MethodHandle initialization process - problem with JVM TI early VM start event

[David Holmes]

Hi Alan,

On 4/01/2017 10:15 PM, Alan Bateman wrote:
> On 04/01/2017 10:39, David Holmes wrote:
>
>>
>> That is my thought too, that the spec needs to give less of the
>> impression that it's okay to access java.base classes at this early VM
>> start event, and basically say that any form of class-loading is not
>> guaranteed to succeed and will quite likely crash the JVM.
> The JVM TI spec has always allowed agents to call any JNI function in
> the start phase. I don't think there was any intention to have agents
> load and execute arbitrary java code but this wasn't fully spelled out.

Yes the spec has always been completely broken in this regard IMHO.

> For JDK 9 then we attempt to preserve this compatibility for existing
> agents by deferring the start phase until after the module system is
> initialized (initPhase2). This has the side effect that they miss out on
> some interesting events during startup. They can of course replay at
> least some of them with GenerateEvents but it's not enough for some
> agents. So this is the reason for the can_generate_early_vmstart
> capability and it's intended for agents that take an oath of carefulness.

Understood.

> So for the spec update then I think the restrictions can be mostly
> limited to when the can_generate_early_vmstart capability is enabled.
> Ideally we should avoid introducing yet another event that signals the
> point in the start phase when it's safe to do things, agents can use the
> VMInit for that.

My precise intent is to modify the text that was added in regard to the 
can_generate_early_vmstart capability, and further reduce expectations 
about what can be expected to work at that time. I have filed a new bug 
to make this correction to the JVM TI spec:

https://bugs.openjdk.java.net/browse/JDK-8172261

Thanks,
David

> -Alan