embedded-forge

A hands-on bare-metal embedded systems study track in C and ARM assembly, targeting Cortex-M3 on QEMU. Five parts, each building on the last: a libc reimplementation, bare-metal startup, an interrupt-driven driver, a bootloader, and a context-switching scheduler.

Update: A follow-up project focussed on custom RTOS has been built: https://github.com/Athyo30/Forge-RTOS/tree/main

Structure

part1_libc/          libc reimplementation + a bump and free-list allocator
                      (host build, no hardware/emulator needed)
part2_startup/        bare-metal reset handler: .data copy, .bss zero, vector table
part3_drivers/        interrupt-driven UART RX with a lock-free ring buffer
part4_bootloader/     vector table relocation + jump to a second application image
part5_concurrency/     cooperative round-robin scheduler with a hand-rolled context switch
setup.sh              installs the toolchain
DEBUGGING_JOURNAL.md  full bug-by-bug write-up

What I did vs. what was scaffolded

I implemented every function in every part — the allocators, the UART driver, the bootloader's validation and VTOR relocation logic, the scheduler's switching policy — and debugged the real bugs that came up along the way (a segfault from unsigned integer underflow, a linked-list corruption bug that passed its own test suite, a race condition between an ISR and main, among others).

Build tooling was AI-scaffolded so study time went into the concepts rather than environment setup: linker scripts, Makefiles, the test harness, and a few pieces of ARM assembly boilerplate too fiddly to be a good place to learn by trial and error (the naked jump-to-application function in Part 4, the register save/restore skeleton in Part 5). Full write-up of every bug, with root cause and fix, is in DEBUGGING_JOURNAL.md.

A preemptive scheduling extension (SysTick + PendSV, forcing a context switch without an explicit yield()) was attempted and deliberately shelved rather than shipped half-working — noted at the bottom of the journal.

Key learnings

• void* used to hold a byte count or an integer is a type-category error, not a style nit — it showed up three separate times (an allocator, a stack offset, a peripheral register write) before the pattern was obvious.
• Unsigned loop counters can't go negative — a backward loop bounded by i >= 0 on a size_t never terminates; it underflows and wraps instead, which is a segfault waiting to happen, not a warning to ignore.
• Passing tests only prove what they exercise. A free-list split bug that truncated the block list was invisible for several test runs because every allocation happened to land on the list's tail.
• Lock-free doesn't mean uncoordinated — a single-producer/single- consumer ring buffer is safe without a mutex only because each shared field has exactly one writer; a shared counter updated from both an ISR and main breaks that invariant immediately.
• A bootloader is a program that fakes a hardware reset — jumping to an application means setting SP/PC and relocating SCB->VTOR, two independent mechanisms that are easy to only do one of.
• AAPCS, not compiler magic, is what lets C and hand-written assembly cooperate — arguments in r0/r1, callee-saved r4-r11 — the same convention underwrites the bootloader's asm jump and the scheduler's context switch.

Toolchain

• gcc / make - host build for Part 1 (no target hardware involved)
• arm-none-eabi-gcc - cross-compiler for Cortex-M3, Parts 2-5
• qemu-system-arm - emulates the lm3s6965evb board (no physical hardware needed for any part)

Tested on WSL2 (Ubuntu) and native Linux. Run ./setup.sh once, then cd partN_* && make test (Part 1) or make qemu (Parts 2-5).