> tl-dr:
> - maximize OCI cache hit ratio across nightlies/releases/PRs/etc;
> - publish simple `Version:`'s that don't include a crazy hash in repo and images
> - introduce `output/packages-hashed` directory
> - radically change the `output/debs` directory structure
- simplify artifact's `prepare_version()` method for `deb` and `deb-tar` artifacts:
- `artifact_base_dir` and `artifact_final_file` will now be auto-calculated; thus removed from each artifact (except `rootfs`)
- `artifact_deb_repo` ("global", "jammy", "bookworm") is now required; "global" means common across all RELEASES
- `artifact_deb_arch` is now required, "all" is arch-independent, otherwise use `${ARCH}`
- `artifact_map_debs` is now auto-calculated based on the above, and shouldn't be specified manually
- `artifact_final_version_reversioned` is optional, and can force the final version of the artifact (specific for the `base-files` case)
- artifacts that need special handling for reversioning can add function names to `artifact_debs_reversion_functions` array (`base-files` and `bsp-cli` cases)
- artifacts `prepare_version()` should set `artifact_version`, but _never_ include it in other variables; `artifact_version` is now changed by framework after `prepare_version()` returns
- no longer use/refer/mention `${REVISION}` when building packages. All packages should be `${REVISION}`-agnostic.
- `${REVISION}` (actually, `artifact_final_version_reversioned`) will be automatically swapped in the `control` file during reversioning
- `fakeroot_dpkg_deb_build()` now takes exactly two arguments: the directory to pack, and the deb ID (key of `artifact_map_packages` dict); add this change in all the artifact's code for this
- `obtain_complete_artifact()`:
- automatically adds `-Rxxxx` "revisioning-hash" to `artifact_version`, by hashing the revisioning functions and any `artifact_debs_reversion_functions` set
- calculates more complex subdirectory paths for both the `output/packages-hashed` and `output/debs`/`output/debs-beta` directories
- with the new subdirectories we can be sure a re-version is already done correctly and can skip it (eg, for partial `download-debs` re-runs)
- in the future we can automatically clean/remove old versions that are no longer relevant based on the dir structure
- exports a lot more information to JSON, including the new subdirectory paths
- comment-out code that implemented `skip_unpack_if_found_in_caches`, I'm very unsure why we had this in the first place
- `obtain_artifact_from_remote_cache()`
- for `deb` type artifacts, OCI won't preserve the subdirectory structure, so move downloaded files to the correct subdirectory manually
- this is not needed for `deb-tar`, since that can preserve the dir structure itself
- introduce `artifacts-reversion.sh` and its main function `artifact_reversion_for_deployment()`
- this has the logic for reversioning .deb's, by `ar`-unpacking them, changing `control.tar` (and possibly `data.tar`), handling `.xz` compression, etc.
- also handles hashing those functions, for consistency. Any changes in reversioning code actually change the artifact itself so we're not caught by surprise
- by default, it changes `control` file only:
- replace `Version:` (which is the hash-version originally) with `artifact_final_version_reversioned` (which is mostly just `${REVISION}`)
- add a custom field `Armbian-Original-Hash:` with the original hash-version
- `artifact_reversion_for_deployment()` is called by
- new CLI wrapper `cli_obtain_complete_artifact()`, used for CLI building of specific artifact, but also for `download-artifact`
- `build_artifact_for_image()` used during image build
- `armbian-bsp-cli-deb.sh`: move `${REVISION}` related stuff from the main package build to new reversioning functions.
- `artifact-armbian-base-files.sh`: move `${REVISION}` related stuff from the main package build to new reversioning functions.
- `kernel`:
- add some custom fields to `DEBIAN/control`:
- `Armbian-Kernel-Version:` / `Armbian-Kernel-Version-Family:` (for future use: cleanup of usage of `Source: ` field which should be removed)
- declutter the `Description:` field, moving long description out of the first line
- obtain `IMAGE_INSTALLED_KERNEL_VERSION` from the reversioned deb (this is still a hack and has not been fixed)
- `uboot`:
- declutter the `Description:` field, moving long description out of the first line
- use the reversioned .deb when deploying u-boot to the image
- `main_default_build_packages()` now stores reversioned values and complete paths to reversioned .deb's
- `list_installed_packages()` now compares custom field `Armbian-Original-Hash: `, and not the `Version:` to make sure debs in the image are the ones we want
- `install_artifact_deb_chroot()` is a new wrapper around `install_deb_chroot()` for easy handling of reversioned debs
- use it everywhere `install_deb_chroot()` was used in `distro-agnostic.sh` and `distro-specific.sh`
- change deployment dir to not include REVISION or ARCH
- get rid uboot's CHOSEN_UBOOT, REVISION and ARCH in the directory name.
