RFR(xxxs): 8245707: Increase Metaspace reserve alignment

[Thomas Stüfe]

Hi  Colleen,

On Tue, Jun 2, 2020 at 5:53 PM <coleen.phillimore at oracle.com> wrote:

>
> Yes, version 01 looks good.


thanks for the review.


>   How do we tell if
> Metaspace::reserve_alignment() is used instead of
> os::allocation_granularity() accidently?


By tripping off asserts - hopefully - which test for the correct alignment.

This change was motivated by the difference between CDS alignment
(MetaspaceShared::reserve_alignment()) and metaspace alignment. I
originally wanted to do this as part of the remodel change - as a success
check - but Ioi asked be to do this separately.

I checked all current usage of metaspace reserve alignment, all is
currently good. This is more of a future trap for new code using metaspace
reserve alignment. I know its far from perfect, but still better than being
page sized.


> I find the use of
> os::vm_page_size(), os::vm_allocation_granularity() and
> Metaspace::reserve_alignment() seem to appear almost arbitrary in
> places, for example in ReservedSpace::initialize().
>
>
Well, I do not have the sources in front of me
but ReservedSpace::initialize() hopefully does not use metaspace alignment!
that would make no sense.

As for os::vm_page_size() vs os::vm_allocation_granularity(), that is an
ongoing annoyance. As I wrote before, we may want to clean up code which
uses os::vm_allocation_granularity(). It should only be used for memory
attach addresses, I do not think it is even necessary for reserve size.

Thanks, Thomas


> Thanks,
> Coleen
>
> On 5/28/20 2:03 AM, Thomas Stüfe wrote:
> > Thanks Ioi.
> >
> > On Thu, May 28, 2020 at 7:17 AM Ioi Lam <ioi.lam at oracle.com> wrote:
> >
> >> The new version looks fine to me.
> >>
> >> Thanks
> >> - Ioi
> >>
> >> On 5/26/20 10:57 PM, Thomas Stüfe wrote:
> >>
> >> New version:
> >>
> >>
> >>
> http://cr.openjdk.java.net/~stuefe/webrevs/8245707-increase-metaspace-reserve-alignment/webrev.01/webrev/
> >>
> >> Thanks, Thomas
> >>
> >> On Wed, May 27, 2020 at 7:42 AM Thomas Stüfe <thomas.stuefe at gmail.com>
> >> wrote:
> >>
> >>> Hi Ioi,
> >>>
> >>> On Wed, May 27, 2020 at 7:04 AM Ioi Lam <ioi.lam at 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
> platforms
> >>> 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 Stüfe 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-reserve-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
> new
> >>>>> metaspace (JDK-8221173) its reserve alignment requirement will be
> much
> >>>>> 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
> trigger
> >>>> 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
> >>>>
>
>