Jsdos

Experimental AMD64 OS.

Summary: Load nearly all kernel and driver code from bitcode files and JIT them into memory for native machine execution. Bitcode loader will verify memory safety and abort loading if proved unsafe.

Build:

• cd jsdos/
• ./compile.sh && ./run.sh

Requirements:

• Linux x86_64 OS to build on
• GCC with C99 support
• QEMU with AMD64 emulation support
• Your linux user account must be able to sudo
• Your linux user account must be added to the disk group to use losetup

Included:

• MyJIT ( http://myjit.sourceforge.net/ )

Current status:

• Woefully incomplete
• MyJIT library included into binary KERNEL64.SYS
• Only required stdlib.h functions are implemented; just scratch code for now
• No IR bitcode format defined yet
• No drivers implemented

Roadmap:

• Runtime Options:
  • Possibly make use of LLVM's bitcode with its SSA form
  • Each SSA value maps directly to a MyJIT virtual register
  • MyJIT will use its register mapping system to map virtual registers onto machine registers
  • JIT considerations:
    • Need I/O opcodes (IN AL, DX, OUT AL, DX, etc.)
  • Need verifiable safety system
    • Define multiple levels of code access
      • Level 0 is kernel level, including H/W access opcodes
      • Level 1 is driver level, interface directly with kernel to access H/W via communication queues
      • Level 2 is driver abstraction level, providing abstracted interfaces for driver communication
      • Level 3 is application level, no H/W access
      • Each level may only communicate with its direct parent level
    • All OS functions passed via interfaces
      • Declarative security model
      • Any interface's function may be nil-ed out if permission is denied and/or not declared
  • Functions are expressed as coroutines (maybe?)
  • No heavy threads.
  • No shared state whatsoever.
  • No blocking OS calls; all async I/O.
  • Each CPU core runs an async event loop to invoke/continue function execution.
  • Best-case cooperatively scheduled drivers/applications where each yields at defined points, worst-case time-sliced scheduling.
• Language Options:
  • Requirements:
    • Need memory safety (no pointer arithmetic and no arbitrary pointer values)
    • All system interaction should go through interfaces (actual vtables)
    • Cross-core communication primitives (e.g. queues)
  • Seriously considering using Google's Go language
    • Its interface system looks clean and simple
    • Might have to add type-safe options for memory-mapped hardware I/O
    • GC is essential