RFR: 8353564: Fail fatally if os::release_memory or os::uncommit_memory fails [v5]

[Robert Toyonaga]

On Wed, 28 Jan 2026 21:24:33 GMT, Robert Toyonaga <duke at openjdk.org> wrote:

>> This PR is a follow up to JDK-8341491. See original discussion: https://github.com/openjdk/jdk/pull/24084#issuecomment-2752513700
>> 
>> This PR makes `os::release_memory`, `os::uncommit_memory`, `os::release_memory_special`, and `os::unmap_memory` fail fatally if they encounter an error. These methods require obtaining the NMT lock. Fatally failing in these places would potentially allow for the tightening of NMT virtual memory locking scopes (future work, if this PR is accepted).  Already in most cases, the callers fail fatally or assert(false) when these os:: methods fail. Another reason to fatally fail is that if the OS memory operation fails, it can be difficult to know for sure what state the OS left the memory in and recover. 
>> 
>> `release_memory`/`uncommit_memory`/`release_memory_special`/`unmap_memory` can fail due to ① Bad arguments, or ② The OS encountered an issue out of control of the JVM.
>> 
>> ①
>> If there is a JVM bug, it's probably reasonable to fatally fail here. Or the caller could be intentionally passing arguments that may or may not be valid. I don't think there is any code like that currently, and this is probably a bad pattern to be following anyway.
>> 
>> ②
>> In platform dependent code: 
>> With regard to mmap/munmap, the only errors that aren't due to bad arguments are ENOMEM and ones related to file descriptors (which are not applicable to uncommit or release).
>> On Windows, VirtualFree only fails due to bad arguments.
>> On AIX, shmdt and disclaim64 only fail due to bad arguments. msync could spontaneously fail with EIO: "An I/O error occurred while reading from or writing to the file system."
>> On BSD, it seems like mprotect and madvise fail only due to bad arguments or invalid privileges.
>> 
>> In the [original discussion](https://github.com/openjdk/jdk/pull/24084#issuecomment-2752513700), the main question was whether ENOMEM upon os::uncommit_memory was recoverable. This may be possible if we uncommit the middle of a region - splitting it in two. This could exceed the limit of the number of mappings resulting in ENOMEM. 
>> 
>> If none of the scenarios in ② are recoverable, then perhaps fatally failing is OK. 
>> 
>> Testing:
>> 	- Tier 1.
>> 	- Manual testing to make sure we fatally fail and the correct messages are printed.
>
> Robert Toyonaga has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Don't pass false, rely on default argument. Punctuation. Whitespace.

Thank you for the review Thomas! 
>The patch uncovers a hidden error in ZGC tests. It uses os::uncommit_memory without checking the return type (ZForwardingTest::TearDown). It probably never worked before; now we assert. I actually like the new behavior better. Releasing not-matching memory ranges is dangerous (NMT would catch that too, though).
>
>I think what happens here is that the underlying memory we operate on was reserved (in ZVirtualMemoryReserver -> ZVirtualMemoryReserverSmallPages::reserve -> ZMapper::reserve) via VirtualAlloc2, and that for whatever reason os::uncommit does not work here. Maybe the range does not match ? Or we need to call a different API.

Yes, the problem was that the ZGC API reserves using `VirtualAlloc2` placeholders and then `os::commit_memory` was called on the same region - which uses `VirtualAlloc` and doesn't support placeholders. The commit fails, so later the `os::uncommit_memory` fails too.

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

PR Comment: https://git.openjdk.org/jdk/pull/29240#issuecomment-3819250968