RFR: JDK-8312492: Remove THP sanity checks at VM startup [v2]

[David Holmes]

On Sat, 22 Jul 2023 07:34:09 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

>> When starting the JVM with -XX:+UseTransparentHugePages, we need to know if the OS supports THPs:
>> 
>> A) We could run on a super-old kernel (very unlikely, THPs are supported since 2.6.38 which is EOL since 2011)
>> B) We could run on a kernel built without THP support, e.g. an embedded device
>> C) We could run on a system with THPs disabled by the admin
>> 
>> The JVM does a little sanity test dance at startup: it mmaps an area the size of a huge page - but in small pages - then calls madvise(MADV_HUGEPAGE) on it. The JVM assumes if the madvise() worked, all is fine.
>> 
>> That sanity test is expensive since we allocate (reserve AND commit) an area, then signal khugepaged by advising to fold it into a huge page. That area could be large if the kernel supports 1 GB THP pages. We may end up committing 1 GB space at startup. And depending on the THP defrag settings, that allocation may have to wait for the khugepaged to allocate a 1 GB page. Even if it does not wait but relies on khugepaged to work concurrently its a bad idea.
>> 
>> The sanity test is also of not much use since it does not detect if THPs are disabled on the system (case (C)). Arguably, that is the majority of cases on existing Linux installations.
>> 
>> The correct - and much cheaper - way is to check the proc fs whether THPs are enabled. But we already do that since [JDK-8310233](https://bugs.openjdk.org/browse/JDK-8310233): "Fix THP detection on Linux" [[1]](https://github.com/openjdk/jdk/blob/37ca9024ef59d99cae0bd7e25b2e6d3c1e085f97/src/hotspot/os/linux/os_linux.cpp#L3762). And that test detects all cases (A)(B)(C). Therefore we can completely remove the sanity checks for THPs.
>> 
>> I tested the patch manually in a number of settings with THPs enabled, disabled, and not compiled into the kernel at all. There is no behavioral difference because since [JDK-8310233](https://bugs.openjdk.org/browse/JDK-8310233) we catch all relevant cases before ever entering this sanity test.
>> 
>> [1] https://github.com/openjdk/jdk/blob/37ca9024ef59d99cae0bd7e25b2e6d3c1e085f97/src/hotspot/os/linux/os_linux.cpp#L3762
>
> Thomas Stuefe has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Fix regression in tests

Seems quite reasonable.

Thanks.

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

Marked as reviewed by dholmes (Reviewer).

PR Review: https://git.openjdk.org/jdk/pull/14967#pullrequestreview-1542447144