RFR: JDK-8313319: [linux] mmap should use MAP_FIXED_NOREPLACE if available [v2]

[David Holmes]

On Wed, 16 Aug 2023 14:24:44 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

>> os::attempt_reserve_memory_at() is used to map memory at a specific address that we fancy. It is not used to replace an existing mapping (hence we omit MAP_FIXED). 
>> 
>> But it may result in the kernel creating a mapping at a different address if our wish address is blocked; we then have to unmap again. That mmap-munmap cycle is unnecessary.
>> 
>> MAP_FIXED_NOREPLACE exists since Linux 4.17 and prevents that unnecessary cycle . We should use that if available. We need to tiptoe around a bit to deal with building on old machines/running on old kernels, though.
>
> Thomas Stuefe has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains two additional commits since the last revision:
> 
>  - Merge branch 'openjdk:master' into JDK-8313319-MAP_FIXED_NOREPLACE
>  - DK-8313319-MAP_FIXED_NOREPLACE

This seems reasonable. I'd like to take it for a spin through our CI though. Thanks.

src/hotspot/os/linux/os_linux.cpp line 3485:

> 3483:   // If a requested address was given:
> 3484:   //
> 3485:   // The POSIX-conform way is to *omit* MAP_FIXED. This will leave existing mappings intact.

s/conform/conforming/

src/hotspot/os/linux/os_linux.cpp line 3489:

> 3487:   // somewhere else. On Linux, that alternative address appears to have no relation to the
> 3488:   // requested address.
> 3489:   // Unfortunately, this is not what we need - if we specified a request address, we'd want

s/request/requested/

src/hotspot/os/linux/os_linux.cpp line 3494:

> 3492:   // Since Linux 4.17, the kernel offers MAP_FIXED_NOREPLACE. With this flag, if a pre-
> 3493:   // existing mapping exists, the kernel will not map at an alternative point but instead
> 3494:   // return an error. We can therefore save that unnessassary mmap-munmap cycle.

typo: unnecessary

test/hotspot/gtest/runtime/test_os.cpp line 930:

> 928: TEST_VM(os, reserve_at_wish_address_shall_not_replace_mappings_smallpages) {
> 929:   char* p1 = os::reserve_memory(M, false, mtTest);
> 930:   ASSERT_NE(p1, (char*)nullptr);

Do you need casts on nullptr?

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

PR Review: https://git.openjdk.org/jdk/pull/15170#pullrequestreview-1590816377
PR Review Comment: https://git.openjdk.org/jdk/pull/15170#discussion_r1302523819
PR Review Comment: https://git.openjdk.org/jdk/pull/15170#discussion_r1302524294
PR Review Comment: https://git.openjdk.org/jdk/pull/15170#discussion_r1302524875
PR Review Comment: https://git.openjdk.org/jdk/pull/15170#discussion_r1302526703