Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> writes: > On Wed, Nov 13, 2024 at 3:25 AM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote: >> >> The selftests build four kernel modules which use copy-pasted Makefile >> targets. This is a bit messy, and doesn't scale so well when we add more >> modules, so let's consolidate these rules into a single rule generated >> for each module name, and move the module sources into a single >> directory. >> >> To avoid parallel builds of the different modules stepping on each >> other's toes during the 'modpost' phase of the Kbuild 'make modules', we >> annotate the module copy target as .NOTPARALLEL, which makes all >> its *dependencies* execute sequentially regardless of whether make is >> doing parallel builds or not. This means the recursive make calls into >> the test_kmods directory will be serialised, and when the first one >> actually builds all four modules in the subdirectory, make will >> correctly skip the three other calls, so we end up with just one build >> of the subdir modules. >> >> Acked-by: Viktor Malik <vmalik@xxxxxxxxxx> >> Signed-off-by: Toke Høiland-Jørgensen <toke@xxxxxxxxxx> >> --- >> Changes in v4: >> - Rebase on bpf-next >> - Link to v3: https://lore.kernel.org/r/20241111-bpf-selftests-mod-compile-v3-1-e2e6369ed670@xxxxxxxxxx >> >> Changes in v3: >> - Use .NOTPARALLEL annotation instead of creating a modules.built file > > Is it just me, or did this make everything non-parallel? When I > applied this locally, even .bpf.c compilation and skeleton generation > was sequential despite `make -j$(nproc)`. > > We can't do that, it's too much of a regression. Huh, no, that was certainly not the intention! Will take another look... -Toke
