Class unloading in ZGC

[Liang Mao]

Hi Erik,

Your answer is a great help. Thanks a lot for sharing the valuable thinking.

Liang




------------------------------------------------------------------
From:Erik Österlund <erik.osterlund at oracle.com>
Send Time:2020 Dec. 7 (Mon.) 23:18
To:"MAO, Liang" <maoliang.ml at alibaba-inc.com>; zgc-dev <zgc-dev at openjdk.java.net>
Subject:Re: Class unloading in ZGC

 Hi Liang,

On 2020-12-07 15:56, Liang Mao wrote:
Hi Erik,

Thank you! It's exactly what I want to know!

> The contract with a SATB collector like G1 is that we need to apply barriers when loading
> a weak oop. The nmethod oops are weak. So not applying nmethod entry barriers, does seem
> like a violation of the SATB invariant, for G1. However, people are arguing that it is okay,
> as all oops embedded in nmethods, that are reachable by mutators during concurrent marking,
> will have their oops marked through. 
Do you mean the weak reference in G1 has the implicit *weak_pointer = null, and we
 need to record the previous value to make sure the SATB invariant?   
 Something like that. Except instead of clearing weak pointers, we make the whole nmethod
 unloaded, and throw away the whole nmethod. All other weak references are cleared during
 reference processing, when they are found to no longer be alive. But nmethods are special,
 resulting in more drastic action than clearing, if they have an embedded oop that is dead.

But for those nmethod which is only on-stack during concurrent marking, we didn't
enqueue the previous value. It is safe so far because the limitation of embeded oop,
right?   
 Exactly. And that's what I am referring to as fragile. It has been debated many times among
 our GC engineers, if it has holes or not. We wanted something more robust.


BTW, if ZGC has an explicit load barrier while accessing an oop from nmethod embeded,
is the nmethod entry barrier still necessary? Does nmethod entry barrier play the role
of such load barrier? 

 It's a swiss army knife, serving multiple roles. It is also used to coordinate lazy cleaning of
 inline caches. In particular, when we reach mark end, nmethods will have dead oops in them.
 But inline caches from other nmethods are still pointing at said nmethods with dead oops.
 What this implies is that the nmethod with the dead oop will get unloaded, but has not been
 unloaded yet. Yet threads that wake up from the safepoint absolutely must not perform calls
 into such nmethods. Normally inline caches are cleaned in the safepoint operation that unloads
 the code cache. But we don't have time for that and let calls into dying nmethods go ahead,
 because we know they will take the slow path in our nmethod entry barrier, and then re-resolve
 the call to something less dead. So apart from dealing with oops, we really do need this also
 for dealing with inline caches (and similarly static calls embedded as direct calls).

 Another thing we have to consider, which is not inherent, but is currently true, is that the
 oops embedded into the code stream on x86_64 are misaligned in memory. This puts particular
 constraints on what mechanism is used to heal the pointers in the nmethod. We could not
 simply use a load barrier with a CAS, as the oop could cross two cache lines, which at the
 very least the specification does not allow. In practice it might work due to luck, but induce
 costs that are huge, similar to inter-processor interrupts (IPI). So that should be avoided at
 all cost. With nmethod entry barriers, we can take mutators into a path where the nmethod
 oops are protected by a per-nmethod lock. One thread will heal the oops, and no other thread
 will concurrently read them.

 It is also the case that performing a load barrier into the nmethod would not necessarily work,
 even if the oops were aligned, as the data and instruction caches are not necessarily in sync.
 That is why the check for the nmethod entry barrier is notified of being disarmed via writes to
 the instruction stream. x86_64 machines have an explicit guarantee that instruction modification
 is observed by the execution, in the order the modifications were written. That means that if the
 nmethod entry barrier executed the instruction that perceived the nmethod as disarmed, then
 that guarantees that instruction executions with immediate oops will also observe the updated
 immediate oop.

 So yeah, there are indeed multiple things that would not work very well without the nmethod
 entry barrier.

 Hope this helps understanding why we do the things that we do.

