On Tue, Aug 08, 2017 at 10:06:04AM +0200, Markus Armbruster wrote: > Stefan Hajnoczi <stefa...@gmail.com> writes: > > > On Wed, Jul 26, 2017 at 02:24:02PM -0400, Cleber Rosa wrote: > >> > >> > >> On 07/26/2017 01:58 PM, Stefan Hajnoczi wrote: > >> > On Tue, Jul 25, 2017 at 12:16:13PM -0400, Cleber Rosa wrote: > >> >> On 07/25/2017 11:49 AM, Stefan Hajnoczi wrote: > >> >>> On Fri, Jul 21, 2017 at 10:21:24AM -0400, Cleber Rosa wrote: > >> >>>> On 07/21/2017 10:01 AM, Daniel P. Berrange wrote: > >> >>>>> On Fri, Jul 21, 2017 at 01:33:25PM +0100, Stefan Hajnoczi wrote: > >> >>>>>> On Thu, Jul 20, 2017 at 11:47:27PM -0400, Cleber Rosa wrote: > >> >>>> Without the static capabilities defined, the dynamic check would be > >> >>>> influenced by the run time environment. It would really mean "qemu-io > >> >>>> running on this environment (filesystem?) can do native aio". Again, > >> >>>> that's not the best type of information to depend on when writing > >> >>>> tests. > >> >>> > >> >>> Can you explain this more? > >> >>> > >> >>> It seems logical to me that if qemu-io in this environment cannot do > >> >>> aio=native then we must skip those tests. > >> >>> > >> >>> Stefan > >> >>> > >> >> > >> >> OK, let's abstract a bit more. Let's take this part of your statement: > >> >> > >> >> "if qemu-io in this environment cannot do aio=native" > >> >> > >> >> Let's call that a feature check. Depending on how the *feature check* > >> >> is written, a negative result may hide a test failure, because it would > >> >> now be skipped. > >> > > >> > You are saying a pass->skip transition can hide a failure but ./check > >> > tracks skipped tests. See tests/qemu-iotests/check.log for a > >> > pass/fail/skip history. > >> > > >> > >> You're not focusing on the problem here. The problem is that a test > >> that *was not* supposed to be skipped, would be skipped. > > > > As Daniel Berrange mentioned, ./configure has the same problem. You > > cannot just run it blindly because it silently disables features. > > > > What I'm saying is that in addition to watching ./configure closely, you > > also need to look at the skipped tests that ./check reports. If you do > > that then you can be sure the expected set of tests is passing. > > > >> > It is the job of the CI system to flag up pass/fail/skip transitions. > >> > You're no worse off using feature tests. > >> > > >> > Stefan > >> > > >> > >> What I'm trying to help us achieve here is a reliable and predictable > >> way for the same test job execution to be comparable across > >> environments. From the individual developer workstation, CI, QA etc. > > > > 1. Use ./configure --enable-foo options for all desired features. > > 2. Run the ./check command-line and there should be no unexpected skips > > like this: > > > > 087 [not run] missing aio=native support > > > > To me this seems to address the problem. > > > > I have mentioned the issues with the build flags solution: it creates a > > dependency on the build environment and forces test feature checks to > > duplicate build dependency logic. This is why I think feature tests are > > a cleaner solution. > > I suspect the actual problem here is that the qemu-iotests harness is > not integrated in the build process. For other tests, we specify the > tests to run in a Makefile, and use the same configuration mechanism as > for building stuff conditionally.
The ability to run tests against QEMU binaries without a build environment is useful though. It would still be possible to symlink to external binaries but then the build feature information could be incorrect. Stefan
signature.asc
Description: PGP signature