RFR: 8235324: Dying objects are published from users of CollectedHeap::object_iterate
stefan.karlsson at oracle.com
Wed Dec 4 12:14:39 UTC 2019
On 2019-12-04 12:23, Per Liden wrote:
> Hi Stefan,
> On 12/4/19 11:34 AM, Stefan Karlsson wrote:
>> Removed the unused oopDesc::keep_alive function:
> zHeap.cpp should include zBarrier.inline.hpp now.
> I would maybe like to restructure the ZGC keepalive barrier a bit (or
> at least discuss some ideas). For example, I'm thinking the fast path
> could be "is_null_or_not_during_mark". But let's take that as a
> separate discussion/change.
I was also thinking about not using the barrier(...) infrastructure at
all, just calling the appropriate "slow path" function.
I also think the other keep_alive_* functions are poorly named, they
just fix the oop but don't keep the object alive.
I'd be happy to deal with that in a separate discussion.
> So, other than the missing include, looks good. No need for a new webrev.
>> On 2019-12-04 09:50, Stefan Karlsson wrote:
>>> Hi all,
>>> Please review this fix to make sure objects published from closures
>>> to CollectedHeap::object_iterate are kept alive.
>>> Caution needs to be taken when CollectedHeap::object_iterate is
>>> used. It may:
>>> - Return non-reachable objects - for example, when G1 walks the prev
>>> - Return the Reference.referent without an appropriate resurrection
>>> - Return an object loaded without a "strong"/"marking" barrier
>>> These objects should not escape the confines of the active safepoint
>>> that calls object_iterate. If they do we run the risk of various
>>> crashes when the objects get garbage collected.
>>> Unfortunately, there are a few closures that do this. The once we
>>> have identified are:
>>> - HeapInspection::find_instances_at_safepoint
>>> - JVMTI object tagging
>>> There might be some merit to solving this by altering the
>>> object_iterate API, but for now we'd like to fix the two problematic
>>> areas listed above. The suggestion is to introduce a
>>> CollectedHeap::keep_alive function that notifies the GC that the
>>> object should be kept alive the current GC cycle (if active).
>>> This solution will also work with code that iterates over object
>>> fields without using object_iterate. For example, with this fix
>>> we'll be able to read the Reference.referent field in JDK-8234508,
>>> with AS_NO_KEEPALIVE, and only keep the object alive if it actually
>>> gets registered in the tag map. Without that, every single object
>>> reachable through Reference.referent would be kept alive.
>>> The fix introduces a keep_alive() function for ZGC, G1, and
>>> Shenandoah. It would be good to get reviews from maintainers of all
>>> these GCs.
More information about the hotspot-gc-dev