Firmware Boot Chain

Overview

A boot chain is a relay of stages, each responsible for setting up just enough state to hand control to the next: ROM to firmware (SBI/UEFI), firmware to bootloader, bootloader to OS. Every handoff has a contract: where the next stage lives, what registers/tables it expects, and what memory is already set up.

Core principle: Each stage owns a contract with the next. Most boot failures are a broken contract at exactly one handoff, so isolate which handoff fails before theorizing about the stage itself.

When to Use

• Writing or porting firmware (RISC-V SBI, UEFI services, ACPI table provision)
• Chaining a bootloader (Limine, GRUB, U-Boot) into an OS kernel
• A stage loads but the next never starts, or starts and immediately faults
• Adding measured boot / TPM PCR extension to the chain
• Discovering peripherals from a device tree or ACPI at firmware time

Map The Handoffs First

Write down the relay before debugging:

ROM -> firmware (SBI/UEFI) -> bootloader -> OS kernel
        provides: SBI calls,    loads:        expects: a0=hartid,
        memory map, ACPI/DTB     kernel+initrd  a1=DTB/ACPI ptr, MMU off

For each arrow, name: the entry address, the register/pointer contract, and the memory/translation state. The failing arrow is your bug location.

Firmware Responsibilities

• Provide the platform description. Hand the next stage a device tree (DTB) or ACPI tables describing memory, CPUs, and peripherals. Probe peripherals from this description rather than hardcoding addresses, so one firmware serves multiple board memory maps.
• Set the entry contract precisely. RISC-V convention passes hartid and a pointer to the platform description in fixed registers; get them exactly right. The next stage trusts them blindly.
• Build-time configure the memory base. RAM base and the firmware's own load address differ per board (for example external DRAM at a high base on one board, on-chip SRAM on another). Make these build-time parameters, not constants buried in one file.

Bootloader Handoff Gotchas

These bite when chaining a general loader (for example Limine) into an OS:

• Filesystem format constraints. The loader may require a specific boot filesystem (FAT16, not FAT32) and a specific layout. Get this wrong and the loader silently finds nothing.
• Timeout and entry config. A nonzero menu timeout can stall an automated boot; a missing or misnamed entry just drops to a prompt.
• Ramdisk/module placement. The loader places initrd/modules in memory; make sure that placement doesn't collide with where the kernel expects to run or with the stack (see bare-metal-bringup).

Measured Boot (TPM)

If the chain is measured:

• Drive the TPM over its real interface (TIS for TPM 2.0) and gate the whole probe on the platform description actually advertising a TPM. Don't assume presence.
• Measure before you transfer control. Each stage extends a PCR with a hash of the next stage (and relevant config) before jumping to it. Measuring after handoff measures nothing useful.
• Use the standard protocol surface (for example EFI_TCG2_PROTOCOL) and emit a TCG2 event log so the measurements are verifiable later.
• Test against a software TPM (swtpm) on the bench before trusting real silicon.

Red Flags

Smell	Do instead
Hardcoded peripheral addresses	Probe from DTB/ACPI
RAM base as a constant	Build-time parameter per board
"It doesn't boot" with no stage isolated	Identify the failing handoff first
TPM probe with no presence gate	Gate on the platform description
Extending a PCR after the jump	Measure-then-transfer
A rebuild that has no effect	Confirm the flashed slot is the one the ROM jumps to
Silent boot from a quad-read flash	Fall back to standard 0x03 to isolate flash config
Writing optional CSRs unconditionally	Probe each under a temp trap handler, skip absent ones

Midstall House Style

• Weir is the reference: pure-Zig RISC-V firmware (SBI/UEFI/ACPI), Limine to NixOS, measured boot with a TPM 2.0 TIS driver and TCG2 event log. Peripherals are discovered from the DTB; RAM base is build-time.
• See boot-handoff-traps.md in this directory: the first-stage failures that boot silent or run a stale image (reset-into-flash XIP needs a BRAM maskrom, a stale image at the real boot slot, a quad/QE flash mismatch, FSBL scratch colliding with the copy destination, a device-tree cell-size parser mismatch, and probing optional CSRs).
• Write docs and comments in ASD-STE100 Simplified Technical English. No em dashes, no emoji. Pairs with bare-metal-bringup for the early-output and translation rungs.