Repository navigation
Releases
YAK 1.1.0
1.1.0 (2026-11-06)
The course release: chapters 1–20 of the course – parts I to V and the start of part VI – each a complete project
in course/NN-slug/ with its own make test, all of them built and tested by CI. The reference kernel is the one of
1.0.0, with two test fixes. I decided on 2026-11-06 to stop the course at chapter 20 for now (docs/ROADMAP.md).
Course
- Chapter 20 – Ring 3 and the first system call, the start of part VI: user segments, RSP0 set per thread at every
switch, USER pages,iretqinto ring 3,int 0x80withwrite(addresses checked) andexit, and faults in
ring 3 that end the program instead of the kernel. - Chapter 19 – A block cache, and the disk's interrupt, the end of part V: a write-through sector cache (394 disk
commands for a 100 000-byte file become 196, then 0) and MSI from the SATA controller, so disk commands sleep in
wait_for. - Chapter 18 – Writing FAT32: new files, appended data, directories and removal in a crash-tolerant order;
check-disk.shhands the disk image tofsck.fatand mtools after every test run. - Chapter 17 – Reading FAT32: boot sector, cluster chains (a chain in a circle is an error, not a hang), long names
and the lower-case flags,lsandcat; the test disks are made bymkfs.fatand mtools. - Chapter 16 – PCI and a SATA disk, the start of part V: PCI enumeration, an AHCI driver that reads sectors by DMA,
diskandreadin the shell. - Chapter 15 – One scheduler for all CPUs, the end of part IV: a timer and an idle thread per CPU, the scheduler lock
held through the switch, waiting that cannot lose a wake-up sent from another CPU. Bugs #29 (a timer test window
too narrow for a busy host) and #30 (joinreturned while the ended thread was still on its CPU) found on the way. - Chapter 14 – The other CPUs: a real-mode trampoline below 1 MiB, INIT/STARTUP IPIs, a GDT, TSS and stack per
CPU, andrun_onfor jobs on other CPUs; QEMU runs with 4 CPUs. Bug #28 in chapter 13 fixed on the way (the
boot hung about once in 20 intask::init). - Chapter 13 – Threads and a scheduler: kernel threads with guarded stacks, a context switch in assembly,
round-robin preemption by the local APIC timer, and blocking waits (sleep, keyboard,join) without lost wake-ups;
the heap now locks with interrupts off, since the timer handler allocates. - Chapter 12 – ACPI and the APIC, the start of part IV: the MADT from the ACPI tables, the I/O APIC routing the
timer and the keyboard, and the local APIC timer measured against the PIT; QEMU runs with a local APIC again.
Bug #27 found and fixed on the way (the PIT's square wave ticked twice per period through the I/O APIC). - Chapter 11 – A first shell, part III: a line editor, a splitter for words with quotes, and a table of six
commands;make testtypes four commands with QEMU'ssendkeyand checks the answers. - Chapter 10 – The heap, the end of part II: a free-list allocator behind
#[global_allocator]that grows by
mapping pages with chapter 9'smap;Box,VecandStringwork in the kernel, and a request over the
limit (fault=heap) ends in a panic. - Chapter 9 – Paging: the live page tables read and changed (
translate,map,unmap), and kernel stacks with
a guard page – a real stack overflow ends in the double fault handler. - Chapter 8 – Physical memory, the start of part II: Limine's memory map, a bitmap frame allocator over the
usable RAM, and physical addresses reached through the higher-half direct map. - Chapter 7 – The keyboard, the end of part I: IRQ 1, a ring buffer between handler and kernel, a scancode
decoder with Shift and Caps Lock, and a kernel that echoes what you type.make testtypes on the emulated
keyboard with QEMU'ssendkey. - Chapter 6 – Hardware interrupts and the timer: the 8259 PIC moved to vectors 32–47, a 100 Hz PIT timer, and
sleep_msthat waits inhltand never ends early. From this chapter on QEMU runs with a CPU without local APIC
(bug #25: under QEMU 8.2 no PIC interrupt arrives otherwise). - Chapter 5 – Interrupts and exceptions: an IDT with handlers for six CPU exceptions, a double fault that survives a
broken stack on IST1, and faults on purpose via the kernel command line (KERNEL_CMDLINE=fault=page make run). - Chapters 1–4 checked as a first reader would use them: the pages now show every line a learner has to change
(including the calls inkmain), name thediffcommand against the previous chapter, and QEMU's own messages in
make testgo tobuild/qemu.log(printed when a check fails). - Issue forms "I got stuck in a course chapter" and "Feedback on the course";
CONTRIBUTING.md. scripts/check-course.pyalso checks the course table, the issue form's chapter list and thediffline of each page.
