RFR: 8212585: Clean up CompiledMethod::oops_reloc_begin()
erik.osterlund at oracle.com
Mon Nov 5 10:14:06 UTC 2018
Thanks for the review.
On 2018-11-05 10:51, Per Liden wrote:
> Hi Erik,
> On 10/18/18 11:07 PM, Erik Österlund wrote:
>> I decided to make this less risky for platforms that do not yet need
>> to be scanned concurrently, and hence don't really need to be "fixed".
>> In this new webrev, I decided to check for the frame completed offset
>> if available and further away than verified entry + native jump size.
>> Then it is safe for me to go ahead and scan this stuff concurrently
>> using nmethod entry barriers. Otherwise, the same good old behaviour
>> we used to rely on applies so that we won't get any surprises where
>> we do not want them.
> Looks good!
>> On 2018-10-17 17:36, Erik Österlund wrote:
>>> In CompiledMethod::oops_reloc_begin() there is code to prevent
>>> looking for oops at the verified entry of an nmethod that is not
>>> entrant, as a native jump instruction has been overwritten there.
>>> Except there would *never* be any immediate oops there for any
>>> compiler, even before becoming not entrant, with or without OSR
>>> compilation, so this special handling of not entrant vs entrant
>>> nmethods is seemingly completely unnecessary.
>>> This gets increasingly awkward when oops_do is called concurrently
>>> with concurrent class unloading, where an nmethod is racing to
>>> become not entrant.
>>> To clean this up, I propose to change the boundary for immediate oop
>>> scanning and start looking for oops only after the frame building is
>>> completed, as there is absolutely no point in looking for oops
>>> before that in the first place. This removes unnecessary workarounds
>>> that exist today, and removes unnecessary races going forward.
>>> It seems like Graal as JIT also inlines oops into the code stream,
>>> but never sets the CodeOffsets::Frame_Complete in the nmethod. So I
>>> recognize such nmethods and unconditionally start scanning at the
>>> verified entry + native jump instruction size. I spoke to the Graal
>>> folks, and they said they do not put oops in there either. So I
>>> think everyone involved should be happy with this solution.
More information about the hotspot-dev