update from mickledore to scarthgap

[beroset]

As described in #245, we updated to mickledore last year, but that is already EOL (as of December 2023). The new long term support version is scarthgap which is expected to be released on 30 April 2024 and is a long term support version with EOL April 2028. See https://wiki.yoctoproject.org/wiki/Releases

[MagneFire]

For what it's worth I've done some work upgrading to scarthgap:

• https://github.com/MagneFire/asteroid/tree/f/scarthgap
• https://github.com/MagneFire/meta-asteroid/tree/f/scarthgap
• https://github.com/MagneFire/meta-smartwatch-1/tree/f/scarthgap
• https://github.com/MagneFire/meta-asteroid-community/tree/f/scarthgap

In summary, a lot is broken:

• dbus seems non-functional
• systemd is failing to launch services
• opening the /dev/binder fails.

I've downgraded systemd to the one we currently use with mickledore. Which fixes dbus related issues, but opening the binder interface still fails resulting in the launcher failing to show anything.

Maybe some of these issues are relevant when upgrading to scarthgap.

Update: Now targeting scarthgap but for beluga which seems to perform even worse. The watch does not boot, it seems to force reboot into QHSUSB__BULK. When using the current kernel the bootlogo is visible for a couple of seconds before rebooting to the mentioned mode. Booting the scarthgap kernel shows a OPPO logo and then it boots to QHSUSB__BULK as well.

[casept]

I've done a bit more work and most watches at least build now. I've tested catfish and it boots, but also has the binder issues and asteroid-launcher does not start.

[MagneFire]

I've finally made it all work. The issue seems to originate with libhybris and compatibility issues with 64bit time_t support (introduced in https://github.com/openembedded/openembedded-core/commit/b9e0c5e750c3097e176fdc18b3b58b622f716e71).

For now I've added a workaround by adding the following to https://github.com/AsteroidOS/meta-asteroid/blob/master/recipes-core/libhybris/libhybris_git.bbappend:

TARGET_CC_ARCH:remove = "-D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 -D_TIME_BITS=64"
INSANE_SKIP:append:pn-${PN} = "32bit-time"

See https://github.com/MagneFire/meta-asteroid/commit/db993a49d25cd23162f5cb8f66f0a9436c6a6fdf

I'm not entirely sure why this is required. My current best guess is that it's some kind of incompatibility with Android Bionic time_t and libc time_t. We should probably investigate this further.

[beroset]

Trying to build this now for catfish. Thanks for the awesome effort!

[beroset]

I had three small problems, all of which I was able to fix. After fixing these problems, I flashed the image to my catfish and it seems to be running just fine.

The problems were:

1. Trouble building the SDK due to cracklib issues
2. problem retrieving desktop-file-utils (apparently they changed from http to https)
3. missing/changed symlink: instead of tmp-glibc/deploy/catfish/asteroid-image-catfish.ext4 it was tmp-glibc/deploy/catfish/asteroid-image-catfish.rootfs.ext4

The attached file scarth.txt can be applied to the scarthgap version of oe-core by going to the oe-core directory and doing git apply scarth.txt.

[MagneFire]

Finally figured out the issue where systemd services would fail on older watches (Linux kernel 3.10). Adding debug traces to systemd revealed that processes seem to crash upon calling umount2().

Performing a git blame revealed a logic change in systemd: https://github.com/systemd/systemd-stable/commit/ea0f3289a288affd5f13c83849b984c8fad63e90

Reverting the logic in mount_switch_root_pivot() seems to workaround this issue. Interestingly the previous code would fail:

Failed to pivot root to new rootfs '/run/systemd/mount-rootfs': Device or resource busy

I've tested this old implementation on catfish (Linux kernel 3.18) as well where it results in the same error. However in case of failure mount_switch_root_move() is called which works in both (3.10 and 3.18) cases. With the new implementation this error does not occur instead resulting in successfully pivoting root on 3.18 while on 3.10 the process crashes at umount2.

I've traced the git tree for differences between the two kernel versions specifically for the umoun2 and namespace related syscalls. Finally the commit https://android.googlesource.com/kernel/msm/+/8033426e6bdb2690d302872ac1e1fadaec1a5581%5E%21/ seems to solve this issue for at least sparrow. I'll test other watches in the coming days.

[MagneFire]

Alright, I've tested the current work on all my watches.

The following worked fine: sturgeon, sparrow, swift, smelt, hoki, beluga, narwhal, dory, tetra, sawfish, pike, ray, lenok While minnow still suffers from a non-working USB and stays on the bootlogo.

I think we're ready for a PR and get this merged.

[MagneFire]

See https://github.com/AsteroidOS/asteroid/pull/292 for the main PR and references to the other relevant PRs.

[beroset]

Merged! Thanks to all for your efforts.