s390: select ARCH_WANT_HUGETLB_PAGE_OPTIMIZE_VMEMMAP

Enable HUGETLB_PAGE_OPTIMIZE_VMEMMAP for s390.

With this, vmemmap pages used to back struct pages for compound tail
pages of hugetlb pages are freed and remapped to compound head page
frame as RO, see also Documentation/vm/vmemmap_dedup.rst.

For 1M hugetlb pages, this results in freeing 3 of 4 vmemmap pages,
saving 12K of memory for each 1M hugetlb page (~1.2%).
/sys/kernel/debug/kernel_page_tables will show the impact:

---[ vmemmap Area Start ]---
[...]
0x0000037202d84000-0x0000037202d85000         4K PTE RW NX
0x0000037202d85000-0x0000037202d88000        12K PTE RO NX

For 2G hugetlb pages, this results in freeing 8191 of 8192 vmemmap
pages, saving 32764K of memory for each 2G hugetlb page (~1.6%)
/sys/kernel/debug/kernel_page_tables will show the impact:

---[ vmemmap Area Start ]---
[...]
0x000003720a000000-0x000003720a001000         4K PTE RW NX
0x000003720a001000-0x000003720c000000     32764K PTE RO NX

The memory savings come with some costs:
- vmemmap mapping for compound hugetlb pages is not a PMD mapping any
  more, but split to 4K PTE mappings, and it will not be coalesced back
  to PMD mapping after freeing hugetlb pages from the pool.
  Apart from theoretical performance impact, this will also (slightly)
  relativize the memory savings because of additional 2K PTE pagetable
  allocations.
- Workload using "on the fly" hugetlb allocations via
  "nr_overcommit_hugepages" instead of using the hugetlb pool via
  "nr_hugepages" will suffer from considerably increased fault handling
  time, see also description from commit 78f39084
  ("mm: hugetlb_vmemmap: add hugetlb_optimize_vmemmap sysctl").
- Freeing hugetlb pages from the pool will require re-allocation of the
  freed struct pages, and therefore needs some memory available to the
  kernel. This might fail in memory constrained scenarios.
- For the same reason, memory offline might fail even for ZONE_MOVABLE
  when hugetlb pages are present (but not for s390, since we do not
  support ARCH_ENABLE_HUGEPAGE_MIGRATION, and therefore cannot have
  hugetlb pages in ZONE_MOVABLE).
- General increased complexity and overhead in kernel handling of
  compound (head) pages.

Therefore, this feature is disabled by default, and has to be enabled
explicitly either by adding "hugetlb_free_vmemmap=on" kernel parameter,
or during run-time via "/proc/sys/vm/hugetlb_optimize_vmemmap" sysctl.
Acked-by: Heiko Carstens <hca@linux.ibm.com>
Signed-off-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>

arch/s390/Kconfig

@@ -121,6 +121,7 @@ config S390
 	select ARCH_WANTS_NO_INSTR
 	select ARCH_WANT_DEFAULT_BPF_JIT
 	select ARCH_WANT_IPC_PARSE_VERSION
+	select ARCH_WANT_HUGETLB_PAGE_OPTIMIZE_VMEMMAP
 	select BUILDTIME_TABLE_SORT
 	select CLONE_BACKWARDS2
 	select DMA_OPS if PCI