RFR(S): 8058461: serviceability/dcmd/CodelistTest.java and serviceability/dcmd/CompilerQueueTest.java SIGSEGV
vladimir.kozlov at oracle.com
Thu Sep 18 00:59:15 UTC 2014
CompileQueue::add() path already has lock acquired.
I think you need the lock only in CompileBroker::print_compile_queues().
On 9/17/14 1:54 AM, Nils Eliasson wrote:
> Hi Albert,
> On 2014-09-16 10:45, Albert wrote:
>> Hi Nils,
>> please see comments inline.
>> On 09/16/2014 10:16 AM, Nils Eliasson wrote:
>>> I would like review of this change that includes three fixes:
>>> 1) Let Dcmd Compiler.codelist only print alive-nmethods. We ran into
>>> crashes when listing zombies and unloaded too. Alive nmethods
>>> includes not-entrants so it still gives a pretty good idea about
>>> whats in the code cache and what has been used recently.
>> What was the cause of these crashes? Is the reason that
>> nmethod->method() is null?
> In one case it crashed when nmethod->method()->constants() was null when
> printing the name.
> I see that the segmented code cache push already changed to only
> iterating the alive nmethods so I removed that from the patch.
>> Would it make sense to print information about zombie methods that
>> don't lead
>> to a crash?
> Probably not. Not_entrant nmethods may stay around for a while, but once
> zombies they are reclaimed pretty fast.
>>> 2) Take CompileQueue lock when printing queue. It is not enough to be
>>> at a safepoint - the compiler threads may still mutate the list
>>> causing crashes.
>> Are we guaranteed to not run into a deadlock when we acquire the lock?
> Yes. It doesn't use any other lock.
> New webrev: http://cr.openjdk.java.net/~neliasso/8058461/webrev.05/
> Thanks Albert and Christian for your feedback.
>>> 3) Relax the parsing of long hex-numbers in the test of codelist.
>>> High addresses (sparc) casues NumberFormatExceptions.
>>> bug: https://bugs.openjdk.java.net/browse/JDK-8058461
>>> webrev: http://cr.openjdk.java.net/~neliasso/8058461/webrev.04/
More information about the hotspot-compiler-dev