Use hardlinks, bind mount and systemd-tmpfiles to simplify the logic

[VannTen]

Hi folks,

I propose to simplify the logic by relying on bind mount and systemd-tmpfiles instead of cleaning up ourselves.

High-level view of new workflow :

On kernel upgrade / or removal :

• cp the whole /usr/lib/modules/$(kver) tree using hardlinks before the transaction to /usr/lib/modules/backup-$(kver)
• after the transaction, bind mount /usr/lib/modules/backup-$(kver) to /usr/lib/modules/$(kver)
• On next boot, bind mount is gone so : -> we delete recursively dir matching /usr/lib/modules/backup-* -> we delete empty dir matching /usr/lib/modules/* (installed kernel is not emtpy so it will not be deleted, but the mount point for the old kernel is)

That's a pretty big change ; I'm ok forking if you think it does not fit that project.

Thanks.

[isjerryxiao]

Brilliant idea. But there's one little flaw: what if the user would like to revert back to the running kernel without a reboot? The bind mounted old kernel would essentially block the downgrade path.

[VannTen]

Good point.

Maybe it can be solved with another hook which removes the bind mount on upgrade if there is one matching the current kernel.

I'll think about that, a bit busy with life right now.

[VannTen]

Hi o/

So I reworked a bit the hooks:

The pre-hook:

• check if the path exist -> detect whether we install or remove
• only copy if the copy does not exist yet (multiple install/uninstall case)
• checks if there is already a mount only when triggered by "Install" (-> the case where the user reinstall the running kernel after having installed another version)
• Also, we only copy the kernel/ directy and all modules* files, which should supersede #12 (when other scripts have problem with duplicate pkgbase)

WARNING: I haven't tested it yet.

Thoughts ?

[isjerryxiao]

Great job!

[isjerryxiao]

Your script is faulty in several ways fix: https://github.com/archlinux-jerry/pkgbuilds/commit/512e59c17da6b2e35fc053ac4cc45adda27317b3 and https://github.com/archlinux-jerry/pkgbuilds/commit/d49e7113e283159fc3448acca875e4091b4c064e

[VannTen]

It's actually worse than that, because rollback is really problematic. You end up with files present conflicting with files pacman want to place in the fs.

A proper solution should be to use depmod to tell the kernel where the modules are, but I need to look at documenation for that and I've been missing the time.

[isjerryxiao]

I don't see any file conflict.

[isjerryxiao]

should have removed the mount point after umount https://github.com/archlinux-jerry/pkgbuilds/commit/f69c1abf352613767dc458a61b12b46d3b582b59