- no two u-boot debs can be installed in the same machine anyway
- this should avoid (late) patching errors that might happen during a point release bump like `6.4.5` -> `6.4.6` cos we'd be using the wrong cached drivers patch
- using the SHA1 will instead (possibly) trigger the "real patching failure", during drivers-harness when building a new driver patch cache
- also try to cleanup old caches in the old format so we've not many leftovers -- each patch is ~150mb
- `ANSI_COLOR=none` is used when we're driving compile.sh from Python
- some debugging messages had newlines in them, thus making Python mark those as `[LEAKED]`
- also avoid log archiving during individual download jobs via SKIP_LOG_ARCHIVE=yes
- I've tested with PARALLEL_DOWNLOADS_WORKERS=16 and it saturates my gigabit link, ghcr.io is great at reads
- more than ~16-ish might be too much though
> How to use:
>
> `./compile.sh inventory` - does just the board inventory; look for output in `output/info`
>
> `./compile.sh targets-dashboard` - does inventory, targets compositing, and images info; look for output in `output/info`, read the instructions output by the command if you want to load the OpenSearch dashboards.
>
> `./compile.sh targets` - does the full targets compositing and artifacts, look for output in `output/info`
>
> If you don't have a `userpatches/targets.yaml`, _one will be provided for you_ defaulting to Jammy minimal CLI
> and Jammy xfce desktop, for all boards in all branches. You can pass filters via `TARGETS_FILTER_INCLUDE=...` to narrow.
>
- board JSON inventory:
- more generic regex parsing of variables from board files:
- all top-level (non-indented) variables are parsed and included in the JSON board inventory
- this allows us to add new variables to the board files without having to update the parser
- variables can be bare, `export` or `declare -g`, but **_must_ be quoted** (single or double) and UPPER_CASE
- some special treatment for certain variables:
- `KERNEL_TARGET` is parsed as a _comma-separated_ list of valid BRANCH'es
- `BOARD_MAINTAINER` is parsed as _space-separated_ list of valid maintainer GH usernames as `BOARD_MAINTAINERS: [...]` in the JSON
- script complains if `BOARD_MAINTAINER` is not set in core boards. Empty is still allowed.
- `HAS_VIDEO_OUTPUT="no"` causes `BOARD_HAS_VIDEO: false` in the JSON (for desktop-only inventorying, see below)
- introduce `not-eos-with-video` in `items-from-inventory` at the targets compositor
- the same as `not-eos`, but with added `BOARD_HAS_VIDEO: true` filter, see above
- introduce `TARGETS_FILTER_INCLUDE` for targets compositor
- this filters the targets _after_ compositing (but before getting image info), based on the board inventory data
- it's a comma-separated list of `key:value` pairs, which are OR-ed together
- new virtual info `BOARD_SLASH_BRANCH` post-compositing inventory for filtering of a specific BOARD/BRANCH combo (e.g. `odroidhc4/edge`)
- some interesting possible filters:
- `TARGETS_FILTER_INCLUDE="BOARD:odroidhc4"`: _only_ build a single board, all branches. JIRA [AR-1806]
- `TARGETS_FILTER_INCLUDE="BOARD_SLASH_BRANCH:odroidhc4/current"`: _only_ build a single board/branch combo
- `TARGETS_FILTER_INCLUDE="BOARD:odroidhc4,BOARD:odroidn2"`: _only_ build _two_ boards, all branches.
- `TARGETS_FILTER_INCLUDE="BOARD_MAINTAINERS:rpardini"`: build all boards and branches where rpardini is a maintainer
- `TARGETS_FILTER_INCLUDE="BOARDFAMILY:rockchip64"`: build all boards and branches in the rockchip64 family
- image-info-only variables like `LINUXFAMILY` is **not** available for filtering at this stage
- rename `config/templates` `targets-all-cli.yaml` to `targets-default.yaml`
- this is used when no `userpatches/targets.yaml` is found
- new default includes all boards vs branches for non-EOS boards
- also desktop for all boards that _don't_ have `HAS_VIDEO_OUTPUT='no``
- introduce simplified `targets-dashboard` CLI:
- does only inventory, compositing, and image info, but not artifact reducing, etc.
