On Fri, 6 Sep 2024 17:51:15 GMT, Jorn Vernee <jver...@openjdk.org> wrote:
>> As discussed in the JBS issue: >> >> FFM upcall stubs embed a `Method*` of the target method in the stub. This >> `Method*` is read from the `LambdaForm::vmentry` field associated with the >> target method handle at the time when the upcall stub is generated. The MH >> instance itself is stashed in a global JNI ref. So, there should be a >> reachability chain to the holder class: `MH (receiver) -> LF (form) -> >> MemberName (vmentry) -> ResolvedMethodName (method) -> Class<?> (vmholder)` >> >> However, it appears that, due to multiple threads racing to initialize the >> `vmentry` field of the `LambdaForm` of the target method handle of an upcall >> stub, it is possible that the `vmentry` is updated _after_ we embed the >> corresponding `Method`* into an upcall stub (or rather, the latest update is >> not visible to the thread generating the upcall stub). Technically, it is >> fine to keep using a 'stale' `vmentry`, but the problem is that now the >> reachability chain is broken, since the upcall stub only extracts the target >> `Method*`, and doesn't keep the stale `vmentry` reachable. The holder class >> can then be unloaded, resulting in a crash. >> >> The fix I've chosen for this is to mimic what we already do in >> `MethodHandles::jump_to_lambda_form`, and re-load the `vmentry` field from >> the target method handle each time. Luckily, this does not really seem to >> impact performance. >> >> <details> >> <summary>Performance numbers</summary> >> x64: >> >> before: >> >> Benchmark Mode Cnt Score Error Units >> Upcalls.panama_blank avgt 30 69.216 ± 1.791 ns/op >> >> >> after: >> >> Benchmark Mode Cnt Score Error Units >> Upcalls.panama_blank avgt 30 67.787 ± 0.684 ns/op >> >> >> aarch64: >> >> before: >> >> Benchmark Mode Cnt Score Error Units >> Upcalls.panama_blank avgt 30 61.574 ± 0.801 ns/op >> >> >> after: >> >> Benchmark Mode Cnt Score Error Units >> Upcalls.panama_blank avgt 30 61.218 ± 0.554 ns/op >> >> </details> >> >> As for the added TestUpcallStress test, it takes about 800 seconds to run >> this test on the dev machine I'm using, so I've set the timeout quite high. >> Since it runs for so long, I've dropped it from the default `jdk_foreign` >> test suite, which runs in tier2. Instead the new test will run in tier4. >> >> Testing: tier 1-4 > > Jorn Vernee has updated the pull request incrementally with one additional > commit since the last revision: > > remove PC save/restore on s390 Fun facts (measured on PPC64le fastdbg build which contains some asm asserts): simple upcall stub: 520 Bytes (regardless of selected GC) upcall_stub_load_target with G1: 72 Bytes upcall_stub_load_target with ZGC: 1204 Bytes (more than 2x the upcall stub + the G1 version!) upcall_stub_load_target with Shenandoah: 1364 Bytes Great to have the upcall_stub_load_target only once :-) ------------- PR Comment: https://git.openjdk.org/jdk/pull/20479#issuecomment-2335182768
