On Mon, Dec 16, 2019 at 8:33 AM Prarit Bhargava <prarit(a)redhat.com> wrote:
On 12/13/19 11:35 AM, Justin Forbes wrote:
> On Fri, Dec 13, 2019 at 8:14 AM Laura Abbott <labbott(a)redhat.com> wrote:
>>
>> On 12/12/19 9:10 AM, Justin Forbes wrote:
>>> On Thu, Dec 12, 2019 at 3:48 AM Peter Robinson <pbrobinson(a)gmail.com>
wrote:
>>>>
>>>> Hey All.
>>>>
>>>> In digging through some pieces around CPU_IDLE I noticed that
>>>> NO_HZ_IDLE is explicitly disabled on x86_64 but not on all other
>>>> architectures.
>>>>
>>>> Doing a "git log --follow
>>>> configs/fedora/generic/x86/x86_64/CONFIG_NO_HZ_IDLE" it goes all
the
>>>> way back to 2016 when we changed the way the configs were handled.
>>>>
>>>> The upstream kernel's opinion [1] on it is "Most of the time
you want
>>>> to say Y here." so I'm wondering if there's a reason why
we're
>>>> difference on x86_64 or is it just lost in the winds of time?
>>>>
>>>> Peter
>>>>
>>>> PS was digging around CPU_IDLE_GOV_TEO for those curious.
>>>>
>>>> [1]
https://cateee.net/lkddb/web-lkddb/NO_HZ_IDLE.html
>>>
>>>
>>> commit 3836faf6e68495fc70316229a3540506f7ce4c98
>>> Author: Kyle McMartin <kyle(a)fedoraproject.org>
>>> Date: Wed Sep 17 13:10:12 2014 -0500
>>>
>>> re-enable RCU_FAST_NO_HZ, enable NO_HZ_FULL on x86_64
>>>
>>> - I also like to live dangerously. (Re-enable RCU_FAST_NO_HZ which
>>> has been off
>>> since April 2012. Also enable NO_HZ_FULL on x86_64.)
>>
>> Yeah I wouldn't quite say it's been "lost" but the real
question
>> is if it still makes sense. I don't have a strong opinion without
>> data. Prarit, any opinion here?
>
> Oh, I wasn't pointing out that it wasn't just lost, I was pointing out
> that NO_HZ_IDLE is not set because we run NO_HZ_FULL. We were one of
> the first distros to do so, and it has worked well for us. I have a
> fairly strong opinion about not dropping back to IDLE without good
> reason.
Getting back to the original question, I had to go back through my history to
see if I could find a reason why there is a discrepancy between x86 and the
other arches.
AFAICT in *RHEL8* we have NO_HZ_FULL on all arches except s390x. S390x has
NO_HZ_IDLE. Additionally s390 upstream has:
[prarit@prarit linux]$ git grep NO_HZ_IDLE arch/s390/
arch/s390/configs/debug_defconfig:4:CONFIG_NO_HZ_IDLE=y
arch/s390/configs/defconfig:4:CONFIG_NO_HZ_IDLE=y
arch/s390/configs/zfcpdump_defconfig:2:CONFIG_NO_HZ_IDLE=y
On Fedora, as noted,
[prarit@prarit fedora]$ find ./ -name *NO_HZ_IDLE* | xargs grep ^
./generic/x86/x86_64/CONFIG_NO_HZ_IDLE:# CONFIG_NO_HZ_IDLE is not set
./generic/CONFIG_NO_HZ_IDLE:CONFIG_NO_HZ_IDLE=y
[prarit@prarit fedora]$ find ./ -name *NO_HZ_FULL* | xargs grep ^
./generic/x86/x86_64/CONFIG_NO_HZ_FULL:CONFIG_NO_HZ_FULL=y
./generic/CONFIG_NO_HZ_FULL:# CONFIG_NO_HZ_FULL is not set
FWIW I think the correct thing to do for performance reasons is use NO_HZ_FULL
on all arches except s390x which requires NO_HZ_IDLE.
Yes, I do believe this is the correct thing to do, as to how we got
into the current state, when NO_HZ_FULL was introduced, it was x86_64
only. Other architectures came in eventually, but as they were already
set to NO_HZ_IDLE, it didn't prompt us, and to be honest, we were
paying less attention to the other architectures back then. It has
been a while. I will get the changes made in rawhide with tomorrow's
builds.
Justin