On Mon, Dec 16, 2019 at 02:05:49PM -0800, Brendan Higgins wrote:
This patchset adds a centralized executor to dispatch tests rather than relying on late_initcall to schedule each test suite separately along with a couple of new features that depend on it.
## What am I trying to do?
Conceptually, I am trying to provide a mechanism by which test suites can be grouped together so that they can be reasoned about collectively. The last two patches in this series add features which depend on this:
RFC 5/6 Prints out a test plan right before KUnit tests are run; this is valuable because it makes it possible for a test harness to detect whether the number of tests run matches the number of tests expected to be run, ensuring that no tests silently failed.
RFC 6/6 Add a new kernel command-line option which allows the user to specify that the kernel poweroff, halt, or reboot after completing all KUnit tests; this is very handy for running KUnit tests on UML or a VM so that the UML/VM process exits cleanly immediately after running all tests without needing a special initramfs.
The approach seems sensible to me given that it separates from a semantics perspective kernel subsystem init work from *testing*, and so we are sure we'd run the *test* stuff *after* all subsystem init stuff.
Dispatching, however is still immediate, and with a bit of work, this dispatcher could be configurable to run at an arbirary time after boot. If there are not immediate use cases for that though, then I suppose this is not a requirement for the dispatcher. But since there exists another modular test framework with its own dispatcher and it seems the goal is to merge the work long term, this might preempt the requirement to define how and when we can dispatch tests post boot.
And, if we're going to do that, I can suggest that a data structure instead of just a function init call be used to describe tests to be placed into an ELF section. With my linker table work this would be easy, I define section ranges for code describing only executable routines, but it defines linker tables for when a component in the kernel would define a data structure, part of which can be a callback. Such data structure stuffed into an ELF section could allow dynamic configuration of the dipsatching, even post boot.
I think this is a good stepping stone forward then, and to allow dynamic configuration of the dispatcher could mean eventual extensions to kunit's init stuff to stuff init calls into a data structure which can then allow configuration of the dispatching. One benefit that the linker table work *may* be able to help here with is that it allows an easy way to create kunit specific ordering, at linker time. There is also an example of addressing / generalizing dynamic / run time changes of ordering, by using the x86 IOMMU initialization as an example case. We don't have an easy way to do this today, but if kunit could benefit from such framework, it'd be another use case for the linker table work. That is, the ability to easilly allow dynamically modifying run time ordering of code through ELF sections.
In addition, by dispatching tests from a single location, we can guarantee that all KUnit tests run after late_init is complete, which was a concern during the initial KUnit patchset review (this has not been a problem in practice, but resolving with certainty is nevertheless desirable).
Indeed, the concern is just a real semantics limitations. With the tests *always* running after all subsystem init stuff, we know we'd have a real full kernel ready.
It does beg the question if this means kunit is happy to not be a tool to test pre basic setup stuff (terminology used in init.c, meaning prior to running all init levels). I suspect this is the case.
Other use cases for this exist, but the above features should provide an idea of the value that this could provide.
## What work remains to be done?
These patches were based on patches in our non-upstream branch, so we have a pretty good idea that they are useable as presented; nevertheless, some of the changes done in this patchset could *definitely* use some review by subsystem experts (linker scripts, init, etc), and will likely change a lot after getting feedback.
The biggest thing that I know will require additional attention is integrating this patchset with the KUnit module support patchset. I have not even attempted to build these patches on top of the module support patches as I would like to get people's initial thoughts first (especially Alan's :-) ). I think that making these patches work with module support should be fairly straight forward, nevertheless.
Modules just have their own sections too. That's all. So it'd be a matter of extending the linker script for modules too. But a module's init is different than the core kernel's for vmlinux.