Re: [RFC 1/1] acceptance tests: rename acceptance to system

[Philippe Mathieu-Daudé]

On 5/21/21 3:03 PM, Alex Bennée wrote:
> Philippe Mathieu-Daudé <philmd@redhat.com> writes:
>> On 5/21/21 2:28 PM, Willian Rampazzo wrote:
>>> On Fri, May 21, 2021 at 4:16 AM Thomas Huth <thuth@redhat.com> wrote:
>>>>
>>>> On 20/05/2021 22.28, Philippe Mathieu-Daudé wrote:
>>>>> On 5/20/21 9:53 PM, Willian Rampazzo wrote:
>>>>>> Conceptually speaking, acceptance tests "are a series of specific tests
>>>>>> conducted by the customer in an attempt to uncover product errors before
>>>>>> accepting the software from the developer. Conducted by the end-user 
>>>>>> rather
>>>>>> than software engineers, acceptance testing can range from an informal
>>>>>> “test drive” to a planned and systematically executed series of scripted
>>>>>> tests" [1]. Every time Pressman refers to the term "acceptance testing," 
>>>>>> he
>>>>>> also refers to user's agreement in the final state of an implemented 
>>>>>> feature.
>>>>>> Today, QEMU is not implementing user acceptance tests as described by 
>>>>>> Pressman.
>>>>>>
>>>>>> There are other three possible terms we could use to describe what is 
>>>>>> currently
>>>>>> QEMU "acceptance" tests:
>>>>>>
>>>>>>    1 - Integration tests:
>>>>>>        - "Integration testing is a systematic technique for constructing 
>>>>>> the
>>>>>>           software architecture while at the same time conducting tests 
>>>>>> to
>>>>>>           uncover errors associated with interfacing. The objective is 
>>>>>> to take
>>>>>>           unit-tested components and build a program structure that has 
>>>>>> been
>>>>>>           dictated by design." [2]
>>>>>>        * Note: Sommerville does not have a clear definition of 
>>>>>> integration
>>>>>>          testing. He refers to incremental integration of components 
>>>>>> inside
>>>>>>          the system testing (see [3]).
>>>>
>>>> After thinking about this for a while, I agree with you that renaming the
>>>> "acceptance" tests to "integration" tests is also not a good idea. When I
>>>> hear "integration" test in the context of the virt stack, I'd rather expect
>>>> a test suite that picks KVM (i.e. a kernel), QEMU, libvirt and maybe
>>>> virt-manager on top and tests them all together. So we should look for a
>>>> different name indeed.
>>>>
>>>>>>    2 - Validation tests:
>>>>>>        - "Validation testing begins at the culmination of integration 
>>>>>> testing,
>>>>>>           when individual components have been exercised, the software is
>>>>>>           completely assembled as a package, and interfacing errors have 
>>>>>> been
>>>>>>           uncovered and corrected. At the validation or system level, the
>>>>>>           distinction between different software categories disappears. 
>>>>>> Testing
>>>>>>           focuses on user-visible actions and user-recognizable output 
>>>>>> from the
>>>>>>           system." [4]
>>>>>>        - "where you expect the system to perform correctly using a set 
>>>>>> of test
>>>>>>           cases that reflect the system’s expected use." [5]
>>>>>>        * Note: the definition of "validation testing" from Sommerville 
>>>>>> reflects
>>>>>>          the same definition found around the Internet, as one of the 
>>>>>> processes
>>>>>>          inside the "Verification & Validation (V&V)." In this concept,
>>>>>>          validation testing is a high-level definition that covers unit 
>>>>>> testing,
>>>>>>          functional testing, integration testing, system testing, and 
>>>>>> acceptance
>>>>>>          testing.
>>>>>>
>>>>>>    3 - System tests:
>>>>>>        - "verifies that all elements mesh properly and that overall 
>>>>>> system
>>>>>>           function and performance is achieved." [6]
>>>>>>        - "involves integrating components to create a version of the 
>>>>>> system and
>>>>>>           then testing the integrated system. System testing checks that
>>>>>>           components are compatible, interact correctly, and transfer 
>>>>>> the right
>>>>>>           data at the right time across their interfaces." [7]
