Skip to content

mlugg/gapxos

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

60 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

gapxos

gapxos (Generic All-Purpose eXtensible Operating System) is an operating system built for the 64-bit amd64 architecture. It is a microkernel design, where most or all drivers are run as userspace applications.

gapxos is compliant with GNU Multiboot 2, and the iso target in this repository will generate a bootable ISO image using GRUB 2. If you wish to use another loader, it is important that the kernel file is loaded as a module.

gapxos is NOT meant to be a POSIX-compliant operating system. It is also not compatible with any legacy technologies such as APM and 32-bit processors.

The system is currently in a very early state, and as such cannot be used for most applications. Here are some important features on the roadmap:

  • Interrupt handling
  • System calls
  • Filesystem structure
  • Task scheduling

System Outline

gapxos boots using a Multiboot 2 compliant bootloader, such as GNU GRUB. As Multiboot only supports running 32-bit images in x86 protected mode, it boots a small x86 "loader" and loads the actual 64-bit kernel in a boot module from an ELF file. The loader sets up a 4k stack and simple GDT. It then traverses the Multiboot structure to find the loaded kernel, and parses the ELF file, setting up the appropriate page tables. It also identity maps the first 8MiB of memory, so the loader can keep executing when it later switches to long mode - this is NOT CERTAIN to work, as Multiboot makes no guarantees about the position in memory of the loaded image, and should be changed in future. The kernel is mapped in the higher half of virtual memory, in address 0xffff800000000000 and above. After it finishes parsing the ELF file, the loader prepares to jump to 64-bit code. It enables PAE, and switches to IA32e compatibility mode, with paging enabled. It loads a small x64 GDT, and jumps to x64 code. It sets several registers to act as parameters for the kernel_main function, and jumps to the loaded kernel code.

The kernel_main function take several arguments. The first is the location in memory of the Multiboot structure, and the rest describe the location of reserved areas of memory which should not be marked as available: specifically, the stack, and the kernel code itself. Firstly, this function recreates the GDT; it is essentially identical to the one created by the loader, but must be recreated as the existing one could be overwritten later. The next job of the kernel is to parse the Multiboot structure and get any useful information from it. Every element of the structure is traversed, and 4 types of tags are actually read: the memory map, any present framebuffers, and the RSDP for either version of ACPI. In future, this will likely also read a second boot module to use as the initramfs. After reading the entire multiboot structure and e.g. validating ACPI checksums, the system begins to initialize the physical memory manager. The PMM is implemented very simply as a stack of free 4k memory pages, so the system first reads the memory map to identify how much memory will be needed to store the stack itself. Once this is calculated, it traverses the map again to find a range of memory which is of sufficient size to store this list and is not already taken up by the program stack, the kernel code, the page tables or the memory map itself, which is the only part of the Multiboot structure still required by the system at this point. After the stack of free memory pages is created and populated, the system finishes its inital boot preparations and runs the system_main function.

The system_main function takes a single argument, which is a structure in memory containing information such as the ACPI RSDT. This function starts by setting up a virtual memory manager for the kernel memory. The virtual memory manager is implemented using a self-balancing binary tree (AVL tree) to store free and allocated ranges of memory. This virtual memory manager begins at a virtual address which is a safe range above the kernel code, such as 0xffff810000000000, and runs until a long range later, but not to the end of memory so as to keep the stack uninterrupted (the precise numbers do not really have an effect as long as there is a sufficient range of virtual memory). It next begins to read ACPI tables.

About

A hacky hobby AMD64 OS.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published