A possible JEP to replace SecurityManager after JEP 411

[Sean Mullan]

Hi David,

Thanks for the feedback and spending some time on this proposal. Some 
specific comments below.

On 4/5/22 9:52 AM, David Lloyd wrote:
> Here at Red Hat there have been serious discussions about the impacts
> of security manager removal on our users, and whether there is an
> actual value impact, and if so, whether it can be mitigated or
> reversed somehow. We are interested in exploring whether we can come
> up with a way in which vendors and projects that wish to continue
> using SecurityManager (or something like it) would be able to do so,
> while still removing the majority of the ongoing maintenance burden
> from the OpenJDK project.
> 
> Before we make a decision on whether or not we think there is
> sufficient justification for working up a formal JEP, we have decided
> that the best first step would be to socialize the idea in a more
> general form so that we can know whether the upstream OpenJDK team
> would even be amenable *at all* to the solution (or something like
> it), particularly in light of the observation that previous threads
> about retaining SecurityManager in any form have been looked upon in a
> fairly negative light.
> 
> The primary idea behind this proposal is that, while all of the points
> in JEP 411 relating to the lack of what most experts might refer to as
> "actual security" are certainly true, the SecurityManager mechanism
> itself does nevertheless have some inherent value. The challenge,
> then, is to strike a balance between the value provided by retaining
> some semblance of the mechanism versus the costs inherent in retaining
> it; we would want as much of the former as possible, for as little of
> the latter as possible.

With this proposal, as I understand it, the JDK would still be 
responsible for maintaining and preserving essentially all of the 
existing calls to the Security Manager (SM). All new code and APIs would 
still need to be evaluated and determined if permission checks were 
needed as well as making appropriate specification changes to note the 
behavior when an SM is enabled (throwing a SecurityException, etc). Any 
missing checks would need to be treated as security issues. And we would 
still need to test the code and APIs to ensure that it worked properly 
and complied with the API specification. This would likely mean 
implementing and maintaining an internal SM implementation in OpenJDK.

The proposal also includes retaining calls to doPrivileged (but later 
potentially replacing them with some other mechanism TBD). The JDK 
source code includes over 1000 calls to doPrivileged. Each of these need 
to be carefully reviewed to ensure that they do not contain security 
issues and any new code needs to be evaluated to see if new calls to 
doPrivileged are necessary.

Retaining doPrivileged (or something similar) means that there can be 
domains of code with different permissions running within the VM, which 
retains much of the complexity of the current SM model.

In this proposal, how privileges are established or propagated is 
implementation-specific. But how could applications or libraries depend 
on the APIs and still have some confidence that the code is behaving 
consistently and securely?

Today, the cost of buying into the SM model is high for libraries and 
applications. Not many third party libraries support the SM and have 
modified their code to perform permission checks and call doPrivileged 
in the right places. If there were pluggable SMs each behaving 
differently, there would likely be less incentive.

Although it sounds beneficial to be able to delegate the SM 
implementation to a 3rd-party, in reality, I think very few people would
take the time to implement it securely, and instead would mostly
leverage its power to do things that aren't at all security related. 
Sure, removing the default SM and Policy implementation reduces the 
complexity a little, but there would still be a fairly significant 
maintenance overhead and an additional drawback that it would make it 
more difficult for applications and libraries to depend on any type of 
consistent behavior.

