RFR: 8263017: Read barriers are missing in nmethod printing code
Erik Österlund
eosterlund at openjdk.java.net
Thu Mar 11 16:53:11 UTC 2021
On Wed, 10 Mar 2021 17:59:30 GMT, Vladimir Ivanov <vlivanov at openjdk.org> wrote:
> OOP accesses from nmethod printing code miss read barriers.
> It causes sporadic crashes when the code stumbles upon a stale OOP.
>
> When `nmethod::print_nmethod` is invoked, the corresponding nmethod is already installed in the code cache and there was a safepoint check in between (see `ThreadToNativeFromVM` usage in `CompileBroker::invoke_compiler_on_method`).
>
> The fix adds missing read barriers in `oop_Relocation::oop_value()` and introduces `oop_Relocation::raw_oop_value()` for cases when read barriers aren't needed.
>
> Along the way, there are some cleanups applied (e.g, `nmethod::print_[oops|metadata]` is superseded by `nmethod::print_recorded_[oops|metadata]`).
>
> Testing:
> - [x] failing test w/ `-XX:+UseZGC` `-XX:ZCollectionInterval=0.01` `-XX:ZFragmentationLimit=0`
> - [x] hs-tier1 - hs-tier4
>
> Thanks!
The load barrier on misaligned oop relocations will cause intermittent crashes. This is not safe. I can give you a fix tomorrow hopefully.
-------------
Changes requested by eosterlund (Reviewer).
PR: https://git.openjdk.java.net/jdk/pull/2919
More information about the hotspot-gc-dev
mailing list