PS-10595 [trunk]: Add innodb_large_page_populate to control huge page pre-population

[polchawa-percona]

https://perconadev.atlassian.net/browse/PS-10595

This patch is a contribution from Alexey Bychko abychko@gmail.com.

Introduce a new InnoDB boolean system variable innodb_large_page_populate
(Linux only) that controls whether huge pages and page-aligned allocations
used by the buffer pool and the page allocator are pre-populated at
allocation time.

When OFF (default), large_page_aligned_alloc() and page_aligned_alloc()
skip MAP_POPULATE and the explicit prefault step by masking the caller's
populate request with the global flag. This reduces startup time and
NUMA-related overhead on large systems with huge pages. When ON, the
previous behavior (honor the caller's populate request) is preserved.

The backing global is named srv_large_page_populate, following the srv_
naming convention used by other InnoDB globals. On non-Linux platforms the
flag is a compile-time constant true, so behavior there is unchanged.

Buffer pool memory is only reserved at buf_pool_init() time; actual
physical allocation happens on first page fault. Huge pages are never
migrated by the kernel's NUMA balancer, and the NUMA interleave memory
policy is only a hint honored at fault time: if a node is briefly out of
free memory when a page is first touched, the page lands on a different
node and stays there permanently, skewing the buffer pool's NUMA placement
for the life of the server. To guard against this, a startup warning is
emitted when innodb_numa_interleave and large_pages are both enabled
while innodb_large_page_populate is OFF, suggesting to drop the page
cache before starting mysqld and to enable innodb_large_page_populate
in that configuration so pages are faulted in immediately under the
interleave policy.