RFR: 8226228: Make Threads_lock an always safepoint checked lock.
Daniel D. Daugherty
daniel.daugherty at oracle.com
Thu Aug 8 19:32:56 UTC 2019
On 8/6/19 2:29 AM, Robbin Ehn wrote:
> Hi guys,
> Let's pick this up :) Been a while...
> I think that we were pretty happy after some explaining !?
> Here is full, let's call it v2:
L4139: // This thread will not be safepoint check, since it has
nit: s/not be/not do a/
L190: // Safepoint checking logically implies active java_thread.
You may want this comment:
// The VM is going to die so no need to unlock Thread_lock.
Thumbs up. I only have comment changes/additions so no need for
a new webrev if you choose to make those changes.
> Passes t1-5.
> Thanks, Robbin
> On 2019-06-17 13:21, Robbin Ehn wrote:
>> Hi all, please review.
>> A safepoint checked lock always check for safepoint if the locker is an
>> JavaThread. If it's not a JavaThread, the thread are not
>> participating in the
>> safepoint protocol thus safepoints are not checked.
>> A JavaThread that is not on _the_ ThreadsList is not participating in
>> the safepoint protocol and for most purposes should be treat as
>> This applies to whether we should check for a safepoint.
>> This changeset suggest adding a method for checking this.
>> We can use that method in the mutex code to determine if we should
>> treat locker
>> as JavaThread or a NonJavaThread.
>> The only problematic scenario is when we are an active JavaThread in
>> vmError.cpp but don't want to safepoint, but want the ThreadsList to
>> be stable.
>> I choose a best effort approach.
>> Alternatives could be stop freeing the backing array if there is an
>> an error or
>> letting the thread running the error code become an inactive
>> JavaThread thus
>> avoiding safepointing.
>> This fix also helps converting Heap_lock to an always lock (Heap_lock
>> changes not included).
>> Passes t1-3.
>> Thanks, Robbin
More information about the hotspot-runtime-dev