Re: [Qemu-devel] [PATCH 14/18] Boot Linux Console Test: add a test for ppc64 + pseries

["Philippe Mathieu-Daudé" <philmd@redhat.com>]

On 1/31/19 3:37 AM, Cleber Rosa wrote:
> On 1/22/19 11:07 AM, Alex Bennée wrote:
>> Cleber Rosa <crosa@redhat.com> writes:
>>
>>> Just like the previous tests, boots a Linux kernel on a ppc64 target
>>> using the pseries machine.
>>
>> So running this on my rather slow SynQuacer I get:
>>
>>  (04/16) /home/alex/lsrc/qemu.git/tests/acceptance/boot_linux_console.py:BootLinuxConsole.test_ppc64_pseries: INTERRUPTED: Test reported status but did not finish\nRunner error occurred: Timeout reached\nOriginal status: ERROR\n{'name': '04-/home/alex/lsrc/qemu.git/tests/acceptance/boot_linux_console.py:BootLinuxConsole.test_ppc64_pseries', 'logdir': '/home/alex/lsrc/qemu.git/te... (60.93 s)
>>
>> which I'm guessing is a timeout occurring.
>>
> 
> Yes, that's what's happening.  It's hard to pinpoint, and control, the
> sluggishness points in such a test running on a different environment.
> For this one execution, I do trust your assessment, and it's most
> certainly caused by your "slow SynQuacer", spending too much running
> emulation code.
> 
> But, I'd like to mention that there are other possibilities.  One is
> that you're hitting a "asset fetcher bug" that I recently fixed in
> Avocado[1] (fix to be available on 68.0, to be released next Monday, Feb
> 4th).
> 
> Even with that bug fixed, I feel like it's unfair to test code to spend
> its time waiting to download a file when it's not testing *the file
> download itself*.  Because of that, there are plans to add an (optional)
> job pre-processing step that will make sure the needed assets are in the
> cache ahead of time[2][3].
> 
>> I wonder if that means we should:
>>
>>   a) set timeouts longer for when running on TCG

I hit the same problem with VM tests, and suggested a poor "increase
timeout" patch [1].

Then Peter sent a different patch [2] which happens to inadvertently
Dictionary resolve my problem, since the longer a VM took to boot on the
Cavium ThunderX I have access is 288 seconds, which is closely below the
300 seconds limit =) I understood nobody seemed to really care about
testing the x86 TCG backend this way, so I didn't worry much.

[1] http://lists.nongnu.org/archive/html/qemu-devel/2018-07/msg03416.html
[2] http://lists.nongnu.org/archive/html/qemu-devel/2018-08/msg04977.html

>>   or
>>   b) split tests into TCG and KVM tests and select KVM tests on appropriate HW
>>
> 
> I wonder the same, and I believe this falls into a similar scenario
> we've seen with the setup of console devices in the QEMUMachine class.
> I've started by setting the device types defined at the framework level,
> and then reverted to the machine's default devices (using '-serial'),
> because the "default" behavior of QEMU is usually what a test writer
> wants when not setting something else explicitly.
> 
>> The qemu.py code has (slightly flawed) logic for detecting KVM and
>> passing --enable-kvm. Maybe we should be doing that here?
>>
> 
> I'm not sure.  IMO, the common question is: should we magically (at a
> framework level) configure tests based on probed host environment
> characteristics?  I feel like we should attempt to minimize that for the
> sake of tests being more obvious and more easily reproducible.

I agree we shouldn't randomly test different features, but rather
explicitly add 2 tests (TCG/KVM), and if it is not possible to run a
test, mark it as SKIPPED.

An user with KVM available would then have to run --filter-out=tcg, or
build QEMU with --disable-tcg.

> And because of that, I'd go, *initially*, with an approach more similar
> to your option "b".
> 
> Having said that, we don't want to rewrite most tests just to be able to
> test with either KVM or TCG, if the tests are not explicitly testing KVM
> or TCG.  At this point, using KVM or TCG is test/framework
> *configuration*, and in Avocado we hope to solve this by having the
> executed tests easily identifiable and reproducible (a test ID will
> contain a information about the options passed, and a replay of the job
> will apply the same configuration).
> 
> For now, I think the best approach is to increase the timeout, because I
> think it's much worse to have to deal with false negatives (longer
> execution times that don't really mean a failure), than having a test
> possibly taking some more time to finish.
> 
> And sorry for extremely the long answer!
> - Cleber.
> 
> [1] - https://github.com/avocado-framework/avocado/pull/2996
> [2] -
> https://trello.com/c/WPd4FrIy/1479-add-support-to-specify-assets-in-test-docstring
> [3] - https://trello.com/c/CKP7YS6G/1481-on-cache-check-for-asset-fetcher
>