[bnunu]

Summary

The native builds (Windows, Linux, Android) now host system link games of up to 128 players on up to 128 machines. The Xbox game allows 16 players on up to 4. Split screen stays at 4 players per machine.

• The limits are two constants in port/linux/include/halo_port_limits.h. The memory they need is set in port/linux/include/halo_port_capacity.h.
• Every change is under #ifdef HALO_LINUX, so the byte-matching MSVC build keeps the Xbox limits. All 621 objects compare identical to main's.
• The protocol version is now 2, so these builds neither see nor are seen by the Xbox game or older native builds. Every machine in a game needs the same build: the game runs in lockstep, and pool sizes affect the simulation.

This PR is rebased onto the current main (13d1ae2) and no longer depends on #1. That PR clamped colours that the x87 maths pushed a hair past 1. Now that 13d1ae2 has put Windows back on SSE maths, those asserts don't fire in a 128-machine game on Blood Gulch, so #1 is closed.

What changed

• Networking
  • All limits follow halo_port_limits.h.
  • The 13 KB game settings record is sent in four pieces and reassembled, since a message header holds at most 4095 bytes. The per-tick update of 128 players (3,857 bytes) fits in one message.
  • Player counts are bytes.
  • Each machine's input carries 4 players' actions.
  • Buffers, queues and FD_SETSIZE are larger.
  • Stream writes send a whole message or drop a peer that stopped reading, instead of corrupting the stream.
  • Lobby changes go out at most every 250 ms.
  • In-game joins wait in a queue.
  • The random name and colour pickers terminate.
• Host behaviour at scale (found with 128 machines)
  • The listen backlog is 128, not 32.
  • Socket buffers are at least 1 MB, enough for the updates a machine hasn't acknowledged yet.
  • Machines that finished loading get a keep-alive every 5 s while slower machines load (the host waits up to 60 s).
  • Logging no longer grows with the square of the session, and debug.txt stays open between lines.
• Game engine
  • Per-player arrays hold 128 players.
  • Places past 16th get ordinals (17th, 21st, ...).
  • Free-for-all teams are the player's index.
  • Team masks are sized and checked.
• Interface
  • The lobby's three remote panels show the first three remote machines.
  • The HUD teammate buffer is sized for 128 players.
  • The motion sensor checks range before a unit takes a blip slot.
• Memory
  • The game state is 16 MB at 0x81A00000 (3.3 MB on the Xbox).
  • Object, effect, particle, contrail, light, sound, cluster-reference and render-state pools are larger.
• Platform (port/linux/src/xnet.c)
  • HALO_NET_ADDRESS gives each copy of the game its own address, so several copies can play together on one computer.
  • HALO_NET_BROADCAST sends the game search to a list of addresses.
• tools/system_link_bots.py
  • Stand-in machines that join a host, acknowledge every tick and send input, for testing large sessions without 128 copies of the game.
• Docs
  • A "System link" section in port/linux/README.md, and new rows in the settings and HALO_LINUX edits tables.

Testing

All on one Windows computer, with the Windows build of this branch on the current main (13d1ae2):

• 128 machines, 128 players, Blood Gulch Slayer. A host (HALO_NET_ADDRESS=127.0.0.200), a second copy of the game as a real client (127.0.0.201), and 126 stand-ins played for about 90 seconds.
  • The update stream held a steady 30 updates per second, and every machine acknowledged every tick.
  • The real client simulated all 128 players for about 2,600 ticks, 0 ticks behind the host, with no out-of-sync events (the per-tick random seed check passed).
  • No assertion fired on either copy.
  • The host uploaded about 118 Mbit/s.
• Earlier runs on the previous main (a06ef74).
  • The same session ran for about 3,200 ticks.
  • A host with 127 stand-ins handled all of them leaving at once within a second.
  • With 8 stand-ins taking 25 s to load, the host's own client stayed connected past the 15 s silence timeout.
• Byte matching. An object-by-object comparison of all 621 MSVC objects against the current main (sections, relocations, symbols) found 0 differences.

Not covered:

• The Linux and Android builds were not built: this machine has no Linux sysroot or Android NDK. The changes are shared C under HALO_LINUX plus the shared xnet.c. A Linux or Android build check would be welcome.
• The stand-ins don't play. A game with many real players, for example spawn points on maps made for 16, hasn't been played.

🤖 Generated with Claude Code