- ignore desktop builds in the OpenSearch indexer
- update the OpenSearch Dashboards, including new information now available
- invert the logic used for `CLEAN_INFO` and `CLEAN_MATRIX`
- defaults to `yes` now, so new users/CI don't get hit by stale caches by default
- repo pipeline CLI stuff is usually run on saved/restored artifacts for `output/info`, so don't clean by default via the CLI
- `USE_TMPFS=no` disables usage of generic tmpfs mechanism (still possibly used for rootfs/image building, which is unrelated), for last-resort cases
- use better/more descriptive `temp_dir_id`'s for kernel build than `k` (now `kernel_dest_install_dir`) and `kd` (now `kernel_debs_temp_dir`)
- specific image/dtb/headers packaging already had decent names, same for other .deb's
- replace `mktemp -d` with `mktemp -d --tmpdir "${temp_dir_id}-XXXXX"` in `prepare_temp_dir_in_workdir_and_schedule_cleanup()`, so we know what's using what in tmpfs
The /dev/mapper directory created by devtmpfs lacked entry for
armbian-root thereby breaking the cryptsetup configuration generated
within initrd file. Use bind mount as that doesn't seem to suffer from
that issue.
We were not monitoring /usr/share/initramfs-tools before where most
of extra hooks gets installed. While testing builds with CRYPTROOT
I created build with dropbear ssh key unlock support first and then
went for password only, but it still used initrd image with dropbear
files and older keys. Including /usr/share/initramfs-tools fixes the
same. Also as dropbear keys were autogenerated, they needed to be
monitored as well.
* Added `shell.nix` definition for temporary development environment for Nix(OS)
* Set uuidgen and other binaries check to not rely on hard-codded paths
With this change, setting EXTRAWIFI=no will disable all wireless
patches applied within drivers_network.sh script. Also since #5265
the rtl88x2cs patches were suppose to be not applied to 6.1+ kernel
onwards, instead they were disabled entirely. As this was done by
adding EXTRAWIFI=no, its now replaced with kernel version limit.
Keeping EXTRAWIFI=no there would have made those patches to apply
which would have changed the meaning of the flag.
Cope with the fix in stable 6.3.13 bf353116d1bf and 6.5-rc1 e8c2af660ba0
"wifi: cfg80211: fix regulatory disconnect with OCB/NAN".
That is the removal of REGULATORY_IGNORE_STALE_KICKOFF
from the wireless regulator internal API to fix any driver
that allowed OCB/NAN.
Note this code will need to be expanded once and if 6.4 include the
above fixup.
Signed-off-by: Alban Browaeys <alban.browaeys@gmail.com>
Use multiple consecutive reads in rtw_sdio_read_port() to limit the number of bytes which are copied by the host from the card in one MMC/SDIO transfer. This allows receiving a buffer that's larger than the hosts max_req_size (number of bytes which can be transferred in one MMC/SDIO transfer). As a result of this the skb_over_panic error is gone as the rtw88 driver is now able to receive more than 1536 bytes from the card (either because the incoming packet is larger than that or because multiple packets have been aggregated).
Signed-off-by: Patrick Yavitz <pyavitz@xxxxx.com>
- introduced at 228354ed69
- simply creating the dir solves it
- reported to author, let's hope for an -rc2 fix.
- better logging when DEBUG=yes (don't pass "-s"(ilent) to make clean)
This brings the patch set up to 2023-06-22 wireless-next
Drivers tested;
8822CS. 8821CU and 8723DS
Notables;
Not seeing random dmesg spam `rtw_8822cs failed to get tx report from firmware` which I would see on both the CS and CU from time to time.
Signed-off-by: Patrick Yavitz <pyavitz@gmail.com>
Patch set includes a complete backport of RTW88 as it currently stands, along with extras
Extra:
Patch: RTL8723DS (SDIO)
Patch: _rtw_download_firmware() warn: missing unwind goto?
Hack: SDIO RX Aggregation Limiting
I only have two units available to me with an 8822CS module and in my testing this is only required on the BPI-CM4IO. With out LIMITING the unit will either kernel panic or not be able to send or receive.
This is currently being looked into:
https://lore.kernel.org/linux-wireless/CAFBinCBaXtebixKbjkWKW_WXc5k=NdGNaGUjVE8NCPNxOhsb2g@mail.gmail.com/T/#u
It may be possible to just set LIMITING across all builds, but that to me seems like a poor choice. This requires testing.