--Sean


> So, here's the idea. It is assumed (for the sake of common
> understanding) that as things stand, all of the classes and members
> marked as "deprecated for removal" as a part of JEP 411 are intended
> to be completely removed without replacement at the end of the term of
> deprecation.  The proposals here are based on this assumption.
> 
> The center of this proposal is that, at the end of the term of
> deprecation, all of the deprecated classes, members, and behavior are
> still removed (including, and especially, AccessController and Policy
> and related classes) /except/ as mentioned here:
> 
>   * Rather than completely removing SecurityManager,
>       * The SecurityManager class becomes abstract and non-deprecated,
> with all of its methods being removed, except as follows
>       * SecurityManager.getSecurityContext() becomes abstract (this is
> the one that returns Object, *not* the stack walking one)
>       * SecurityManager.checkPermission() (both of them) become abstract
>   * Rather than removing the SecurityManager-related methods from System,
>       * System.getSecurityManager() is retained and de-deprecated
>       * [Optional] System.setSecurityManager() is retained and
> de-deprecated (we would want to explore whether it is feasible to
> replace this (and the system property lookup mechanism) using
> ServiceLoader, if bootstrap allows it)
>   * [Optional] Rather than /immediately/ removing all of AccessController,
>       * Retain its deprecation-for-removal status
>       * Retain only doPrivileged(PrivilegedAction) and
> doPrivileged(PrivilegedExceptionAction) as simple pass-throughs (no
> JVM semantics other than being present on the call stack like any
> method) since they are pervasively used, to allow frameworks time to
> transition to (for example) a third-party alternative.
> 
> The burden of permission verification would lie completely with the
> security manager implementation.  The JDK would not have a
> 'SecurityManager' implementation of any kind, outside of the internal
> test suite.
> 
> The other part of this proposal can come in one of two possible flavors.
> 
> ### Option 1: Authorization interfaces
> 
> Each point in the JDK where there presently is a permission check is
> classified into an authorization category of related operations. An
> interface is introduced for each category which contains the methods
> encapsulating the relevant check, in a package that is deemed most
> appropriate for that particular grouping.  For example, there might be
> a 'SocketAuthorization' interface in the 'java.net' package, with
> methods like 'checkConnect(SocketAddress from, SocketAddress to)' and
> 'checkAccept(SocketAddress addr)'.
> 
> At the point where a permission check previously would take place, a
> check like this is performed instead:
> 
>      if (System.getSecurityManager() instanceof SocketAuthorization sa) {
>          sa.checkAccept(addr);
>      }
> 
> Any public or protected method with such a check should include
> @throws Javadoc explaining that a SecurityException may be thrown.
> 
> The Permission subclasses previously used specifically by these
> operation sites *may* in this case be deprecated for removal
> immediately or at some point in the future, if desired.
> 
> It is the sole responsibility of the SecurityManager implementer to
> implement the various necessary interfaces, and any third-party
> authorization interfaces that would also be relevant.
> 
> ### Option 2: Retain permission system
> 
> Under this option, the existing authorization checks are mostly
> retained, however, since the SecurityManager class only has a general
> 'checkPermission()' method, the logic previously found in the
> 'SecurityManager' class which expands specific check calls into
> general 'checkPermission()' calls (for example, calls to
> 'checkConnect' for sockets) would necessarily become the
> responsibility of the site of the permission check.  Some work would
> be undertaken
> to refactor this code accordingly.
> 
> With this solution, the corresponding Permission subclasses would be
> retained indefinitely.
> 
> In either case it is the responsibility of the implementer of
> SecurityManager to utilize these checks appropriately for
> authorization decisions, based on whatever factors are deemed
> appropriate, which may include contextual information such as a
> currently-authenticated identity or the call stack, or (for example) a
> context object utilizing the ScopeLocal mechanism.
> 
> ### Other changes
> 
> It would be worth exploring whether the SecurityManager installation
> could be refitted to use the ServiceLoader mechanism (for example at
> first call to getSecurityManager()) based on the class loader of the
> application class or module path.  This would allow the
> 'System.setSecurityManager' method, and support for the corresponding
> system property, to be removed at the end of the term of deprecation.
> 
> Testing
> 
> Neither solution would ease the burden of testing from the JDK quite
> as much as complete removal, of course. The necessary testing for the
> individual checks should be limited to ensuring that the permission
> check calls are happening with correct arguments and that any thrown
> SecurityException is propagated.  The policy for testing
> SecurityManager installation would depend on whether, and to what
> extent, the more recent changes restricting the installation of the
> security manager are reversed.  Other testing issues may arise as
> well.
>