RFC (csr): 8221528: Introduce compatibility mode with VM option -XX:AllowRedefinitionToAddOrDeleteMethods

[Thomas Stüfe]

On Thu, Mar 28, 2019 at 11:10 AM David Holmes <david.holmes at oracle.com>
wrote:

> On 28/03/2019 4:55 pm, Thomas Stüfe wrote:
> > On Thu, Mar 28, 2019 at 8:40 AM David Holmes <david.holmes at oracle.com
> > <mailto:david.holmes at oracle.com>> wrote:
> >
> >     Hi Thomas,
> >
> >     On 28/03/2019 5:02 pm, Thomas Stüfe wrote:
> >      > Hi Serguei,
> >      >
> >      > This is planned to be introduced in jdk 13?
> >      >
> >      > The only concern I have is that both paths (old and new behavior)
> >     should
> >      > be well tested, and hiding the old behavior behind an optional
> >     switch
> >      > may reduce its test coverage considerably.
> >
> >     The old behaviour is hardly ever used - I'm not even sure we
> currently
> >     have a test that tries to use it - so it's already not "well tested"
> >     and
> >     there's no intent here of increasing test coverage for the code we
> plan
> >     to remove. The new path will be tested as we do have a test that
> >     expects
> >     to get the error. And Serguei will of course have a simple test that
> >     checks the flag with both values.
> >
> >
> > I remember at least one case causing crashes in our code base where the
> > method ordering array was not reordered on RedefineClasses. May have
> > been this one: JDK-8149743.
> >
> > My point is, I have seen crashes in the field coming from bytecode
> > agents redefining classes with method added or substracted, so it is no
> > theoretical problem.
>
> My point is that we don't really care if the code works well or not,
> we're intending to kill it off not improve it. It's a capability that
> really should never have snuck in to the code.
>
>
If you provide a backward compatibility switch which is not well tested,
the coding behind it may bitrot; then, when you want to use it, you would
run into errors. That is worse than not having that switch.

Jdk head development goes at quite a pace, the chance that something which
is not tested well bitrots is not that absurd.

I am not idly bike shedding. I remember distinctly running into problems
associated with RedefineClasses method adding/removal (I even remember us
talking about it). So I know this is a thing which can happen. I do not
want to improve it but I would like it to reliably work for as long as this
switch exists.

Cheers, Thomas


> Cheers,
> David
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.java.net/pipermail/serviceability-dev/attachments/20190328/afd29eed/attachment-0001.html>