode client of the real-mode NetNub TSR. Every body verified against the shipped binary (part_007.c cluster @004501d4..@00452378); every error string extracted byte-for-byte from BTL4OPT.EXE; the authentic interface header (CODE/RP/MUNGA_L4/NetNub/NETNUB.HPP, NETCOM_VERSION 11) supplied every struct and function code, and each binary literal resolved to one of its macros (0x63c == MAX_SEND_DATA_SIZE, 0x2c == MULTIPLE_SEND_BUFFER_SIZE(0), 0x640 == MAX_RECEIVE_DATA_SIZE, 10 == the frame's destination limit).

The real-mode bridge

NetNub::SendCommand (@00452100) is the gateway: setRMBuff() pushes the PM Netcom's Function/Status/Buffer_Length + buffer down to the real-mode shared block, DPMI int 31h/0300h fires the TSR's interrupt (vector from getRMNumbers()), getRMBuff() pulls the reply back. The bridge routines are NOT recompiled — the SURVIVING 1995 netshare.OBJ (CODE/RP/MUNGA_L4) has been a mungal4.lib member all along via the prebuilt-fallback (netshare in ORDER_mungal4); this file only had to supply the C-side globals it imports (Function_Ptr, Buffer_Length_Ptr, baseadr:offs, RMFunction_Off, RMBuffer_Length_Off — binary homes 0x522b84..0x522b98) and extern "C" prototypes (TASM uppercased the OBJ's symbols; tlink32 matches caselessly — no /c). BC4.52's 32-bit <dos.h> really does have int386x/segread — the binary's own error text ("int386x():ERROR=") named the RTL call. The register image zeroes exactly the fields the binary zeroes (reserved/flags/fs/gs/ss:sp).

The topology (all decomp-proven)

Console listens on TCP 1501, pod mesh on TCP 1502.
StartConnecting: hosts BEFORE ours in the egg list get an active OPEN (they already listen); our entry is recognized by address and adopted as the local host; hosts AFTER ours get a LISTEN. That ordering rule forms the peer-to-peer mesh without collisions. A slave then Sends AcknowledgeEggFile to the console; MasterMode Fails ("host is in an illegal startup mode" — 4.10 pods are always slaves).
CheckBuffers (@004518b8) is the pump: deliver a whole buffered packet; else advance ONE connection state machine per call (OPEN must match its address — Verify; the console LISTEN is a wildcard that learns its address; a mesh peer landing on the WRONG listen gets its socket SWAPPED to the record that expects it — the race healer), then one MULTIPLE_RECEIVE drains the wire into per-host pad buffers (pad_size/pad_head/pad_tail — the surviving L4HOST.HPP names them).
Wire frame = NetworkPacketHeader{clientID, gameID, fromHost, timeStamp}
- message; the netnub strips the leading socket handle. GetNextMungaPacket round-robins hosts from lastHostIteratorPosition, copies out one whole packet, memmove-compacts the pad. RemovePacket is EMPTY in the binary (@004520f8) — delivery already dequeued.
Send discipline: buffer empty → direct send, and only a NETNUB_DATA_DUMPED (would-block) result queues the message; buffer non-empty → everything queues BEHIND (per-host FIFO) + one drain attempt. ExecuteBackground drains one request per idle slot. The MessageBuffer copies each message (SendRequest owns the copy), tables queues by host, and round-robins them. A disconnect during send posts the internal HostDisconnected event through GetNetworkClientPointer(0)-> ReceiveNetworkPacket(NULL, &msg) — the self-dispatch idiom the decomp's FUN_0041d7f0(this,0) resolves to (ClientID 0 == the manager).
Shutdown: game hosts close + destroy; THE CONSOLE CONNECTION SURVIVES the mission (pod returns to the console-only lobby). The binary zeroes BOTH connection counters even though the console stays online — reproduced faithfully (the console's own disconnect path recreates its listener rather than decrementing through zero).
Mode: reliable == the netnub BLOCKS in send (SET_SWITCHES blocking_switch=1); unreliable == non-blocking, dumps join the retry buffer. The ctor forces ReliableMode; the engine flips to Unreliable during map load.

Sources & drift notes

RP411's L4NET.CPP is the 2007 WinSock port — its 1995 NetNub code survives only in comment blocks (used line-for-line for the ctor RM plumbing). CODE/RP's L4NET.HPP + NETNUB.HPP + NETSHARE.ASM are authentic-form.
Ghidra dropped register/trailing args at several sites (GetRemoteHost's HostID, SendMessageToNetnub's host argument, CreateWeaponExplosions-style stack args) — the HPP signatures and HostManager's Find-based GetRemoteHost settle each.

Runtime status

Single-user path regression-clean on the rig (mission soak, no faults). The NETWORKED path compiles + links (0 unresolved) but awaits its live test: a console + NetNub-TSR session (net_loop-style conf) driving OUR BTL4REC — the console handshake (egg → Ack → connections → LoadMission), a two-pod mesh, and mission teardown back to the lobby. Two-agreeing-runs rule applies.

5.0 KiB Raw Blame History