Skip to content
IronFlockPublic

About

No description or website provided.

Topics

Resources

Security policy

Stars

0 stars

Watchers

1 watching

Forks

Latest commit

 

History

122 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

FlockFlasher

All you need to debug the application is:

npm run dev

Rolling out a new version

main takes changes only through pull requests whose checks are green. just release-patch (or -minor, -major) on a feature branch bumps the version in package.json and package-lock.json, opens the release pull request and waits for its merge. Changes that release nothing go in with just merge.

The release pull request also builds the Linux and Windows installers as a dry run (nothing published), so a broken build never reaches main. Its merge runs release.yml on GitHub Actions:

  1. The Linux and Windows installers are built, one runner each, unsigned.
  2. They are uploaded, with the latest*.yml files the in-app updater reads, to a draft GitHub Release. The updater does not see drafts.
  3. The tag v<version> is set, and the draft becomes the published, latest release.

Nothing is built or published from a laptop. A failed release run is re-run from the Actions tab. A version given up after step 2 leaves a draft behind: delete it with gh release delete v<version> --yes.

just package builds the installer of your own OS into dist/ the same way, without publishing.

macOS is not released

macOS blocks apps that Apple has not notarized, and notarization needs a membership in the Apple Developer Program, which has ended. Releases therefore hold no macOS installer. Installed Mac apps stay on v2.1.2: their updater finds no latest-mac.yml in a newer release and offers nothing. On a Mac, just package still builds an app for local use.

Windows installers are unsigned

The Windows installer carries no code signature. On first start, Windows SmartScreen shows "Windows protected your PC": choose More info, then Run anyway. Auto-updates work without a signature, because the installed app has no publisher name to check updates against.

Release secrets

The repository's release environment (deployments from main only) holds the release GitHub App's key, RELEASE_APP_PRIVATE_KEY, and its variable RELEASE_APP_CLIENT_ID. The release needs nothing else. gh secret list -R IronFlock/flasher -e release shows the names and dates, never the values.

Board catalog

FlockFlasher writes nothing but the IronFlockOS images of one catalog, the signed IronFlockOS catalog. A device is added by its device file (.flock); the user picks the hardware (menu "Hardware") and then its variant (menu "IronFlockOS variant"). The main process keeps the verified catalog, looks up the image there by its key (<image_id>@<version>, imageKey in src/shared/images.ts), downloads it, and checks the catalog's size and SHA-256 before it writes it. A flash request can name nothing else: no local image file, path or URL (assertFlashRequest in src/main/security/validation.ts, checked by the flash IPC handler and again in flashDevice). Files the app opens (file dialog, drag and drop, file association) are device files only.

The signed IronFlockOS catalog

https://instance-registry.ironflock.com/dl/ironflockos/manifest.json, published by every IronFlockOS release (IronFlockOS ADR-0009, schema). It is the device-agent's envelope, {"payload": "<base64 of the payload JSON>", "signature": "<base64 of a 64-byte Ed25519 signature>"}, at most 64 KiB, signed over the lines ironflockos-manifest:v1, ironflockos/manifest.json and the payload's SHA-256 (lowercase hex), joined by \n without a trailing newline. src/main/security/catalog.ts verifies it with Node's crypto:

  • The trusted keys are compiled in as IRONFLOCKOS_CATALOG_PUBLIC_KEYS (the Cloud KMS key ironflock-release/ironflockos-catalog, version 1, key id 036ef3bd; PEM form). To rotate, release a FlockFlasher that lists both keys before IronFlockOS signs with the new one, and drop the old line in a later release.
  • A catalog that is unsigned, tampered with, signed by another key, larger than 64 KiB, not JSON, of another schema, malformed, internally inconsistent (a board's image differs from its release) or issued before the one accepted last is refused: the app shows the error and offers no board, or, after a failed refresh, the boards verified before. It never falls back to unverified data.
  • Forward compatible: an image of a variant, kernel or architecture this FlockFlasher does not know (from a newer release) is skipped with a log line, and a board left without a known image is hidden. Unknown fields are ignored.
  • The catalog is fetched at start and again whenever the board list is shown (adding a device, opening the board list), at most every 5 minutes, and at once with Retry after an error, so a release published while the app is open appears without a restart. Idle devices keep their board and variant on the new release.
  • The last verified catalog is kept in ~/.Reflasher/ironflockos-manifest.json. It is used only when the catalog cannot be reached (network or HTTP error), with a notice. The rollback check is against the newer of it and the catalog in use.

