[prev in list] [next in list] [prev in thread] [next in thread] 

List:       openjdk-hotspot-runtime-dev
Subject:    Re: RFR(xxxs): 8245707: Increase Metaspace reserve alignment
From:       Thomas_Stüfe <thomas.stuefe () gmail ! com>
Date:       2020-05-28 6:03:49
Message-ID: CAA-vtUwBMaAWS+We6LRYm5xV0jFHMwF5f0OOWdYAKJpaEbi6WQ () mail ! gmail ! com
[Download RAW message or body]

Thanks Ioi.

On Thu, May 28, 2020 at 7:17 AM Ioi Lam <ioi.lam@oracle.com> wrote:

> The new version looks fine to me.
>
> Thanks
> - Ioi
>
> On 5/26/20 10:57 PM, Thomas St=C3=BCfe wrote:
>
> New version:
>
>
> http://cr.openjdk.java.net/~stuefe/webrevs/8245707-increase-metaspace-res=
erve-alignment/webrev.01/webrev/
>
> Thanks, Thomas
>
> On Wed, May 27, 2020 at 7:42 AM Thomas St=C3=BCfe <thomas.stuefe@gmail.co=
m>
> wrote:
>
>> Hi Ioi,
>>
>> On Wed, May 27, 2020 at 7:04 AM Ioi Lam <ioi.lam@oracle.com> wrote:
>>
>>> Hi Thomas,
>>>
>>> If (UseLargePages && UseLargePagesInMetaspace) is true, then
>>> _reserve_alignment will be 4*os::large_page_size(). Is this what you
>>> want?
>>>
>>>
>> No :) That is why I limited this to cases where the page size is 4K.
>>
>> However I see that on Windows this would still land us at 4*allocation
>> granularity instead, so 256K. I'll change that.
>>
>>
>>> Or, do you want to do the *4 only when large page is not used?
>>>
>>>
>> Only if page size is 4K. The reason is that I want to trip over
>> anyone asserting and calculating alignment (eg. CDS) but not trigger any
>> tests which do test things like memory counters - on some of our platfor=
ms
>> we have 64K and I don't want to adjust a bunch of tests.
>>
>>
>>> Why did you choose *4? Would *2 also accomplish what you want?
>>>
>>>
>> *2 would also be acceptable. I'll post an update.
>>
>> Thanks! ..Thomas
>>
>>
>>> Thanks
>>> - Ioi
>>>
>>> On 5/25/20 10:57 PM, Thomas St=C3=BCfe wrote:
>>> > Hi all,
>>> >
>>> > may I have reviews for this tiny change:
>>> >
>>> > JBS: https://bugs.openjdk.java.net/browse/JDK-8245707
>>> > webrev:
>>> >
>>> http://cr.openjdk.java.net/~stuefe/webrevs/8245707-increase-metaspace-r=
eserve-alignment/webrev.00/webrev/
>>> >
>>> > One of the motivations behind JDK-8243392 was to decouple CDS
>>> reservation
>>> > alignment from metaspace alignment and prevent misuse of metaspace
>>> reserve
>>> > alignment outside metaspace. This is needed since with the upcoming n=
ew
>>> > metaspace (JDK-8221173) its reserve alignment requirement will be muc=
h
>>> > higher than CDS alignment.
>>> >
>>> > For the time being, until JDK-8221173 finds its way into jdk mainline=
,
>>> we
>>> > should increase metaspace reserve alignment in jdk mainline to trigge=
r
>>> any
>>> > remaining illegal uses of metaspace reserve alignments and to prevent
>>> new
>>> > ones from creeping in.
>>> >
>>> > All CI tests at SAP went through successfully, so at the moment we
>>> seem to
>>> > be clean.
>>> >
>>> > Thanks, Thomas
>>>
>>>
>
[prev in list] [next in list] [prev in thread] [next in thread]