Re: [PATCH] spapr: Fix VSMT mode when it is not supported by the kernel

[Laurent Vivier]

On 20/11/2019 13:47, Greg Kurz wrote:
> On Wed, 20 Nov 2019 12:28:19 +0100
> Laurent Vivier <address@hidden> wrote:
> 
>> On 20/11/2019 10:00, Laurent Vivier wrote:
>>> On 20/11/2019 05:36, David Gibson wrote:
>>>> On Tue, Nov 19, 2019 at 04:45:26PM +0100, Greg Kurz wrote:
>>>>> On Tue, 19 Nov 2019 15:06:51 +0100
>>>>> Laurent Vivier <address@hidden> wrote:
>>>>>
>>>>>> On 19/11/2019 02:00, David Gibson wrote:
>>>>>>> On Fri, Nov 08, 2019 at 05:47:59PM +0100, Greg Kurz wrote:
>>>>>>>> On Fri,  8 Nov 2019 16:40:35 +0100
>>>>>>>> Laurent Vivier <address@hidden> wrote:
>>>>>>>>
>>>>>>>>> Commit 29cb4187497d sets by default the VSMT to smp_threads,
>>>>>>>>> but older kernels (< 4.13) don't support that.
>>>>>>>>>
>>>>>>>>> We can reasonably restore previous behavior with this kernel
>>>>>>>>> to allow to run QEMU as before.
>>>>>>>>>
>>>>>>>>> If VSMT is not supported, VSMT will be set to MAX(8, smp_threads)
>>>>>>>>> as it is done for previous machine types (< pseries-4.2)
>>>>>>>>>
>>>>>>>>
>>>>>>>> It is usually _bad_ to base the machine behavior on host capabilities.
>>>>>>>> What happens if we migrate between an older kernel and a recent one ?
>>>>>>>
>>>>>>> Right.  We're really trying to remove instaces of such behaviour.  I'd
>>>>>>> prefer to completely revert Greg's original patch than to re-introduce
>>>>>>> host configuration dependency into the guest configuration..
>>>>>>>
>>>>>>>> I understand this is to fix tests/migration-test on older kernels.
>>>>>>>> Couldn't this be achieved with migration-test doing some introspection
>>>>>>>> and maybe pass vsmt=8 on the QEMU command line ?
>>>>>>>
>>>>>>> ..adjusting the test case like this might be a better idea, though.
>>>>>>>
>>>>>>> What's the test setup where we're using the old kernel?  I really only
>>>>>>> applied the original patch on the guess that we didn't really care
>>>>>>> about kernels that old.  The fact you've hit this in practice makes me
>>>>>>> doubt that assumption.
>>>>>>>
>>>>>>
>>>>>> The way to fix the tests is to add "-smp threads=8" on the command line
>>>>>> (for all tests, so basically in qtest_init_without_qmp_handshake(), and
>>>>>> it will impact all the machine types), and we have to check if it is
>>>>>
>>>>> Ohhh... it isn't possible to initialize Qtest with machine specific
>>>>> properties ? That's a bit unfortunate :-\
>>>>
>>>> Uhh... I don't see why we can't.  Couldn't we just put either -machine
>>>> vsmt=8 or -smp 8 into the cmd_src / cmd_dst printfs() in the
>>>> strcmp(arch, "ppc64") case?
>>>
>>> Yes, but we need to do that to all other tests that fail. test-migration
>>> is not the only one impacted by the problem (we have also pxe-test), so
>>> it's why I thought to fix the problem in a generic place.
>>>
>>> But it seems there are only this couple of tests that are impacted so I
>>> can modify both instead. I think only tests that really start CPU have
>>> the problem.
>>>
>>> I'm going to send a patch to fix that.
>>
>> And again, it's a little bit more complicated than expected: setting
>> vsmt to 8 works only with kvm_hv, but breaks in case of TCG or kvm_pr.
>> So the test must check what is in use...
>>
> 
> AFAICT, migration-test explicitly skip tests if kvm_hv isn't present.
> 
>     /*
>      * On ppc64, the test only works with kvm-hv, but not with kvm-pr and TCG
>      * is touchy due to race conditions on dirty bits (especially on PPC for
>      * some reason)
>      */
>     if (g_str_equal(qtest_get_arch(), "ppc64") &&
>         access("/sys/module/kvm_hv", F_OK)) {
>         g_test_message("Skipping test: kvm_hv not available");
>         return g_test_run();
>     }
> 
> and I don't see any error in pxe-test if I force tcg and vsmt=8.
> 
> What error do you see with your testing ?

In fact, you're right, it works with vsmt=8 and it's better.

Laurent