Re: [RFC] cxl_test: upgrade as a first class citizen selftests capable driver

[Dan Williams <dan.j.williams@intel.com>]

Luis Chamberlain wrote:
> On Fri, Dec 16, 2022 at 08:55:19PM -0800, Dan Williams wrote:
> > In other words the suggestion that the current
> > organization ultimately leads to bit rot has not been substantiated in
> > practice.
> 
> On top of this patch I just added a custom debug patch to my tree which
> enables CXL_BUS and CXL_TEST by default when this is currently allowed
> and it got quite a bit of kernel build warnings. Although some of these
> are specific to my change, some of them do not seem to be related to
> that and likely could benefit from fixing:
> 
> https://gist.github.com/mcgrof/73dce72939590c6edc9413b0384ae4c2
> 
> And so although you may not see some build warnings so far, it does not
> negate my suggestion that having cxl_test as a proper upstream driver strategy
> gets you more build testing / coverage.

If autobuild coverage of test components is the main concern then
cxl_test can copy what nfit_test is doing with CONFIG_NVDIMM_TEST_BUILD.
No need for disruptive redesign of how this facility is integrated.

> > The proposed direction to move tests out of the ndctl.git repo into the
> > kernel solves the wrong problem.
> 
> That's not in any way what I suggested and is not the point to my patch.
> 
> The proposed patch does not suggest to ditch ndctl unit tests but to
> *enable* also sefltests to make use of cxl_test using the selftests
> framework, which is very different. It is not saying "abandon" ndctl
> unit tests, but rather, "also enable linux kernel selftests for CXL with
> cxl_test".

I think centralizing test scripts is a virtue, and right now the
momentum is with those located ndctl.git. This is why I jumped to the
conclusion that the long term direction would be to pick one location
for maintainer regression tests.

> But more importantly, it looks for the value of proper kernel
> integration and making use of kconfig for the actual dependencies
> and requirements. This is of high value.
> 
> In addition to this, one possible area I see of value with this change is the
> ability to also use the wrap feature later, even without cxl_test to enable
> error injection. What would this look like? You simply replace one built in
> routine as you do with another which has sprinkled in should_fail() calls,
> which otherwise would be an eyesore upstream. This shold also then not
> depend the rest of cxl_test stuff, but can make use of building
> alternative wrap routines which could be replacement for upstream ones.
> 
> Another benefit of this strategy is you can also test cxl_test *without*
> the need for for requiring modules, which some folks prefer for testing.
> At LSFMM this came up for instance and one of the biggest grudges with
> testing some frameworks was the dependency on modules.

I do think this is the void that QEMU CXL testing would attempt to fill.