RFR: 8225035: Thread stack size issue caused by large TLS size

[Thomas Stüfe]

Hi Florian,

On Thu, Jun 27, 2019 at 11:09 AM Florian Weimer <fweimer at redhat.com> wrote:

> * Thomas Stüfe:
>
> > Hi Florian,
> >
> > On Thu, Jun 27, 2019 at 9:57 AM Florian Weimer <fweimer at redhat.com>
> wrote:
> >
> >  * Thomas Stüfe:
> >
> >  >> I think
> >  >>
> >  >>   stack_adjust_size += guard_size;
> >  >>
> >  >> is redundant with the get_minstack-based adjustment because it is
> either
> >  >> included in it directly in the get_minstack value, or indirectly in
> the
> >  >> updated stack size computation in glibc (where it adds the guard size
> >  >> internally).
> >  >
> >  > Please do not change this!
> >  >
> >  > To me and I guess many others it is very important that nothing at all
> >  > changes if this workaround is off.
> >
> >  That has already happened because recent glibc adds the guard size to
> >  the stack size internally now.
> >
> >  But what I meant: only perform the += adjustment above for the
> >  !AdjustStackSizeForTLS case.
> >
> > Yes, this is what I meant with "I'd like to keep this coding
> > simple". The switch supposedly has only to do with TLS, so for the
> > casual code reviewer it makes no sense to bring guard page size into
> > this.
>
> In combination with the glibc stack setup code, __pthread_get_minstack
> always accounts for the guard size.  In some glibc versions, the
> accouting happens in the stack allocation, in other versions, in
> __pthread_get_minstack.  But the net effect is that if you get the magic
> number from __pthread_get_minstack, you do not need to add the guard
> size anymore.  It's either in the __pthread_get_minstack value, or glibc
> no longer considers the guard size as part of the guard size.  In both
> cases, Hotspot should not add the guard size *again*.  At least not for
> an option that says it's about TLS size.
>
> Does this make sense?
>

Lets see if I get this right:

Jianglis switch off (default), glibc < 2.28 (? that version right?):
- on pthread_create, glibc does not add guard page size to stack size
- so we add guard page size to stack size beforehand.
This is the state today, nothing changes with Jianglis patch.

Jianglis switch off (default), glibc >= 2.28:
- on pthread_create, glibc now adds guard page size to stack size ?
- so (continue to) add guard page size to stack size beforehand.
This is the state today, nothing changes with Jianglis patch. We may use a
page more than needed though.

Jianglis switch on, glibc < 2.28:
- on pthread_create, glibc does not add guard page size to stack size
- Jiangli calls pthread_get_minstack which in that version adds guard page
size to returned size.
- so you argue we should not add guard page size to stack size.

Jianglis switch on, glibc >= 2.28:
- on pthread_create, glibc now adds guard page size to stack size
- Jiangli calls pthread_get_minstack which in this version does ignore
guard page size
- same argument, we should not add guard page size to stack size then.

Is that what you mean?

..Thomas


> Thanks,
> Florian
>