Signed-off-by: Patrick Yavitz <pyavitz@gmail.com>
```
# Brief detour. Turns out that HardKernel's vendor odroidxu4 kernel already has this driver
# "slipstreamed" into it, complete with a bunch of PDF files and other junk.
# See https://github.com/hardkernel/linux/tree/odroid-5.4.y/drivers/net/wireless/rtl8812au
# If we remove them here, the resulting patch will contain binary diffs which are unsupported by patch(1).
# So if building for the odroidxu4/current, we'll leave the original files in place, and just overwrite
# the possibly-updated source files (not PDFs and such). Thanks, HardKernel.
```
- even Rich'er patch output by colorizing certain strings green/yellow/red
- BASE_GIT_TAG now very sneakily also accepts a branch name
- IMPORTANT: this includes: BREAKING CHANGE: patches failing to apply now break the build. fixes#4958
- also break on legacy `process_patch_file()` failure, remove `EXIT_PATCHING_ERROR`
- split `image` apt .list deployment into `image-early` (after rootfs extract, before building) and `image-late` (after building)
- rationale: having the repo enabled during the image build can cause 'apt upgrade' and others to do inconsistent stuff depending on the (possibly random) state of the repo
- this way it's consistent across all artifacts and images
- debs-to-repo-download: fix: use DEB_STORAGE instead of hardcoding output/debs (so BETA=yes works)
also:
- fix `Duplicate oci_target` message in mapper-oci-uptodate
also:
- get rid of comments and `SUBREVISION` and `RC` variables
- better debugging in calculate_hash_for_variables() do_normalize_src_path=yes/no
- tag a few places where output/debs might wrong in face of BETA=yes/no
- some hashed variables might contain "${SRC}", so hashes never match, unless built in Docker
- this strips away SRC from all vars and adds debugging so we can detect more later.
- `artifact-uboot`: include more variables into hash, for ATF & rk stuff
- `artifact-armbian-desktop`:
- hashed vars actually contain /armbian in different context, skip normalization in this case
- include results of relevant aggregation into artifact_input_variables
- otherwise: desktops with different appgroups/configs lead to build failures in pipeline
- will cause warnings in JSON preparation step, if more than one appgroups/config combo is targeted, since repo can only have one
- bsp-cli: now depends on `base-files (>= ${REVISION})`, this way upgrading the bsp-cli causes our base-files to be installed
- bsp-cli no longer does gymnastics with /etc/os-release et al, all done in armbian-base-files now
- general/apt-utils.sh: introduce `apt_find_upstream_package_version_and_download_url()`
- base-files: add release to version, in order to comply with repo restrictions (valid repos can't have two different debs with same name and version, md5 must match)
- rationale: we don't want an eternal chicken-egg problem with rootfs vs repo.
- but, desktop rootfs require some parts of repo. case in point: `system-monitoring-center`
- so only add certain components of repo (-desktop, -utils) to rootfs so that is honored
- introduce `custom_apt_repo()` hook for extensions to add their repos as well
- same chicken-egg-avoiding is possible via param `CUSTOM_REPO_WHEN`
- otherwise we get 4 errors for each "error" in CI
- one for the real error
- one for "wait for cleanups"
- one for "Docker build failed"
- one from GHA itself since command failed (we can't get rid of this)
- `odroidxu4`'s firmware hook for vendor kernel. such is life, no use warning.
- `sun50iw9`: vendor kernel with extension, proven working, stop warning
- old kernels don't have working headers, it's a fact of life, remove warning
- `PACKAGE_LIST_BOARD_REMOVE` is used only in `helios64` board
- the fact is `PACKAGE_LIST_BOARD_REMOVE` removes the packages _from the rootfs_ and not _from the image_
- this only accounts for that fact, but in the future we should actually change how this works
- partially revert 6ef394d95d (thus bring back the TUI/dialog for selecting KERNEL_CONFIGURE=yes/no)
- partially revert d890b418c7 (thus bring back the capacity to config & build image in one go)
- stop after configuring kernel, but only if command is `kernel-config`, not regular image-build KERNEL_CONFIGURE=yes
- rootfs is built with BOARD= so just live with it (idea about this warn was that families might hook rootfs build process)
- count, instead of du, apt caches. seems some empty dirs are reported as "1Mb" by du?
- revert #5139 Revert "Apt lists is needed for synpatic index generation"
This reverts commit 63e9bd3baa
- this actually only left junk in rootfs cache for the future to find
