RFR(S): 8170409: CMS: Crash in CardTableModRefBSForCTRS::process_chunk_boundaries

[Volker Simonis]

On Tue, Nov 29, 2016 at 11:58 AM, Mikael Gerdin
<mikael.gerdin at oracle.com> wrote:
> Hi Volker,
>
> On 2016-11-28 19:33, Volker Simonis wrote:
>>
>> Hi,
>>
>> can I please have a review and sponsor for the following fix submitted
>> by gunter.haug at sap.com:
>>
>> http://cr.openjdk.java.net/~simonis/webrevs/2016/8170409/
>
>
> I haven't looked too closely on the change but one thing I'm curious about
> just to further my understanding of the issue:
> Would an alternate solution be to always acquire the ParGCRareEvent_lock
> before reading _last_LNC_resizing_collection[i] since in that case the mutex
> semantics would ensure proper ordering of the memory accesses?
>

Hi Mikael,

we've discussed that as well and we think you are right in that always
acquiring the ParGCRareEvent_lock before reading
_last_LNC_resizing_collection[i] would solve the problem as well. But
as ParGCRareEvent_lock is potentially a heavy-weight OS lock that
would not be a practical solution.

Thanks,
Volker

> Thanks
> /Mikael
>
>
>> https://bugs.openjdk.java.net/browse/JDK-8170409
>>
>> We've observed a crash (see bug report for a stack trace) in
>> CardTableModRefBSForCTRS::process_chunk_boundaries() from time to time
>> since several years now, but only on non TSO platforms:
>>
>> - It only happens in opt builds.
>> - Analysis of the assembly code revealed the actual crash site to be
>> an array store to a pointer (_lowest_non_clean) which is an argument
>> to process_chunk_boundaries()
>> - The pointer is actually calculated in
>> CardTableModRefBS::get_LNC_array_for_space() and passed as argument to
>> CardTableModRefBSForCTRS::process_chunk_boundaries()
>> - CardTableModRefBS::get_LNC_array_for_space() doesn't enforce TSO on
>> _last_LNC_resizing_collection[i] so the pointer to an uninitialized
>> structure (i.e._lowest_non_clean) could become visible to other
>> threads before the value of _last_LNC_resizing_collection[i].
>>
>> Solution:
>>
>> Use OrderAccess::load_acquire and OrderAccess::release_store for
>> accessing _last_LNC_resizing_collection[i] in
>> CardTableModRefBSForCTRS::get_LNC_array_for_space()
>>
>> Thanks you and best regards,
>> Volker
>>
>