Are security manager policies modifiable on the fly? Just wondering if someone decides they want to disallow or allow something, will they need to restart their vms/geode node?
I think Dan pointed this out earlier in the thread, but just wanted to have us consider the original cve that led to the heavy handed deny all method invocations: CVE-2017-9795 Apache Geode OQL method invocation vulnerability Description: A malicious user with read access to specific regions within a Geode cluster may execute OQL queries that allow read and write access to objects within unauthorized regions. In addition a user could invoke methods that allow remote code execution I think Juan's proposal would still allow us to provide multiple solutions that may or may not reopen that hole, but it would be up to the user to decide what they are willing to accept. The choice for what should be default would still be up for debate... On Tue, Jul 2, 2019 at 1:07 PM Jacob Barrett <jbarr...@pivotal.io> wrote: > > > > On Jul 2, 2019, at 11:58 AM, Juan José Ramos <jra...@pivotal.io> wrote: > > > > Hello Jake, > > > > I've been doing some reading about the *Java Security Manager* and, even > > when it might work for our use case, I don't think is a good fit due to > the > > following reasons: > > 1). We already have chosen *Shiro* for authentication and authorization, > > adding yet another security framework (and mapping between roles and > > permissions between the two of them) for OQL method invocation security > > seems overkilling to me. I'd rather spend some time improving the > > *ResourcePermissionBasedMethodAuthorizer > > The security manager doesn’t have to be as fined grained as individual > users. Do we really intend to support actions on objects base on current > user too? > > > [1] *and/or adding *Permissions*/*Roles* to our current security > framework > > than introducing a new security framework altogether. > > Again, we don’t have to add anything. If the user wants to restrict access > to certain activities they could add a policy file to the JVM to do so. It > doesn’t have to be user based. We could simply have two hard coded roles, > UserCode and SystemCode. SystemCode has AllPermissions, UserCode has none. > Now we run all queries under the context of UserCode and it can’t execute > any File, Runtime, Socket, etc. > > > 2). There can only be one *Security Manager* per JVM at a given time > Eh, sorta but no, you can use thread contexts and class loader isolation, > but there really doesn’t need to be more than one. > . > > 3). Customers already using a custom *Security Manager* on their own will > > be in trouble... we can certainly wrote our own implementation as a > > composite and parse multiple policy files, but this will probably break > the > > backward compatibility for these customers (and requires yet more > > configuration and things to keep in mind when upgrading). > Their custom SecurityManager would conform to the same rules as the > default file based one in the JRE. They would be able to control code > access using their existing framework. Seems like win to me. > > > 4). The current set of default *Permissions* (*PropertyPermission*, > > *FilePermission*, etc.) don't apply to our use case, so we'll need to > > create our own implementations and do some plumbing to map to what we > > currently have in terms of principals and roles (more boilerplate). > > How do they not apply? > > > 5). In order to check a permission we basically need to check whether a > > Security Manager is installed (*System.getSecurityManager() != null*) and > > execute the check afterwards (*securityManager**.checkPermission(perm)*); > > the *Security Manager* then looks at the classes of all methods currently > > within the call stack (roughly speaking) and returns or throws an > > exception... a lookup through a simple Map is probably faster, and easier > > to maintain and read. > > All the things we need to be protected from have already been coded with > these checks in the JRE. Do you think we need to protect other activities? > > -Jake > >
