RFR: 8202184: Reduce time blocking the ClassSpecializer cache creating SpeciesData
paul.sandoz at oracle.com
Tue Apr 24 19:01:39 UTC 2018
This looks good. Took me a while to understand the interactions: you need to replace not update otherwise there is a race on isResolved (which currently queries multiple state, there is no singular guard here). We could push isResolved into the synchronized block and simplify but every findSpecies call may take a small hit (or there are potentially other ways looking more closely at how ClassSpecializer.Factory initializes state, i.e. a fenced static field write, but we go further down the rabbit hole)
I think this might fix some rare and intermittent recursive exceptions from ConcurrentHashMap cache we have been observing, where a collision occurs on keys for recursive updates (rather than for the same key).
> On Apr 24, 2018, at 9:57 AM, Claes Redestad <claes.redestad at oracle.com> wrote:
> the current implementation of ClassSpecializer.findSpecies may cause
> excessive blocking due to a potentially expensive computeIfAbsent, and
> we have reason to believe this might have been cause for a few very
> rare bootstrap issues in tests that attach debuggers to VM in the middle
> of this.
> Breaking this apart so that code generation and linking is done outside
> of the computeIfAbsent helps minimize the risk that a thread gets
> interrupted by a debugger at a critical point in time, while also removing
> a few limitations on what we can do from a findSpecies, e.g., look up other
> Bug: https://bugs.openjdk.java.net/browse/JDK-8202184
> Webrev: http://cr.openjdk.java.net/~redestad/8202184/open.00/
> This implementation inserts a placeholder, unresolved SpeciesData,
> and then replaces it with another instance that is linked together
> fully before publishing it, which ensure safe publication. There might
> be alternatives involving volatiles and/or careful fencing, but I've not
> been able to measure any added startup cost from this anyway.
More information about the core-libs-dev