RFR: 8360048: NMT crash in gtest/NMTGtests.java: fatal error: NMT corruption: Block at 0x0000017748307120: header canary broken [v2]

[Aleksey Shipilev]

On Mon, 14 Jul 2025 20:56:16 GMT, David Holmes <dholmes at openjdk.org> wrote:

>> This is a clone of https://github.com/openjdk/jdk/pull/25950 that we need to get integrated ASAP.
>> 
>> ---
>> 
>> The canary header test failed since there were concurrent remove and free() from the tree. The remove operations are synch'ed with corresponding NMT lock. The ReserveMemory::reserve() uses the same lock internally and is not included in the locked code block.
>> 
>> ---
>> 
>> I'm re-testing with tiers 1-4
>
> David Holmes has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - Update test/hotspot/gtest/runtime/test_virtualMemoryTracker.cpp
>    
>    Co-authored-by: Aleksey Shipilëv <shipilev at amazon.de>
>  - Update src/hotspot/share/nmt/virtualMemoryTracker.cpp
>    
>    Co-authored-by: Aleksey Shipilëv <shipilev at amazon.de>

So replacing the uses of (sigleton, shared) `VirtualMemoryTracker::Instance` with local `VirtualMemoryTracker vmt(true);` dodges the locking issue, right?

What confuses me in this patch is why are we doing replacements like:


-  site->commit_memory(rgn->committed_size());
+  site->commit_memory(VirtualMemoryTracker::Instance::committed_size(rgn));


Doesn't that introduce dependencies on that singleton instance? Can you confirm that is sane?

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

PR Review: https://git.openjdk.org/jdk/pull/26284#pullrequestreview-3019298822