 /Erik

Thanks,
 Liang



------------------------------------------------------------------
From:Erik Österlund <erik.osterlund at oracle.com>
Send Time:2020 Dec. 7 (Mon.) 21:26
To:"MAO, Liang" <maoliang.ml at alibaba-inc.com>; zgc-dev <zgc-dev at openjdk.java.net>
Subject:Re: Class unloading in ZGC

 Hi Liang,

 Sorry, I don't know if I understand what you are referring to specifically. I think you
 are talking about what happens when class unloading is enabled, am I right?

 If so, then there is indeed a difference between G1 and ZGC. They both scan the stacks,
 marking through on-stack nmethods. But ZGC also arms nmethod entry barriers, to lazily
 mark through nmethod oops. Here is why.

 ZGC needs to color all nmethod oops as "marked" before exposing them to mutator threads.
 We also think that explicitly marking oops exposed to mutators is the most robust way
 of treating these oops, as they are indeed weak until used. So marking them in the nmethod
 entry barrier during concurrent marking, is in spirit very similar to applying a weak load
 barrier on Reference.get(), which also G1 does.

 The contract with a SATB collector like G1 is that we need to apply barriers when loading
 a weak oop. The nmethod oops are weak. So not applying nmethod entry barriers, does seem
 like a violation of the SATB invariant, for G1. However, people are arguing that it is okay,
 as all oops embedded in nmethods, that are reachable by mutators during concurrent marking,
 will have their oops marked through. That is okay, as long as the compiler knows about SATB,
 and hence what oops it is allowed to embed in the nmethods. If the compiler was to for example
 embed a string from the string table, that might not necessarily be reachable by the holders
 of the inlined method holders, then this approach would crash as the violation of the SATB
 contract would suddenly become more visible. By using nmethod entry barriers, this logic
 becomes more robust, as the compiler does not have to know what oops it may or may not embed
 into the code stream, as we explicitly apply barriers.

 While the robustness reason is one reason to do this dance regardless, we certainly do also
 need to apply the right colors in ZGC to the pointers, regardless of whether we would trust
 the actual objects to be marked or not. And, in order to deal with relocation properly, we
 needed something like nmethod entry barriers anyway, as a mutator really is not allowed to
 see not yet relocated oops. So with this mechanism already in place, it made sense to use it
 for marking as well, solving 3 problems at the same time: 1) ensuring the objects are marked
 in a more robust way, 2) ensuring the colors of exposed nmethods are good during marking, and
 3) dealing with concurrent relocation.

 I have argued that G1 should also use nmethod entry barriers to explicitly enforce its SATB
 invariant, regarding these weak oops, and that the way they are treated today is not robust.
 In fact, that is indeed being done in the loom repo, and is likely to become the standard way
 of dealing with concurrent marking w.r.t. nmethods, for all concurrently marking GCs in HotSpot.

 Hope this helps, and that I got your question right.

 Thanks,
 /Erik

On 2020-12-07 13:47, Liang Mao wrote:
Hi Erik,

If we are only considering the pause time thread root processing in jdk12-15. 
Comparing to G1 which only marks the on-stack nmethod at mark start pause
without nmethod entry barrier, ZGC will mark the on-stack nmethod
at mark start pause and also use nmethod entry barrier to do the marking.
Is the additional marking by nmethod entry barrier a specific behavior because of
 color pointer mechanism? 

Thanks,
Liang 



------------------------------------------------------------------
From:Erik Österlund <erik.osterlund at oracle.com>
Send Time:2020 Dec. 7 (Mon.) 20:08
To:"MAO, Liang" <maoliang.ml at alibaba-inc.com>; zgc-dev <zgc-dev at openjdk.java.net>
Subject:Re: Class unloading in ZGC

 Hi Liang,

On 2020-12-07 12:48, Liang Mao wrote:
 Hi Erik,

Appreciate your comprehensive reply!
I still have few quetion.
> -----Original Message-----
> From: Erik Österlund [mailto:erik.osterlund at oracle.com]
> Sent: 2020年12月7日 18:35
> To: Liang Mao <maoliang.ml at alibaba-inc.com>; zgc-dev <zgc-
> dev at openjdk.java.net>
> Subject: Re: Class unloading in ZGC
> 
> Hi Liang,
> 
> So there are two distict cases. Class unloading enabled (default), and class
> unloading disabled (seemingly for people that just really want to have memory
> leaks for no apparent good reason).
> 
> When class unloading is enabled, the code cache comprises weak roots, except
> oops that are on-stack that are treated as strong. These semantics are the same
> across all GCs.
> When marking starts, ZGC
> lazily processes the snapshot of nmethods that were on-stack when marking
> started, with lazy application of nmethod entry barriers. These barriers will mark

