Operating systems: review and evidence matrix¶
Reviewed during the 2026-10-18 course revision. This matrix distinguishes conceptual/source review from executed experiments. A revised explanation is not automatically a hardware validation. Return to the course guide for the learning sequence.
Lecture coverage¶
| Lecture | Main contract reviewed | Reproducible exercise or remaining boundary |
|---|---|---|
| 1 | Privilege, kernel/BSP scope, sysfs/procfs and driver isolation | Read-only host inspection; exact board release remains target-specific |
| 2 | Thread identity, CoW, reaping, affinity, quota and wait states | SPSC real threads; kernel internal layouts are not UAPI |
| 3 | IRQ/wakeup/execution separation, workqueue contexts and teardown | RT wakeup measurement; no device IRQ measurement claimed |
| 4 | Syscall error contracts, shm sizing, clocks, epoll, timer slack | RT/BPF build labs; shared-memory lifecycle shown as pseudocode |
| 5 | Boot stages, UEFI, DT cells/status, module build and trust scope | Links real DT tools and build-only module; no module load |
| 6 | Fair service versus response deadlines and CPU reservations | RT demo; privileged admission/scheduling is a separate milestone |
| 7 | PREEMPT_RT mechanisms versus measured/analytical bounds | Finite wakeup experiment, not a worst-case proof |
| 8 | Eligible/effective CPUs, IRQ/housekeeping and interference | Obtained-affinity test; no global host isolation changes |
| 9 | Ownership, waiting, completion reuse and locking contexts | Existing RT/SPSC exercises; no live DMA abort test |
| 10 | Publication, slot release, progress and reclamation | Native SPSC plus sanitizer and intentional-overwrite checks |
| 11 | Lock-order graph, PI/ceiling limits, corrected Pathfinder account | Worked blocking arithmetic; no privileged PI timing experiment |
| 12 | Mapping, faults, lifetime and physical-address distinctions | Worked buffer/mapping analysis |
| 13 | Translation counts, TLB cost and huge-page policy | Explicit binary-unit arithmetic; no huge-page speed claim |
| 14 | Allocation, contiguity, CMA and failure paths | API/source review; no device allocation stress test |
| 15 | DMA direction/counts/order/completion, managed-memory APIs | Six-entry/four-segment worked example; no GPU/device run |
| 16 | NUMA topology, policy versus residency and measured costs | Documented placement experiment; no multi-node result invented |
| 17 | Probe publication, dependency errors, devm lifetime, overlays | Linked DT compilation/overlay lab; physical bring-up remains |
| 18 | ioctl dispatch/UAPI, mapping lifetime, V4L2/GPU ownership | Driver fragments explicitly labeled; no camera capture claimed |
| 19 | io_uring terminal completion and DMA-BUF sharing | Bounded-lifetime reasoning; no invented IOPS or liburing runtime |
| 20 | One-way PCIe ceilings, NVMe queue limits, conditional P2P | Worked bandwidth/Little's-law arithmetic; no P2P benchmark |
| 21 | Durable replacement, snapshots, tmpfs and OverlayFS | Read-only storage inspection; no host filesystem remount |
| 22 | RPMB scope, endurance units, storage identity and A/B state | Links actual file-based OTA protocol tests, not firmware/power-cut proof |
| 23 | Quota/mask semantics, toolkit dependencies, MIG/MPS scope | Linked actual CPU/container lifecycle lab; Orin MIG claim removed |
| 24 | Pinned deployed-system claims and full-path timing | Links maintained AGNOS/Jetson material rather than old platform assumptions |
| 25 | Recipe→package→image boundaries and evaluated variables | Reuses actual native package and 922/1105-recipe metadata parses; no target image boot |
| 26 | BPF pairing, event ABI, loss, compilation versus load/attach | Real BPF ELF/skeleton/reader build and malformed CLI checks |
Lectures 6–10, 12–14, 16 and 24 retain the earlier substantive revision in this mission. This pass corrected the other bodies and revisited 11/15 where old historical/device claims remained. All eight consolidation notes were checked for consistency; notes 1–3 and 6–8 were rewritten to remove conflicting short-form claims. The 14-question final practice now checks buffer reuse, durability, resource limits, and build-evidence boundaries as well as scheduling and memory.
Executed OS checks¶
Run from the repository root with the listed prerequisites:
python3 -m unittest discover -s 'Phase 1 - Foundational Knowledge/3. Operating Systems/labs' -p 'test_*.py' -v
The recorded review used an isolated libbpf header directory through COURSE_BPF_INCLUDE; without those optional prerequisites the BPF suite explicitly skips. A skip is not a native build pass.
| Suite | Actual result | What it exercised |
|---|---|---|
test_rt_demo.py |
3 passed | C17 warnings-as-errors; actual default/allowed CPU affinity; malformed options; rejected memory lock |
test_spsc_queue.py |
3 passed | Optimized reference/real-thread transfer; ASan/UBSan; intentional missing-full-check defect detected |
test_bpf_build.py |
3 passed | Clang BPF object; bpftool skeleton; C17 reader; ELF sections; seven invalid argument cases |
The combined run passed all 9 tests. No BPF program was loaded or attached, no kernel module was loaded, and no camera/GPU/flash operation was performed by this OS test suite. Linked labs have their own narrower validation records and should be cited on those terms.
Sources and maintenance¶
Version-sensitive statements link to primary Linux/man-pages, NVIDIA, Yocto/BitBake, liburing, and OP-TEE documentation at the point of use. The Pathfinder correction uses a preserved first-person letter by its flight-software lead, identified as a historical account. Numerical examples state their units and assumptions; unmeasured universal latency/IOPS tables were removed.
For a later target run, add the exact kernel/configuration, BSP/firmware, architecture, device identities, tool versions, workload, permissions, measurement clocks, failures, and artifacts. Keep “source reviewed,” “compiled,” “executed on a host,” and “validated on hardware” separate in the resulting report.