Difference between revisions of "Flashmap"
m (Remark on chip flashing implications of an in-image flashmap)
(Start cataloging work to be done)
|Line 173:||Line 173:|
==How do we fix it (the solution being pursued)==
==How do we fix it (the solution being pursued)==
Revision as of 22:27, 17 February 2015
Toward a unified representation for the layout of coreboot flash images
N.B. The changes described herein are being made as part of the Chromium OS project; as such, they will initially be committed to the project's own fork of the main coreboot repository, which is available at https://chromium.googlesource.com/chromiumos/third_party/coreboot. Unless otherwise noted, the paths and processes described throughout this page are as they exist(ed) in a checkout of the master branch of the Chromium OS sources as they appeared at the beginning of 2015. One of the guiding design principles is to keep the tools general enough that they will be helpful to others, and the resulting work will be upstreamed to the main repository once it has been regression-tested in the context of Chromium OS hardware.
How it's currently done (how the Chromium OS project presently constructs firmware images)
Most Intel-based Chromium OS devices currently use an 8 MB firmware image that includes—among other things—the Intel ME firmware, a copy of coreboot including the ramstage and depthcharge (bootloader) payload, two additional copies of the ramstage and bootloader payload, and a separate SeaBIOS payload. The primary description of this format exists in board-specific flattened device tree files, which are used by a script called cros_bundle_firmware to modify the image produced by the coreboot build system. For instance, the layout of the Panther board's firmware exists at https://chromium.googlesource.com/chromiumos/platform/depthcharge/+/master/board/panther/fmap.dts, and results in a final image that looks like this:
Intel Chromium OS firmware flash layout with component packaging notes
|Section||Offset||Contents||Original source||cros_bundle_firmware flag(s)||Packaging procedure||Coreboot Kconfig entr[yi](es)|
|0x700000||Bootstub (coreboot image)||coreboot.rom (coreboot build system)||--coreboot coreboot.rom
|cros_bundle_firmware helper adds both depthcharge.elf as a payload (which—md5sums confirm—is exactly equivalent to just adding depthcharge.payload) and a compiled (then mod if ied) version of fmap.dts to the existing CBFS|
|0x611000||GBB (Google Binary Block)||/chromeos-config/ entries in fmap.dts||--dt fmap.dts
|cros_bundle_firmware helper generates and inserts it|
|0x610800||RO-FWID (Firmware ID)||/model entry in fmap.dts concatenated with build number||--dt fmap.dts||cros_bundle_firmware helper performs the concatenation|
|0x610000||FMAP (Flash MAP)||fmap.dts itself||--dt fmap.dts||cros_bundle_firmware helper generates it while assembling the ultimate image file|
|0x600000||RO-VPD (Vital Product Data)||cros_bundle_firmware fills this section with 0xff s|
|0x400000||Legacy (SeaBIOS image)||seabios.cbfs||--seabios seabios.cbfs||cros_bundle_firmware helper inserts the file verbatim, as confirmed by an md5sum on the final image|
|0x3f8000||RW-VPD (Vital Product Data)||cros_bundle_firmware fills this section with 0xff s|
|0x3f6000||Vblock-dev (third-party kernel signing keys)||cros_bundle_firmware fills this section with 0xff s|
|0x3f4000||Shared-data (RW firmware calibration data)||cros_bundle_firmware fills this section with 0x00 s|
|0x3f0000||ELOG (Event LOG)||cros_bundle_firmware fills this section with 0xff s|
|0x3e0000||MRC-cache (Memory Reference Code training data)||cros_bundle_firmware fills this section with 0xff s|
|0x3dffc0||FWID-B||(Copy of RO-FWID above)|
|0x300000||Main-B (copy of coreboot ramstage and payload)||(Same as main-A below)|
|0x2f0000||Vblock-B (signing keys)||(Same as vblock-A below)|
|0x2effc0||FWID-A||(Copy of RO-FWID above)|
|0x210000||Main-A (copy of coreboot ramstage and payload)||depthcharge.payload and ramstage.stage||--uboot depthcharge.payload
--add-blob ramstage ramstage.stage
|Because fmap.dts lists its type as "blob boot,ramstage" , cros_bundle_firmware concatenates the blob called ramstage onto the end of coreboot|
|0x200000||Vblock-A (signing keys)||cros_bundle_firmware helper signs the main-A section and puts the signature here|
|0x001000||ME (Intel Management Engine firmware blob)||coreboot.rom||--coreboot coreboot.rom||Present in the original image generated by the coreboot build system (which simply inserted some Intel blob verbatim)|
|0x000000||FD (Intel Firmware Descriptor header)||coreboot.rom||--coreboot coreboot.rom||cros_bundle_firmware helper invokes ifdtool to create the final image, using the original coreboot.rom file's ME and "skeleton" (template) IFD, and concatenating on verbatim the rest of the pieces it assembled to form the (Intel-dubbed) "BIOS" section|
What's so bad about that (the pitfalls of this build model that we hope to solve)
While the layout of the flash image is obvious from a board's fmap.dts file, the actual process by which the image is constructed—and most importantly, the transformation and other processing applied to each input file—is poorly documented and not widely understood. After examining the overall assembly procedure and distilling it into the above tables, the following shortcomings are apparent:
The build system organization favors divergence from upstream coreboot
This implies two subproblems: upstreaming code from Chromium OS to the main coreboot source tree is difficult, and it isn't easy to build upstream coreboot for a Chromium OS device. It's important to note that, while related, these are not identical shortcomings: while the upstreaming difficulty certainly prohibits building some devices, other devices have enough code already pushed that it would be possible to run them if the coreboot build system were able to produce an image that was ready to flash on the hardware.
