ADR-0018: Every make of a build runs with --no-print-directory

Status: accepted Date: 2026-09-21

Context

The first golden build of Home Assistant OS failed twice, in the same package and at the same line: rtl8821cu, a kernel module, stopped with Makefile:243: *** multiple target patterns. Stop. Nothing else failed, so it was not a race.

Buildroot's kernel-module infrastructure passes the kernel release to the module's Makefile as KVER= followed by a backtick command, $(MAKE) ... -C <linux> --no-print-directory -s kernelrelease, whose output the shell substitutes into the command line. The make that runs kernelrelease inherits MAKEFLAGS, and a sub-make gets w (print directory) there whenever its parent prints directories. GNU make 4.3 then prints make: Entering directory ... and make: Leaving directory ... to standard output, even with -s and even with --no-print-directory on that command line. KVER therefore became three lines, and the words with a colon (make[1]:) reached the kernel's Makefile as goals, which it reads as target patterns. Reproduced with a two-line Makefile in the worker image.

Upstream builds do not hit this because they run the top-level make in a way that leaves w out of MAKEFLAGS. Both of ours printed "Entering directory" (the golden script and the Buck2 action), so both would fail, on any project with such a module.

Decision

The golden script and the package action (br2/pkg_action.py, MAKE) pass --no-print-directory to the top-level make. Sub-makes inherit it through MAKEFLAGS, so none of them gets w. It changes only log output, not any file the build produces.

Consequences

• Easier: kernel-module packages that capture kernelrelease build, in the golden and in an action.
• Harder: MAKE is an action input, so every action key changes once (the caches of earlier runs are not reused). The manifests do not change.
• Not solved: any other package that captures the output of a sub-make would be affected by the same class of stray output; this is the only one seen so far.