CI: migrate aarch64 Android to Cuttlefish - #5524
Open
10ne1 wants to merge 2 commits into
Open
Conversation
Replace the 2016-era SDK emulator (slow full-system emulation of an Android 7 image, no KVM) with the same Cuttlefish setup x86_64 already uses, so both Android 64-bit targets test against a current official build instead of one target running a decade-old image. GitHub's arm64 runners expose no /dev/kvm, so the device cannot be accelerated by matching the host architecture. It does not have to be: Cuttlefish boots a foreign-architecture guest through qemu, dropping -accel and swapping -cpu host for -cpu max when the target does not match the host (IsHostCompatible() in host/libs/vm_manager/qemu_manager.cpp), so the guest runs under TCG on an ordinary x86_64 runner. Selecting that is the -vm_manager=qemu_cli flag; crosvm, the default, requires host and guest to share an architecture. The device images are arm64 but every host tool running them must be x86_64, and cvd fetch otherwise takes the host package from the target named by --default_build, which would land arm64 binaries the runner cannot execute. --host_package_build pins those to the x86_64 build of the same drop. The image moves from ubuntu:26.04 to debian:13, following the x86_64 one, which upstream switched to keep the base in step with the Cuttlefish packages it installs (6ae84e4). The two are the same setup now, so putting them on different distributions would only invite drift. launch_cvd's timeout grows from 10m to 45m because a TCG guest boots far slower than an accelerated one: measured arm64 boots ranged from 16 to 22 minutes for the same configuration, against roughly one for x86_64. The timeout only bounds a hang. The job's own budget has to grow too: test_tier2 allows 25 minutes, which the boot alone can spend, so the matrix entry raises it to 50 and every other tier2 target keeps the existing 25. A full run lands around 37 minutes. cuttlefish-entrypoint.sh moves to ci/ because it is architecture-neutral and both targets now use it; nothing in it changes. The clang wrapper stays at the retired emulator's runtime API level (24), matching how x86_64-linux-android is pinned, so the version-gated test skips resolve exactly as before: this migration changes the test infrastructure, not the coverage. Raising both is a separate change. Assisted-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Adrian Ratiu <adrian.ratiu@collabora.com>
launch_cvd -daemon prints nothing between "Launcher Build ID" and a booted guest. On x86_64 that gap is about a minute, but the emulated arm64 guest spends around twenty minutes there, so the job looks hung, and a boot that genuinely hangs produces no evidence at all: the timeout fires and the container goes away with the logs inside it. Follow the device's launcher and kernel logs in the background for the duration of the launch. This is also where the logs actually are: cvd keeps the runtime under a per-group directory beneath /var/tmp/cvd rather than under the fetch directory, so they are not simply sitting in $CUTTLEFISH_DIR to be read afterwards. Enabling the serial console (-console=true) would not do this. It attaches an interactive console device, which under -daemon is just a pty in the instance directory; the kernel messages worth seeing are written to kernel.log either way. Assisted-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Adrian Ratiu <adrian.ratiu@collabora.com>
Collaborator
|
Some changes occurred in an Android module cc @maurer |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Hello, this moves the aarch64 Android CI from the decade old legacy emulator (Android 7) to the modern Android 17 matching the existing x86_64 test infra.
Deliberately not included, migrating these is left for future PRs:
arm-linux-androideabineeds more investigation.aosp_cf_arm64_only_phoneis using a 64-bit-only image, so moving it needs a different image variant and its own investigation.Once the above 2 are also migrated, the legacy emulator logic can be removed entirely.
Running aarch64 tests is still pretty slow because it runs via emulation (no KVM on github arm runners), but it is booting a full modern Android and especially unblocks PRs like #4459.
The approaches compared, time-wise:
main): 11m40s - 14m28s. A full system, but an Android 7 / API 24 image from 2016 which is buggy and blocks PRs like Addmalloptand related constants for Android #4459.arm64-v8a: ~36m, of which 14--21m is the boot depending on load and only ~6m is actually running tests.main, for reference):~6m10s. This is roughly what aarch64 would cost if github's arm64 runners exposed
/dev/kvm. The gap is the price of TCG.Credit goes to @maurer for pointing out that Cuttlefish can boot an arm64 guest on an x86_64 host: it drops
-acceland swaps-cpu hostfor-cpu maxwhen the guest arch does not match the host, so-vm_manager=qemu_cliruns the device under TCG.This worked nicely so we don't need an extra layer like android-emulator-runner which unfortunately only works with images up to 2022.
Please see more details in the commit messages.
cc @tgross35