The boards of the catalog show the production release. The menus of a device:

  1. "Hardware": the board (its model name).
  2. "IronFlockOS variant": the board's images without the installer, base (the default), base with the real-time kernel (rt) and nvidia.
  3. "How to install", only when the board's installer stick carries the chosen variant: write the variant's image to the drive (the default; the drive becomes the device's boot disk), or write the installer stick, which installs the variant onto the device's internal disk. The stick is a way to deliver a variant, not a variant. Its catalog entry lists the image ids it carries (payloads); a catalog without them (v0.10.0) means the target's base image. Payload ids that are not images of the release and target are dropped with a log line.

Every IronFlockOS image is a raw image published as .img.gz (schema 1 allows nothing else); size and SHA-256 are checked on the download, which is then decompressed into ~/.Reflasher and written from there. For an installer stick the main process checks that the chosen variant is one of the stick's payloads in the verified catalog (CatalogStore.plan) before it writes anything.

The device file on the flashed drive

After writing a .flock item, FlockFlasher puts the device file onto the FAT partition labelled IRONFLOCKOS (every IronFlockOS image has one, the installer stick included), or, as a fallback, onto the first FAT partition. It finds the partition by label and file system, not by order (src/main/utils/config-partition.ts), and mounts it if the OS did not:

  • Linux: lsblk (through sudo when the udev database has no file systems yet), udisksctl mount, else mount -t vfat through sudo into a temporary directory.
  • macOS: diskutil list and diskutil info, diskutil mount (through sudo if needed). macOS does not mount an EFI system partition by itself, and IRONFLOCKOS is the EFI system partition on amd64-efi and arm64-efi.
  • Windows: PowerShell's Storage cmdlets list the partitions; a partition without a drive letter gets one from Add-PartitionAccessPath, or from diskpart for an EFI system partition.

The file is synced, and a partition FlockFlasher mounted is unmounted again (on Windows, the drive letter it assigned is removed). The device gets <name>.flock only; its enroll removes the file after use, so no copy with the device's credentials is left behind. An installer stick also gets installer.conf with the one line payload=<image id> (the chosen variant); every other key keeps the installer's default (confirm=yes: it asks on the console before it writes).

Command line

flockflasher installs IronFlockOS without the app, for scripts, servers and machines without a desktop. It is one JavaScript file for Node 18 or later (out/cli/flockflasher.cjs, built by npm run build:cli or just cli) with nothing but Node's own modules, so it runs wherever it is copied. It writes disks on Linux and macOS; on Windows only images works.

node flockflasher.cjs images                 # the images of the signed catalog
node flockflasher.cjs drives                 # the disks it can write (--all: every disk, and why not)
sudo node flockflasher.cjs flash --flock kitchen-pi.flock --device /dev/sdb --image rpi-arm64-base

flash takes the image of the .flock's hardware when its architecture has only one (amd64: the PC), else the one --image names. --installer writes the installer stick that installs that image instead, and --installer-disk, --installer-confirm, --installer-after and --installer-allow-usb set the other keys of installer.conf (an unattended install: --installer-confirm no). --wifi-ssid and --wifi-password put a WiFi network into the copy of the .flock on the disk, as the app's WiFi fields do. It asks before it erases the disk, unless --yes. flockflasher help lists every option.

It shares the app's code for what must not differ: the signed catalog (src/main/security/catalog.ts, src/main/api/catalogs.ts, the same cache in ~/.Reflasher), the image's size and SHA-256, and the device file on the IRONFLOCKOS partition (src/main/utils/device-files.ts). It writes without etcher-sdk, so it needs no native modules (src/cli/write.ts): the .img.gz is decompressed as it is read, written in 4 MiB blocks (to /dev/rdiskN on macOS) and synced, then read back and compared (--no-verify skips that). Before it writes, it checks the disk (src/cli/disks.ts): a whole disk, writable, not the disk of the running system (/, /boot, swap, a live system's medium), not held by LVM, RAID or dm-crypt, and on macOS not an internal disk unless its medium is removable (the built-in SD card reader). It unmounts what the disk has mounted; on Linux it opens the disk with O_EXCL, which fails while anything still uses it. --image-file writes a downloaded .img.gz instead of downloading it, if it is the catalog's (same size and SHA-256).

EtcherSDK

FlockFlasher uses the etcher-sdk to flash and mount drives.

