RFR: 8237143: Eliminate DirtyCardQ_cbl_mon
sangheon.kim at oracle.com
sangheon.kim at oracle.com
Wed Feb 5 22:52:49 UTC 2020
On 1/31/20 2:25 PM, Kim Barrett wrote:
>> On Jan 23, 2020, at 3:10 PM, Kim Barrett <kim.barrett at oracle.com> wrote:
>>> On Jan 22, 2020, at 11:12 AM, Thomas Schatzl <thomas.schatzl at oracle.com> wrote:
>>> On 16.01.20 09:51, Kim Barrett wrote:
>>>> Please review this change to eliminate the DirtyCardQ_cbl_mon. This
>>>> is one of the two remaining super-special "access" ranked mutexes.
>>>> (The other is the Shared_DirtyCardQ_lock, whose elimination is covered
>>>> by JDK-8221360.)
>>>> There are three main parts to this change.
>>>> (1) Replace the under-a-lock FIFO queue in G1DirtyCardQueueSet with a
>>>> lock-free FIFO queue.
>>>> (2) Replace the use of a HotSpot monitor for signaling activation of
>>>> concurrent refinement threads with a semaphore-based solution.
>>>> (3) Handle pausing of buffer refinement in the middle of a buffer in
>>>> order to handle a pending safepoint request. This can no longer just
>>>> push the partially processed buffer back onto the queue, due to ABA
>>>> problems now that the buffer is lock-free.
>>>> mach5 tier1-5
>>>> Normal performance testing showed no significant change.
>>>> specjbb2015 on a very big machine showed a 3.5% average critical-jOPS
>>>> improvement, though not statistically significant; removing contention
>>>> for that lock by many hardware threads may be a little bit noticeable.
>>> initial comments only, and so far only about comments :( The code itself looks good to me, but I want to look over it again.
>> After some offline discussion with Thomas, I’m doing some restructuring that
>> makes it probably not very efficient for anyone else to do a careful review of
>> the open.00 version.
> Here's a new webrev:
Webrev.02 looks really good.
> mach5 tier1-5
> Performance testing showed no significant change.
> I didn't bother providing an incremental webrev, because the changes
> to g1DirtyCardQueue.[ch]pp are pretty substantial. Those are the only
> files changed, except for the suggested move of the comment for
> G1ConcurrentRefineThread::maybe_deactivate and some related comment
> improvements nearby.
> Most of this round of changes are refactoring within G1DirtyCardQueueSet,
> mainly adding internal helper classes for the FIFO queue and for the paused
> buffers, each with their own (commented) APIs. I think that has addressed a
> lot of Thomas's comments about the comments, and I hope has made the code
> easier to understand.
> I've also improved the mechanism for handling "paused" buffers, simplifying
> it by making better use of some invariants.
>> On Jan 22, 2020, at 11:12 AM, Thomas Schatzl <thomas.schatzl at oracle.com> wrote:
>> // The key idea to make this work is that pop (get_completed_buffer)
>> // never returns an element of the queue if it is the only accessible
>> // element,
>> If I understand this correctly, maybe "if there is only one buffer in the FIFO" is easier to understand than "only accessible element". (or define "accessible element”).
> I specifically don't want to say it that way because we could have a
> situation like
> (1) Start with a queue having exactly one element.
> (2) Thread1 starts a push by updating tail, but has not yet linked the old
> tail to the new.
> (3) Thread2 performs a push.
> The buffer pushed by Thread2 is "in the queue" by some reasonable
> definition, so the queue contains two buffers. But that buffer is not yet
> accessible, because Thread1 hasn't completed its push. The alternative is
> to (in the description) somehow divorce a completed push from the notion of
> the number of buffers in the queue, which seems worse to me. I expanded the
> discussion a bit though, including what is meant by "accessible".
>> The code seems to unnecessarily use the NULL_buffer constant. Maybe use it here too. Overall I am not sure about the usefulness of using NULL_buffer in the code. The NULL value in Hotspot code is generally accepted as a special value, and the name "NULL_buffer" does not seem to add any information.
> The point of NULL_buffer was to avoid casts of NULL in Atomic operations,
> and I then used it consistently. But I've changed to using such casts,
> since it turned out there weren't that many and we can get rid of those
> uniformly here and elsewhere when we have C++11 nullptr and nullptr_t.
More information about the hotspot-gc-dev