11 KiB
Rebuilding coreboot image generation
Current situation
Chrome OS (CrOS) probably has the most complex image bundling process in the coreboot ecosystem. To make CrOS features more accessible to the wider coreboot community, we want to move these capabilities into upstream coreboot’s build system.
Right now, the CrOS build system creates coreboot images, and various
instances of the payload (with different configuration options), plus some
more files (eg. EC firmware), then passes them to a CrOS-specific utility
(bundle_firmware.py) to build the final image from that.
bundle_firmware adds a flashmap (fmap) to the final image and creates
additional CBFS filesystems in fmap regions. It then extracts some files from
the original CBFS region (that was put in place carefully to later match to
the default fmap region) and copies some of them into the others, as well as
putting more data (eg. the bitmap data, keys) as raw data into other fmap
regions.
With the recent addition of more files to CBFS, both on the coreboot side (dsdt, FSP, and so on) and with Chrome OS specifics (eg. more files describing boot screens) we either need to expand the scope of bundle_firmware or move the capability to build complex images to upstream coreboot’s build system. This document proposes to do the latter and outlines how this could be achieved.
Problems with the current build system parts
One common sentiment is that it should be possible to reuse some of the existing mechanisms that are supposed to be supplanted by this. The main concern during this design that precluded their use was that none of them provides a comprehensive solution to building complex coreboot based images:
- fmap.dts and fmd provide a flash layout, but no assignment of files of regions
- cbfs-files-y ends up as an internal make variable using
weird|formatting|to|deal|with|make’s|limitations - make isn’t powerful enough to deal with ordering these entries in said variable to guarantee success if there’s enough room for the files. While that could be added, that becomes more make macro work indistinguishable from magic that people fail to understand, break and with good reason complain about to work around such issues, Chrome OS firmware uses a custom tool with even more special cases to finally build the image it needs. If coreboot upstream is to support vboot, it should also be powerful enough not to need magic tools that only live within downstream projects.
Requirements
A complete Chrome OS coreboot image consists of (depending on the device)
- platform specific data in raw fmap regions (eg IFD, ME firmware),
- the bootblock (coming from the bootblock),
- three copies of coreboot, consisting of the stages (verstage, romstage, ramstage) plus data,
- depthcharge plus data (with each of the coreboot copies),
- EC firmware files (with each of the coreboot copies),
- signatures over several parts of the image and
- some final checksumming over parts of the image to satisfy boot ROM tests on ARM
A complete upstream coreboot image (with fallback/normal switch configuration, using a yet to be implemented switching scheme based on fmaps) consists of
- platform specific data in raw fmap regions (eg IFD, ME firmware),
- two copies of coreboot, consisting of
- the bootblock and
- the stages (romstage, ramstage) plus data,
- payload plus data (with each of the coreboot copies),
Since a single platform is potentially built with different payload configurations (eg. modding a Chromebook to not use the verified Chrome OS boot scheme), some concerns need to be kept separate:
- Platform requirements that have nothing to do with the payload or boot schemes
- IFD, ME, … need to copied to the right place
- boot ROM requirements such as checksums must be honored
- Payload/boot scheme requirements
- Having one to three regions with certain files copied into them
Proposal
The proposal is based on manifest files that describe certain aspects of the final image. The number of manifest files may change over time, but this seems to be a reasonable approach for now. As long as coreboot uses fmap and cbfs, there should be few need to change the language, since composition is done through files.
The final image is generated by a utility that is handed a number of manifests
and the size of the flash (derived from CONFIG_ROM_SIZE). These manifest files
deal with different concerns, with the following an example that should match
current use cases:
Chipset manifest
The chipset details if there are any non-coreboot regions, and assigns them names, locations, sizes and file contents and prepares a region for what is “platform visible” (eg. IFD’s BIOS region) that may be of flexible size (depending on the flash chip’s size). For the purpose of this document, that region is called “BIOS”. It can also specify if there’s a post processing requirement on the final image.
coreboot manifest
coreboot provides lists of the files it generates for each category it’s building (eg. bootblock, verstage, romstage, ramstage). They not only contain the stages themselves, but also additional files (eg. dsdt belongs to ramstage since that’s where it is used)
Boot method manifest
The boot method manifest can subdivide the BIOS region, eg. using it directly (for coreboot’s “simple” bootblock), splitting it in two (for coreboot’s fallback/normal) or in many parts (for Chrome OS, which requires two CBFS regions, one for GBB, several for VPD, …). It also specifies which of the file lists specified earlier belong in which region (eg. with verstage verifying romstage, verstage needs to be only in Chrome OS’ RO region, while romstage belongs in RO and both RW regions). It can also specify a post processing step that is executed before the chipset’s.
Payload and additional manifests
External components should also provide manifests to add files to categories. This way the payload and other components (eg. EC firmware) can be developed without needing to touch the central boot method manifest (that likely resides in the coreboot tree, given that coreboot needs to deal with choosing fmap regions already).
coreboot build system
The coreboot build system will be split more distinctly in two phases: The first is about building the files (with results like romstage.elf), while the second phase covers the assembly of the final image.
By having a global picture of the final image’s requirements, we can also avoid issues where files added earlier may prevent later additions that have stricter constraints - without resorting to hacks like https://chromium-review.googlesource.com/289491 that reorder the file addition manually.
Example
As an example, we’ll define an Intel-based board with a postprocessing tool (something that doesn’t exist, but isn’t hard to imagine):
It specifies an IFD region, an ME, and the BIOS region. After the image is built, the entire image needs to be processed (although the tool likely works only on a small part of it)
It’s built in a Chrome OS-like configuration (simplified at places to avoid distracting from the important parts), so it has three CBFS regions, and several data regions for its own purpose (similar to GBB, FWID, VPD, …). After the regions are filled, one data region must be post-processed to contain signatures to enable verifying other regions.
Chipset manifest
================
# A region called IFD, starting at 0, ending at 4K
region IFD: 0 4K
# Add the specified file “raw” into the region.
# If the file is smaller than the region, put it at the bottom and fill up
# with 0xff
raw IFD: build/ifd.bin align=bottom empty=0xff
# Call the postprocessor on the data that ends up in IFD (in this example it
# might lock the IFD)
postprocess IFD: util/ifdprocess -l
# a region called ME, starting at 4K, ending at 2M
region ME: 4K 2M
raw ME: 3rdparty/blobs/soc/intel/xanadu/me.bin align=bottom empty=0x00
# a region called BIOS, starting at 2M, filling up the free space
# filling up fails (build error) if two regions are requested to fill up
# against each other
region BIOS: 2M *
# This would define a region that covers the last 4K of flash.
# The BIOS region specified above will end right before it instead of
# expanding to end of flash
# region AUX: -4K -0
# specify the tool that post-processes the entire image.
postprocess image: util/intelchksum/intelchksum.sh
coreboot manifest
=================
# declare that build/verstage.elf belongs into the group ‘verstage’
# these groups are later referred to by the “cbfs” command.
group verstage: build/verstage.elf stage xip name=fallback/verstage
group romstage: build/romstage.elf stage xip name=fallback/romstage
group ramstage: build/ramstage.elf stage name=fallback/ramstage
compression=lzma
group ramstage: build/dsdt.aml compression=lzma
boot method manifest
====================
# Define RO as region inside BIOS, covering the upper half of the image.
# It’s a build error if the result crosses outside BIOS.
# math expressions are wrapped with ( ),
# and mentions of regions therein always refer to their size
subregion BIOS RO: ( image / 2 ) -0
# Define RW to cover the rest of BIOS.
# The order of RW and RO doesn’t matter except to keep comments clearer.
# Dynamic items like RW (“*”) will be sized to fill unused space after
# everything else is placed.
subregion BIOS RW: 0 *
# It may be necessary to separate the RO/RW definition into another manifest
# file
# that defines the RO configuration of the flash
# Some more subregions, with dynamically calculated sizes
subregion RW RW_A: 0 ( RW / 2 )
subregion RW RW_B: * -0
subregion RW_A FW_MAIN_A: RW_A * -0
subregion RW_A VBLOCK_A: 0 64K
# foo +bar specifies start + size, not (start, end)
# also, start is given as “the end of VBLOCK_A”
# (while using a region in the “end” field means “start of region”)
subregion RW_A FWID_A: VBLOCK_A +64
# To make the example not too verbose, RO only has the CBFS region
subregion RO BOOTSTUB: 0 *
# Postprocess the data that ends up in VBLOCK_A,
# passing the listed regions as additional arguments.
# Circular dependencies are build errors.
postprocess VBLOCK_A(FW_MAIN_A): signtool
# binding files to regions indirectly through groups
cbfs BOOTSTUB: verstage, romstage, ramstage, payload
cbfs FW_MAIN_A: romstage, ramstage, payload
# defining defaults: unless overridden, in all regions that use CBFS (“*”),
# we want all files to come with SHA256 hashes.
# Wildcard defaults have lower priority than specific defaults.
# Other conflicts lead to a build error.
cbfsdefaults *: hash=sha3
payload manifest
================
group payload: payload.elf payload
group payload: bootscreen.jpg name=splashscreen.jpg type=splashscreen
EC firmware manifest
====================
# overrides the cbfsdefault above
group payload: ecrw.bin name=ecrw hash=sha256
group payload: pdrw.bin name=pdrw hash=sha256
manifest parsing
The exact BNF is work in progress.
Some parser rules are
- one line per statement
- '#' introduces a command until the end of line
Some processing rules
- When there’s a conflict (eg. two statements on what to do to a region, overlap, anything that can’t be determined), that is a build error.
- the order of statements doesn’t matter, enabling simple addition of more manifests where the need arises.