Sorry that I need to mention I was looking at the code of 8214897: ZGC: Concurrent Class Unloading.
It handled the on-stack nmethod at pause time. Do you mean the pause processing
is not necessary at that patch and the nmethod walking can be delayed as long as nmethod
entry barrier is there? 
On the other hand, if on-stack nmethod is processed at pause time in mark start,  the nmethod
entry barrier is not necessary?    
 What I was describing is what we do today, as opposed to what we did in JDK12.

 Back then, we did not have concurrent stack processing, which we do have today. Therefore,
 in that patch, I had to process stacks in a safepoint. Moreover, when class unloading is disabled,
 I walked the code cache in a safepoint. I was not feeling very motivated to optimize the case when
 class unloading is disabled, as there is pretty much no reason I can think of why you would want
 to disable it. It's just a memory leak with no benefit, to disable class unloading. For other collectors
 class unloading might come at a latency cost. But for ZGC it does not. So there does not seem to exist
 any form of trade-off.

 Since concurrent stack processing was integrated, there is no longer any need for processing
 the on-stack nmethods in safepoints, so that has been moved out of safepoints and is instead
 concurrently, incrementally and cooperatively applied through lazy nmethod entry barriers as
 the mutators return into frames that have not been processed yet. Since then, we have also made
 the code cache walk when class unloading is disabled concurrent, as it simplified the root processing
 code in the end to have only concurrent roots, instead of distinguising between STW and concurrent
 roots as well as strong vs weak. Now there is only strong vs weak, and no roots are scanned during
 safepoint operations, with or without class unloading.

 Thanks,
 /Erik

Thanks,
Liang

> the objects, and heal the pointers to the corresponding marked color, as
> expected by our barrier machinery. New nmethods that are called go through
> the same processing using nmethod entry barriers. Semantically this ensures that
> on-stack nmethods are treated as strong roots, and the rest of the nmethods
> are treated as weak roots.
> This has the same semantics
> as any other GC.
> 
> When class unloading is disabled, the code cache comprises strong roots.
> That means that the GC will
> during concurrent marking walk all nmethods, and mark the oops as strong.
> However, remember that there are two operations: marking the objects, and
> self-healing the pointers as expected by the barrier machinery.
> The second part of the operation still requires us to lazily apply nmethod entry
> barriers to the stacks as well as arming nmethod entry barriers for calls, during
> concurrent marking, so that the oops in the nmethods are self-healed to the
> corresponding marked pointer color, before they are exposed to the execution
> of mutators, which might for example store this oop into the object graph. So I
> suppose the special thing here compared to G1 is that we both walk the code
> cache marking all the oops, *and* explicitly walk the stacks marking them as
> well, with the main purpose of fixing the pointer colors before the mutator gets
> to use the nmethod. And arming the nmethod entry barriers for calls, for the
> same reason.
> 
> During relocation, we only arm the nmethod entry barriers with and without
> class unloading. The relocation is lazy and won't be performed until either
> someone uses the nmethod (on-stack lazy nmethod entry barrier or a call to a
> new nmethod), or the subsequent marking cycle will walk the code cache and
> make sure that the objects are remapped, when it is performing marking.
> 
> Hope this makes sense and sheds some light on this confusion.
> 
> /Erik
> 
> On 2020-12-06 16:40, Liang Mao wrote:
> > Hi ZGC team,
> >
> > Previously without concurrent class unloading in ZGC, the code cache
> > will be all treated as strong roots. Then concurrent class unloading
> > will only mark the nmethod of executing threads at mark start pause
> > and use the nmethod entry barrier to heal and also mark the oops. That
> > sounds reasonable. But when I looked into the concurrent marking in G1, it
> doesn't threat all code cache as strong roots and of course has no nmethod
> entry barrier. So I'm confused why ZGC need the nmethod entry barrier for
> > marking. Does the difference comes from the different algorithm of SATB vs
> load barrier?
> >
> > Thanks,
> > Liang
> >