[jdk11u-dev] RFR: 8229147: Linux os::create_thread() overcounts guardpage size with newer glibc (>=2.27)

[Thomas Stuefe]

On Mon, 10 Jul 2023 19:47:08 GMT, Lutz Schmidt <lucy at openjdk.org> wrote:

> This backport is for parity with 11.0.21-oracle. 
> Fix didn't apply at all, os/linux/os_linux.cpp was too different in particular. 
> I managed to create a working version by massaging and applying the changes manually. 
> 
> There is some risk involved. It is not caused by the intentional change of behavior but by potential collateral damage. 
> 
> Tests (SAP nightly test suite) do not indicate any issues.
> Reviews pending.

I am unsure about this change. I think in itself it's not that important. Effectively it increases stack size of Non-Java-Threads by one page (Java threads should not use glibc guard pages).

Thing is, JDK-8225035 in itself is a lot more important than this change since it affects stack size for running hotspot embedded in a native app that uses a lot of TLS. And that delta is unpredictable, whereas here we are talking about one page. 

But as @GoeLin writes, it is also more risky. And 11 is long out in the field; all native apps did have time getting used to the current VM behavior.

With downporting JDK-8229147, I don't see why we need the TLS parts of JDK-8225035 at all if all we want to have is "+ 1 page if glibc >= 2.27". You even hardcode AdjustStackSizeForTLS to false. Why not leave it out completely?

-------------

PR Comment: https://git.openjdk.org/jdk11u-dev/pull/2035#issuecomment-1641815191