Re: [PATCH] x86/build: Build VSMP support only if selected

[Juergen Gross <jgross@suse.com>]

On 01/11/2018 16:09, Eial Czerwacki wrote:
> Greetings,
> 
> On 11/01/2018 03:45 PM, Juergen Gross wrote:
>> On 01/11/2018 14:10, Eial Czerwacki wrote:
>>> Greetings,
>>>
>>> On 11/01/2018 12:39 PM, Shai Fultheim (Shai@ScaleMP.com) wrote:
>>>> On 01/11/18 11:37, Thomas Gleixner wrote:
>>>>
>>>>> VSMP support is built even if CONFIG_X86_VSMP is not set. This leads to a build
>>>>> breakage when CONFIG_PCI is disabled as well.
>>>>>
>>>>> Build VSMP code only when selected.
>>>>
>>>> This patch disables detect_vsmp_box() on systems without CONFIG_X86_VSMP, due to
>>>> the recent 6da63eb241a05b0e676d68975e793c0521387141.  This is significant
>>>> regression that will affect significant number of deployments.
>>>>
>>>> We will reply shortly with an updated patch that fix the dependency on pv_irq_ops,
>>>> and revert to CONFIG_PARAVIRT, with proper protection for CONFIG_PCI.
>>>>
>>>
>>> here is the proper patch which fixes the issue on hand:
>>> From ebff534f8cfa55d7c3ab798c44abe879f3fbe2b8 Mon Sep 17 00:00:00 2001
>>> From: Eial Czerwacki <eial@scalemp.com>
>>> Date: Thu, 1 Nov 2018 15:08:32 +0200
>>> Subject: [PATCH] x86/build: Build VSMP support only if CONFIG_PCI is
>>> selected
>>>
>>> vsmp dependency of pv_irq_ops removed some years ago, so now let's clean
>>> it up from vsmp_64.c.
>>>
>>> In short, "cap & ctl & (1 << 4)" was always returning 0, as such we can
>>> remove all the PARAVIRT/PARAVIRT_XXL code handling that.
>>>
>>> However, the rest of the code depends on CONFIG_PCI, so fix it accordingly.
>>>
>>> Signed-off-by: Eial Czerwacki <eial@scalemp.com>
>>> Acked-by: Shai Fultheim <shai@scalemp.com>
>>> ---
>>>  arch/x86/Kconfig          |  1 -
>>>  arch/x86/kernel/vsmp_64.c | 80
>>> +++--------------------------------------------
>>>  2 files changed, 5 insertions(+), 76 deletions(-)
>>>
>>> diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
>>> index c51c989..4b187ca 100644
>>> --- a/arch/x86/Kconfig
>>> +++ b/arch/x86/Kconfig
>>> @@ -524,7 +524,6 @@ config X86_VSMP
>>>         bool "ScaleMP vSMP"
>>>         select HYPERVISOR_GUEST
>>>         select PARAVIRT
>>
>> Do you really still need PARAVIRT and HYPERVISOR_GUEST?
>> Maybe you want IRQ_REMAP instead?
>>
> Better performance is achieved with PARAVIRTed kernel.   Hence we keep
> them both in.

Do you have an explanation for that? Normally PARAVIRT is expected
to have a small negative impact on performance.


Juergen