BoredOS logo BoredOS

Process Management & Scheduling

BoredOS implements a lightweight, symmetric multiprocessing (SMP) capable multitasking environment. This document outlines the architecture of the scheduler, process structures, context switching, and ELF binary loading.

1. Process Structure (process_t)

The core of the process management system is the process_t structure, defined in src/sys/process.h. Due to kernel memory constraints, BoredOS supports a maximum of 16 concurrent processes (MAX_PROCESSES), stored in a statically allocated array.

Key fields include:

  • Identification: pid, parent_pid, pgid (Process Group ID), and name.
  • Memory & Context:
    • rsp: The saved stack pointer during a context switch.
    • pml4_phys: The physical address of the Page Map Level 4 table (VMM root) for this process.
    • kernel_stack & user_stack_alloc: Pointers to allocated stack memory.
  • Scheduler State: ticks, sleep_until, is_idle, cpu_affinity.
  • Resources:
    • fds: File descriptor table tracking open files, pipes, and sockets (up to MAX_PROCESS_FDS = 16).
    • gui_events: A circular queue for Window Manager events (keyboard, mouse).
  • Signals: POSIX-like signal tracking via signal_mask and signal_pending.

2. The Scheduler

BoredOS uses a Preemptive Round-Robin scheduler implemented as a circular linked list.

Symmetric Multiprocessing (SMP)

Each CPU core maintains its own current_process pointer (current_process[my_cpu]). When a new user process is spawned via process_create_elf, the kernel assigns it to an Application Processor (AP) core using a simple round-robin assignment policy (next_cpu_assign), avoiding Core 0 (BSP) which is typically reserved for kernel tasks and driver interrupts.

The process_schedule Loop

When the timer interrupt fires, it calls process_schedule(current_rsp):

  1. It saves the current_rsp into the current process's structure.
  2. It handles cleanup of killed processes (kill_pending).
  3. It traverses the circular linked list (cur->next) looking for a process where cpu_affinity == my_cpu.
  4. It checks if the process is sleeping (sleep_until > now).
  5. It switches the hardware context:
    • Updates the Task State Segment (TSS) ring 0 stack pointer.
    • Switches the page directory by writing the new pml4_phys to CR3.
    • Returns the new process's rsp, which the interrupt handler then pops into registers.

3. Context Switching

Context switching is achieved by manually constructing an interrupt stack frame (IRETQ frame).

When a process is created, the kernel sets up the top of its kernel stack with:

  • SS (Stack Segment: 0x1B for user, 0x10 for kernel)
  • RSP (The process's stack pointer)
  • RFLAGS (0x202 to ensure interrupts are enabled)
  • CS (Code Segment: 0x23 for user, 0x08 for kernel)
  • RIP (The entry point of the binary or function)
  • Zeroed space for General Purpose Registers and a 512-byte fxsave region for FPU/SSE state.

When process_schedule returns the new rsp, the assembly interrupt stub uses pop instructions to restore the general-purpose registers, and finally executes iretq, transitioning execution to the new process seamlessly.

4. ELF Loading

Userland applications in BoredOS are standard 64-bit ELF binaries.

The function process_create_elf orchestrates this:

  1. Memory Allocation: Creates a new PML4 page table for the user process.
  2. Parsing: Calls elf_load(filepath, pml4, &size) to parse the ELF headers, allocate required physical memory, and copy the executable segments (text, data, bss) into the process's virtual address space at the locations specified by the ELF program headers.
  3. Stack Setup: Allocates a 256KB user stack mapped at 0x800000.
  4. Argument Passing: Parses the args_str passed to the executable and pushes an argv array onto the newly allocated user stack.
  5. Execution: Sets the stack frame's RIP to the ELF entry point and links the process into the scheduler's run queue.

5. Process Termination

When a process exits (or is killed), it is not immediately freed. The scheduler sets kill_pending = true. The actual destruction of the PML4 table and stack allocations is deferred to the next tick inside process_schedule to avoid freeing the memory of the code currently executing the cleanup.