On Fri, Jan 10, 2014 at 7:31 PM, Christos Zoulas <chris...@astron.com> wrote: > I would have preferred that the cleanup functionality was implemented > differently, running in the context of the shell that ran the test. > This could have been done by issuing a 'trap test_cleanup 0' before > invoking the test, instead of all the complex stuff that is currently > been done (unless I am missing something).
The reason it works like it does is because atf-run takes care of most cleanup by itself and runs every test case on its own subprocess. The cleanup routine supports exists only to let the test deconfigure global resources that it may have affected, like, for example, kernel modules, file systems, routing table, etc. The tests that require this should be a minority, because tests affecting global system state should be minimized (and, thanks to rump, that can be generally avoided these days). Keep in mind that the feature set of the 3 atf libraries must be identical from the point of view of atf-run. trap might have worked on atf-sh, but there is no way you can get proper cleanup like this on C. If the program crashes, the cleanup won't run -- and that's why cleanup is intended to run separately so that it can be executed no matter what happens to the process that ran the test body. Changing the way this works in the atf libraries is not an option. But with the switch to Kyua it wouldn't matter, because Kyua supports test programs built with different "frameworks" (including "no framework"). We'd investigate using other unit-testing libraries at that point that have more traditional fixture definitions, and even use different approaches for sh and C, all within the same NetBSD test suite. I'll be posting an update to the plan in this area soon. -- Julio Merino / @jmmv