Infrastructure
- The CI course job builds and tests every chapter (20 now, about 6½ minutes) and installs dosfstools and mtools
for the FAT32 test disks of chapters 17–20.
Tests
paging::translateno longer reads the PAT bit of a 2 MiB or 1 GiB page as an address bit (bug #26; no caller
was affected, all of them translate 4 KiB pages).- The userspace integration test no longer fails when a 100 ms sleep wakes up two ticks late on a busy host (bug #24).
YAK 1.0.0
1.0.0 (2026-10-27)
The first stable release: a reliable, text-based x86-64 system for QEMU (docs/GOALS.md). Stages 9 to 12 (docs/ROADMAP.md, phases A–C; docs/PROGRESS.md).
Shell and console
/bin/sh: command substitution$(...),"$@"/"$*",case … esac,localvariables in functions; an
assignment the environment cannot hold (> 4096 bytes) is an error instead of being lost.- Input lines up to 4096 characters, edited over several screen rows (was: 120 characters); the console understands
the cursor movement sequencesESC[nA/B/C/D.
Kernel
- Thread stacks have an unmapped guard below them: a thread's stack overflow kills the process instead of overwriting
the heap (thread_spawnwith stack 0 = kernel-chosen slot). - UDP checksums are verified on receipt.
Release and tests
- Release files: a bootable ISO, the disk image,
run-yak.sh– YAK runs with nothing but QEMU (README, "Try it
without building"); attached to every GitHub release by.github/workflows/release.yml; checklist
docs/RELEASING.md. - README "Supported environment" with measured RAM limits (BIOS from 24 MiB, UEFI from 160 MiB).
- Fault tests: a disk that fails at every single I/O of seven file operations (FAT32 stays mountable, only
fsck-repairable damage); TCP with lost packets, a silent peer and connect timeouts (kernel/tests/netfault.rs);
a misbehaving DNS server.
Name
- The project is now called YAK – Yet Another Kernel (formerly KernelTry): boot banner,
uname, disk label,
docs and website.
Infrastructure
- CI:
make checkon fresh Ubuntu and Fedora machines for every push and pull request
(.github/workflows/check.yml); the website is built there too. - Website: the documentation as a book (mdBook over
docs/,make book), deployed to GitHub Pages frommain. docs/ROADMAP.md: the plan to turn YAK into a course that rebuilds the kernel chapter by chapter.
Fixed (see docs/BUGS.md)
- #21: a program's heap could grow into its stack area.
- #20:
docs/DESIGN.mdsaid there was no block cache (found by an outside review). - #19: a new process could run its first syscall before its process table entry existed (SMP race in
spawn_with);
the user-space panic handler needed the heap and recursed when an allocation failed. - #18:
sleep_mscould end up to one tick early. - #17: the TLB shootdown regression test's time limit was too strict for CI machines.
KernelTry 0.8.0 (pre-release)
First public pre-release of KernelTry, a hobby x86-64 kernel in Rust for learning how an operating system works. Runs in QEMU (BIOS and UEFI). Not 1.0 yet; see "Missing for 1.0" in docs/PROGRESS.md.
Highlights
- Start with README.md and the four walkthroughs (docs/walkthrough/), then docs/EXPERIMENTS.md and docs/HACKING.md
make deps,make check,make stress; the Rust toolchain is pinned/bin/syscheckcalls every syscall with bad arguments; six kernel bugs found this way are fixed (#11–#16 in docs/BUGS.md)- License: GNU GPL v3 or later
Full list: CHANGELOG.md. Tested on Ubuntu 24.04, not yet on Fedora.