RFR: 8280056: gtest/LargePageGtests.java#use-large-pages failed "os.release_one_mapping_multi_commits_vm" [v4]

[Thomas Stuefe]

On Mon, 20 Nov 2023 16:59:21 GMT, Gerard Ziemski <gziemski at openjdk.org> wrote:

> Looking at this I am starting to wonder whether we really need this test at all.
> 
> We seem to be testing whether the underlaying OS does what we want it, as opposed to our VM code, which in this case is just a thin convenience layer.
> 
> I can see this test having been useful when we were first implementing/using these features, but what would we need to get wrong nowadays to cause a regression in this test for example, short of someone doing a complete rewrite and getting something wrong?

I wrote this originally (I believe) because of a very real issue where, on Windows, we would allocate memory in stripes for different NUMA nodes, and then later release the memory. IIRC, this just released the first stripe, so the rest stayed allocated (os::release_memory(p, size) called VirtualFree(MEM_RELEASE, p), ignoring the size parameter). 

Fixing that was not trivial - we changed os::release_memory to, if called over multiple reservations, release them all. It would query all existing reservations in that area with VirtualQuery. That stuff is intricate and it can fail, so a regression test makes actually a lot of sense.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/16240#discussion_r1399563118