The etcher-sdk requires root permissions to access drives and flash drives.

In the current build for MacOS and Linux, we must spawn a subprocess within the FlockFlasher that gets elevated.

The easiest way to do that is to spawn a root process using sudo. This process will be a node process that runs the etcher-sdk code.

To do so, we create an external JavaScript script that uses the etcher-sdk. This script needs to point to the node_modules that are contained within the application (for production). To do so, we point to the compressed (ASAR) node_modules in the code.

To learn more about the compressed ASAR package: https://www.electronjs.org/docs/latest/tutorial/asar-archives

Since we don't want to rely on the user having the node binary installed, we can use the electron binary that comes with the application and run that as a node process to spawn an elevated process.

The electron binary can be accessed on process.execPath within the application and can be put into 'node mode' using the ELECTRON_RUN_AS_NODE environment variable.

Since we spawn a subprocess, we need to be able to read back the flashing progress of this subprocess. To do so, we print the progress data to the stdout in JSON string, which is then read and parsed in the frontend. (Line in code)

Windows

In order to access the USB drives in Windows, the etcher-sdk must include the winusb-driver-generator package. It is an optional dependency limited to Windows, so npm install adds it on Windows only.

Its native module compiles during npm install, which needs Python and the Visual Studio Build Tools (the GitHub-hosted Windows runner has both). The package bundles the WDK co-installer files it needs. npm does not fail when an optional dependency fails to build, so check that node_modules/winusb-driver-generator/build/Release holds a .node file; the release build fails without one.

AppImages (Linux)

For AppImages, it is sadly not straightforward to access the packaged node_modules within the application. Since the AppImage is technically a drive, we must first mount it to a temporary folder (link to code) and then access the packaged node modules within.

Electron / Vue

API

Electron uses an RPC API to facilitate communication between frontend and backend tasks. All operations requiring system API calls must be executed on the Electron backend.

To streamline this process, we have defined a types file (types/index.ts) that contains all the RPC endpoints. These endpoints can be called from the frontend and are executed on the backend.

The RPCs that are executed on the backend are defined in main/ipcHandlers.ts, while the frontend RPCs, defined in preload/index.ts, simply call the RPCs registered on the backend.

Store

We use Pinia, a straightforward and simple global state management tool, to manage global state in the Vue frontend. Event listeners and backend API calls are made and registered within Pinia. All stores are located in the renderer/src/store folder.

Translations

For app-wide translations, we utilize the I18next internationalization framework. Adding, updating, creating, and deleting translations is as simple as creating or editing a file in the renderer/locales folder and then registering the new languages in the renderer/src/main.ts file.

Flashing runtime (why there is a bundled Node binary)

The app runs on Electron 43 with etcher-sdk 10 (Node 22 via .nvmrc). The old Electron-19 pin existed because etcher-sdk breaks on Electron 20+ (balena-io/etcher#4087): Electron's V8 memory cage (Electron 21+) forbids the external buffers that etcher-sdk's @ronomon/direct-io needs for O_DIRECT block writes, so running the flasher via ELECTRON_RUN_AS_NODE aborts (napi_create_external_buffer assertion).

The fix: the elevated flash + unmount subprocesses run under a bundled standalone Node binary (no memory cage), fetched per-platform by scripts/fetch-node.js (postinstall, git-ignored) into resources/binaries/<platform>/. The flash-path native deps that are N-API (drivelist, @ronomon/direct-io, xxhash-addon) work under both runtimes unchanged; mountutils (nan) is rebuilt for the bundled Node's ABI by scripts/rebuild-flash-natives.js (postinstall, after install-app-deps; CI skips these packaging steps, see scripts/postinstall.js). Keep NODE_VERSION in sync between those two scripts. When upgrading Electron, bump both package.json electron and electron-builder.yml electronVersion.

The native modules (drivelist, mountutils) compile from source on macOS, so npm install needs a complete Command Line Tools install. If it fails with 'functional'/'cstdlib' file not found, the CLT is missing its C++ stdlib headers — reinstall it: sudo rm -rf /Library/Developer/CommandLineTools && sudo xcode-select --install.

On Linux, GCC fails on Electron 43's V8 headers with expected identifier before '__attribute__' when deprecation warnings are on, which node-gyp turns on for every addon. scripts/prepare-natives.js turns them off for the Electron build on Linux. The usb module also needs the udev headers: sudo apt-get install libudev-dev.

About

No description or website provided.

Topics

Resources

Security policy

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages