avoid race for git lockfile when syncing modules in parallel about ocrd_all

[bertsky]

It's again unfortunate that all which GNU make offers here is global breakdown of concurrency via .NOTPARALLEL, where we instead want only local serialization.

For a synchronization mechanism with blocking behaviour on the shell, I can only think of using GNU parallel (which we already install under deps-ubuntu when workflow-configuration is not disabled). This would allow for example:

$(OCRD_MODULES): always-update
	sem --id ocrd_all_git git submodule sync $(GIT_RECURSIVE) $@
	if git submodule status $(GIT_RECURSIVE) $@ | grep -qv '^ '; then \
		sem --id ocrd_all_git git submodule update --init $(GIT_RECURSIVE) $@ && \
		touch $@; fi

@stweil @kba opinions?

[stweil]

IMO pull request #124 is a simply and sufficient solution for this issue.

[bertsky]

IMO pull request #124 is a simply and sufficient solution for this issue.

#124 would sacrifice the independence of builds for different modules, using modules as a single dependency for all executables, and would require downloading all modules even when (re-)installing only a single executable.

The sem solution is simple, too. It's also just a tiny change.

I would have started a PR for it already, but I still have difficulty reliably reproducing the above problem. The race seems to be more related to CPU load than network load. I want to be sure to understand the actual cause before rushing to conclusions.