As can be seen from the above table, there's no technical reason that sufficient functionality to build a complete image can't be upstreamed to coreboot itself: with the exception of ramstage.stage—which is actually just another copy of a binary generated by the coreboot build system—everything that needs to be added to the coreboot-generated image is either a binary blob or compiles without depending on any part of coreboot besides libpayload.
There is error-prone duplication in the image layout configuration
Because image assembly is currently a multistage process, several offsets and sizes need to be duplicated across multiple configuration files; if these files are not manually kept in step, the process can result in an obviously broken (e.g. part or all of coreboot or the CBFS is overwritten with another binary) or more subtly broken (e.g. coreboot assumes an FMAP exists at a particular flash location, but when it attempts to find it at runtime, it isn't there). To make matters worse, the redundant configuration parameters are stored in separate repositories, so checking out a different version of one boot component can silently result in a mismatch. Adding insult to injury, one of said configurations is stored in binary form within the IFD, which forces us to check in an IFD blob for each boardmaking it even more difficult to rearrange the flash layout.
Why you should care (how this pertains to all coreboot users)
Short answer: For owners of certain Chromebooks/Chromeboxes: it would make it easier to build upstream coreboot for their devices, as well as easier to upstream more boards. For everyone else: the Chromium OS firmware includes several features that are generally useful, and landing tools for building Chromebook/Chromebox images would enable other users to easily add modular, precompiled components to their image builds. This would allow them to avoid maintaining separate, last-mile build systems that use separate configuration files, and hence make it easier to apply coreboot to advanced and unconventional setups.
The benefits reach beyond the mere construction of a firmware image, however: on many modern systems, integrating flashmap support will make it easier to flash images onto the board. For example, on recent x86 platforms, the Intel Management Engine reads its firmware from a section of the flash memory. Post-boot, access to this portion of the flash chip from the CPU is often completely disallowed; thus, it is impossible to perform actions such as reflashing that portion of the flash without an external debugger. Unless the coreboot build system is able to extract the flash layout from the existing image, crafting a replacement image that can be flashed from the host itself is very difficult. Including a flashmap in a standard format directly in the image would provide the foundations necessary for the build system to automatically construct images that could be flashed on such systems. However, in order to facilitate this future direction, it's important that whatever tools are added use the image's—rather than the source tree's—flashmap representation whenever possible, so that end users can easily construct their own images even if the latter cannot be redistributed to them.
How do we fix it (the solution being pursued)
What needs to be done?
- On Chrome OS devices, some offsets and sizes are currently read from an FMAP section in Google's flashmap format. This is currently controlled by
CONFIG_CHROMEOS, but should be migrated to a separate Kconfig key so that it can be used for other purposes as well. To begin with, the build system can just assume that an FMAP region will be manually inserted after the initial build process.
- The source tree's representation of the flashmap is currently stored in the board directory of the depthcharge codebase and parsed by the separate cros_bundle_firmware script. The ability to generate an FMAP region—and the per-board dts files—should be moved into the coreboot repository, where it can be used to construct the firmware image. There already exists a prototype reimplementation of much of cros_bundle_firmware that might be a good basis for this portion of the work.
- Make the coreboot build system capable of inserting blobs into the firmware image, external to the CBFS. One important question here is whether the FMAP deserves to go in a section of its own, or be placed in the CBFS: it's a bit of a chicken-and-egg problem, because the location of the CBFS could be stored in the FMAP so it didn't need to be baked into the coreboot build itself. This may require changes to cbfstool, but will allow the replacement of cros_bundle_firmware.
- Generate per-board IFD headers based on the per-board plaintext dts files and (possibly) some binary template. This should allow us to eliminate this class of blobs from the repositories. This may require some changes to ifdtool.
- Implement the ability to build images with the same layout as an existing image. This will help users to easily build modified images for devices that already run coreboot. Once again, the issue of where the FMAP should be placed is important, because we need to be able to find it in the existing image.
- Add functionality to copy nonredistributable portions of an existing image into a new one. This way, users who own a hardware device that ships with vender blobs will be able to build firmware images for it without violating vendor agreements.