RFR: 8253683: Clean up and clarify uses of os::vm_allocation_granularity

[Thomas Stuefe]

On Wed, 3 Dec 2025 11:30:32 GMT, Frederic Thevenet <fthevenet at openjdk.org> wrote:

>> Hi everyone,
>> 
>> `os::vm_allocation_granularity()` is meant to describe the alignment restrictions of the operating system when we reserve memory. That is 64 KiB on Windows (`VirtualAlloc`) and 256 MiB on AIX (with `shmat`). On every other platform it happens to match the page size. The page size (available via `os::vm_page_size()`) is what matters when we later commit or protect the reserved pages.
>> 
>> Because the functions are poorly documented and the two numbers are identical on most systems, they have gradually been used more and more interchangeably. We now have many code paths that round **sizes** up to `os::vm_allocation_granularity()` or assert that a size is a multiple of it. That is wrong. Only addresses need that alignment, sizes merely have to be page-aligned. Places that round sizes should instead use `os::vm_page_size()` as they are unrelated to attach alignment.
>> 
>> For this change I have gone over the call sites of `os::vm_allocation_granularity()` and where it was being used to pad or sanity-check a size I have instead replaced it with `os::vm_page_size()`. The calls that genuinely deal with an attach address are left untouched.
>> 
>> Testing:
>> - Oracle tiers 1-8
>
> src/hotspot/os/windows/os_windows.cpp line 2997:
> 
>> 2995:   char * p_buf;
>> 2996:   // note: at setup time we guaranteed that NUMAInterleaveGranularity was aligned up to a page size
>> 2997:   size_t page_size = UseLargePages ? _large_page_size : os::vm_page_size();
> 
> I don't think using `vm_page_size` instead of `vm_allocation_granularity` is appropriate here, as `page_size` is later used to align an address to pass to `virtualAlloc` (`next_alloc_addr`), not just region sizes (see lines 3023-3030)

Hmm, I agree with @fthevenet in that this code looks strange if it were ever called for non numa non largepages. In practice, this is only called for `(UseLargePagesIndividualAllocation || UseNUMAInterleaving)` and maybe we should assert this here, if only as documentation.

For NUMA, we use `NUMAInterleaveGranularity` for chunksize, which is constrained to be at least allocation granularity (see `NUMAInterleaveGranularityConstraintFunc`).

For large pages, these are large 2MB pages. So the value for chunk size will always be >= pagesize since alloc granularity >= page size. And maybe we also should assert that :-)

---

Note that this also could benefit from a rewrite using VirtualAlloc2 and placeholder regions. See https://bugs.openjdk.org/browse/JDK-8376561. Ping @roberttoyonaga - another case of reserve-release-reserve.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/28493#discussion_r2740692794