>>>>>>
>>>>>> The tests implemented inside the QEMU "acceptance" directory depend on 
>>>>>> the
>>>>>> software completely assembled and, sometimes, on other elements, like 
>>>>>> operating
>>>>>> system images. In this case, the proposal here is to rename the current
>>>>>> "acceptance" directory to "system."
>>>>>
>>>>> Are user-mode tests using Avocado also system tests?
>>>>> https://www.mail-archive.com/qemu-devel@nongnu.org/msg782505.html
>>>>
>>>> We've indeed got the problem that the word "system" is a little bit
>>>> overloaded in the context of QEMU. We often talk about "system" when
>>>> referring to the qemu-softmmu-xxx emulators (in contrast to the linux-user
>>>> emulator binaries). For example, the "--disable-system" switch of the
>>>> configure script, or the "build-system" and "check-system" jobs in the
>>>> .gitlab-ci.yml file ... thus this could get quite confusing in the
>>>> .gitlab-ci.yml file afterwards.
>>>
>>> I agree with you here. After I made the changes to the code, I noticed
>>> QEMU has the "system" word spread all over the place. That may confuse
>>> people looking at the "system tests" without much interaction with
>>> software testing terminology.
>>>
>>>>
>>>> So I think renaming "acceptance" to "system" is especially ok if we only
>>>> keep the "softmmu"-related tests in that folder... would it maybe make 
>>>> sense
>>>> to add the linux-user related tests in a separate folder called tests/user/
>>>> instead, Philippe? And we should likely rename the current build-system and
>>>> check-system jobs in our gitlab-CI to build-softmmu and check-softmmu or 
>>>> so?
>>>>
>>>
>>> As I mentioned in Philippe's reply, those tests are still considered
>>> system tests because system testing is the software built and
>>> interacting with external test artifacts in software engineering.
>>>
>>>> Alternatively, what about renaming the "acceptance" tests to "validation"
>>>> instead? That word does not have a duplicated definition in the context of
>>>> QEMU yet, so I think it would be less confusing.
>>>
>>> While at the beginning of your reply, I started thinking if
>>> "validation" would cause less confusion for the QEMU project. Although
>>> validation testing is a broader concept inside the Verification &
>>> Validation process, encompassing unit testing, functional testing,
>>> integration testing, system testing, and acceptance testing, it may be
>>> an option for the QEMU project.
>>>
>>> While system testing would be the correct terminology to use, if it
>>> causes more confusion, using a less strict terminology, like
>>> validation testing, is valid, in my opinion.
>>
>> This works for me:
>>
>> - tests/system/softmmu
>> - tests/system/user
>>
>> Or validation, as you prefer.
> 
> So what are tests/tcg if not user tests? They *mostly* test
> linux-user emulation but of course we have softmmu tests in there as
> well. 

I expect a tests/tcg/ to check a specific TCG feature, which doesn't
have to be user-mode specific (IIRC Xtensa does some sysemu checks).
Also, you control the compiler toolchain, flags, etc... so you can
adapt for a specific feature bit to test, use kludges and so on.

I expect tests in tests/system/ (user/softmmu) to user real-world
binaries, which we aren't modifying. Sometime non-public/released
compiler toolchain has been used.

See for example the test referred tests the bFLT loader (beside
testing userland Linux binary for Cortex-M).

Another example is the Sony PlayStation2 binary testing the
O32 ABI and multiple opcodes from the TX79 SIMD core:
https://www.mail-archive.com/qemu-devel@nongnu.org/msg782493.html

Personally I'm not interested in writing a test for a loader or
multiple opcodes when we have pre-built binaries. For the opcodes
coverage I'd use a TCG plugin to confirm the opcodes have been
used.

If you think these tests belong to tests/tcg/, I am OK to put
them they, but I don't think adding the Avocado buildsys
machinery to the already-complex tests/tcg/ Makefiles is going
to help us...

Regards,

Phil.