If user enabled +UseTransparentHugePages, we do a sanity check to check support for THP:

```
bool os::Linux::transparent_huge_pages_sanity_check(bool warn,
                                                    size_t page_size) {
  bool result = false;
  void *p = mmap(nullptr, page_size * 2, PROT_READ|PROT_WRITE,
                 MAP_ANONYMOUS|MAP_PRIVATE,
                 -1, 0);
  if (p != MAP_FAILED) {
    void *aligned_p = align_up(p, page_size);

    result = madvise(aligned_p, page_size, MADV_HUGEPAGE) == 0;

    munmap(p, page_size * 2);
  }

  if (warn && !result) {
    warning("TransparentHugePages is not supported by the operating system.");
  }

  return result;
}
```

That does not work as expected. The madvice returns success even if THPs are disabled at the system level. As a result, the UseTransparentHugePages flag stays enabled:

```
thomas@starfish$ cat /sys/kernel/mm/transparent_hugepage/enabled
always madvise [never]
thomas@starfish$ ./images/jdk/bin/java -XX:+UseTransparentHugePages -XX:+PrintFlagsFinal | grep Pages
     bool LargePagesIndividualAllocationInjectError = false {develop} {default}
     intx StackRedPages = 1 {pd product} {default}
     intx StackReservedPages = 1 {pd product} {default}
     intx StackShadowPages = 24 {pd product} {default}
     intx StackYellowPages = 2 {pd product} {default}
     bool UseLargePages = true {pd product} {default}
     bool UseLargePagesIndividualAllocation = false {pd product} {default}
     bool UseTransparentHugePages = true {product} {command line}
```

Also verified with gdb.

The real test would be to check /sys/kernel/mm/transparent_hugepage/enabled.

A variant of that problem is that we do workarounds meant for THP mode always, as in JDK-8303215. That also ties in with JDK-8310233.

So it seems THP setup needs some more TLC.