On 05/04/2017 16:26, Gregg Wonderly wrote:
:
At the forefront of the failure of the SecurityManager to be an avidly used
element of Java applications, is the simple fact that the whole infrastructure
is horribly inefficient and full of locks and mutable data which should not be
locked and should instead be immutable. But, because “assume no security
manager in the picture” is the state of “the world”, the details have not been
fixed, officially. In the Jini community, Peter Firmstone (most of the recent
work), and I, have spent quite a bit of time on understand and fixing these
issues. He has tried to push the knowledge of what needs to be done into
various places, which I don’t recall at the moment.
You are right that the performance overhead when running with a security
manager can be a problem. Walking the stack to get the intersection of
the permissions has historically been a big part of this. If there are
proposals and patches to improve the performance then security-dev is
the place to bring those for discussion, maybe hotspot-dev if it gets
into the stack walking in the VM.
In the end because JDK-9 breaks everything (well not quite, but nearly, since
spring and other similar libraries are broken), requiring command line
arguments, why not just finally fix Java by instituting a default
SecurityManager, instead of the command line flags, which enforces all of these
things. Why not just require the added command line argument to be the
permissions file to use?
I assume the breakage you see is just a consequence of strong
encapsulation. There is a lot of existing code that relies on being able
to break into non-public members or access protected members from the
wrong context. These are issues that should be reported to the owners of
these libraries. We know it will be difficult in the interim before the
maintainers release new versions that work with JDK 9.
You mention Spring. There are some very smart people working on it and
tracking JDK 9 closely from what I can tell. Existing versions
SpringBoot fail today for a couple of reasons - one is that it is
helping itself to a non-public constructor for MethodHandles.Lookup.
Another is that it wants to invoke the protected ClassLoader.defineClass
from the wrong context. I'm confident that they will fix this code in
time as there new APIs to address these use cases. In the mean-time
then there are command line options to keep existing versions working.
Yes, command line options are pain but for executable JAR cases like
this then there are JAR file attributes that come some relief from CLI
options.
As regards the security manager then it's hard to see how it fits into
the discussion. To be honest, we don't see a lot of security manager
usage on the server side these days. I look at a lot of bug reports and
error logs that include the command line and I don't see
-Djava.security.manager very often. If the overhead was zero and it was
easier to identify the permissions needed by each component then maybe
it would have got more popular. Another reason is that the access checks
that you see in the discussions here are the access checks that the JLS
and JVMS specifies. They are access checks that have to be done when you
are not running with a security manager. They are also at a completely
different level, meaning can bytecode in C1.m can access a member of C2.
This is very different to the stack based permission checks that a
security manager does.
-Alan
