RFR(M): 8154736: enhancement of cmpxchg and copy_to_survivor for ppc64
erik.osterlund at oracle.com
Tue May 29 12:00:55 UTC 2018
Hi Martin and Michihiro,
On 2018-05-29 12:30, Doerr, Martin wrote:
> Hi Kim,
> I'm trying to understand how this is related to Michihiro's change. The else path of the initial test is not affected by it AFAICS.
> So it sounds like a request to fix the current implementation in addition to what his original intend was.
I think we are just trying to nail down the correct fencing and just go
for that. And yes, this is arguably a pre-existing problem, but in a
race involving the very same accesses that we are changing the fencing
for. So it is not completely unrelated I suppose.
In particular, hotspot has code that assumes that if you on the writer
side issue a full fence before publishing a pointer to newly initialized
data, then the initializing stores and their side effects should be
globally "visible" across the system before the pointer to it is
published, and hence elide the need for acquire on the loading side,
without relying on retained data dependencies on the loader side. I
believe this code falls under that category. It is assumed that the
leading fence of the CAS publishing the forwarding pointer makes the
initializing stores globally observable before publishing a pointer to
the initialized data, hence assuming that any loads able to observe the
new pointer would not rely on acquire or data dependent loads to
correctly read the initialized data.
Unfortunately, this is not reliable in the IRIW case, as per the litmus
test "MP+sync+ctrl" as described in "Understanding POWER
multiprocessors" (https://dl.acm.org/citation.cfm?id=1993520), as
opposed to "MP+sync+addr" that gets away with it because of the data
dependency (not IRIW). Similarly, an isync does the job too on the
reader side as shown in MP+sync+ctrlisync. So while what I believe was
the previous reasoning that the leading sync of the CAS would elide the
necessity for acquire on the reader side without relying on data
dependent loads (implicit consume), I think that assumption was wrong in
the first place and that we do indeed need explicit acquire (even with
the precious conservative CAS fencing) in this context to not rely on
implicit consume semantics generating the required data dependent loads
on the reader side. In practice though, the leading sync of the CAS has
been enough to generate the correct machine code. Now, with the leading
sync removed, we are increasing the possible holes in the generated
machine code due to this flawed reasoning. So it would be nice to do
something more sound instead that does not make such assumptions.
> Anyway, I agree with that implicit consume is not good. And I think it would be good to treat both o->forwardee() the same way.
> What about keeping memory_order_release for the CAS and using acquire for both o->forwardee()?
> The case in which the CAS succeeds is safe because the current thread has created new_obj so it doesn't need memory barriers to access it.
Sure, that sounds good to me.
> Thanks and best regards,
> -----Original Message-----
> From: Kim Barrett [mailto:kim.barrett at oracle.com]
> Sent: Dienstag, 29. Mai 2018 01:54
> To: Michihiro Horie <HORIE at jp.ibm.com>
> Cc: Erik Osterlund <erik.osterlund at oracle.com>; david.holmes at oracle.com; Gustavo Bueno Romero <gromero at br.ibm.com>; hotspot-dev at openjdk.java.net; hotspot-gc-dev at openjdk.java.net; ppc-aix-port-dev at openjdk.java.net; Doerr, Martin <martin.doerr at sap.com>
> Subject: Re: RFR(M): 8154736: enhancement of cmpxchg and copy_to_survivor for ppc64
>> On May 28, 2018, at 4:12 AM, Michihiro Horie <HORIE at jp.ibm.com> wrote:
>> Hi Erik,
>> Thank you very much for your review.
>> I understood that implicit consume should not be used in the shared code. Also, I believe performance degradation would be negligible even if we use acquire.
>> New webrev uses memory_order_acq_rel: http://cr.openjdk.java.net/~mhorie/8154736/webrev.10
> This is missing the acquire barrier on the else branch for the initial test, so fails to meet
> the previously described minimal requirements for even possibly being sufficient. Any
> analysis of weakening the CAS barriers must consider that test and successor code.
> In the analysis, it’s not just the lexically nearby debugging / logging code that needs to be
> considered; the forwardee is being returned to caller(s) that will presumably do something
> with that object.
> Since the whole point of this discussion is performance, any proposed change should come
> with performance information.
More information about the hotspot-gc-dev