User:TomYu/Plugin support improvements

This project is targeted at release 1.9.

This is an early stage project for MIT Kerberos. It is being fleshed out by its proponents. Feel free to help flesh out the details of this project. After the project is ready, it will be presented for review and approval.

The MIT Kerberos implementation has a number of places where modular extensibility would facilitate development and deployment of new technologies. Developers of the core code and authors of extension code have an interest in making it easier for third parties to extend the code in a well-compartmentalized way. Administrators of Kerberos deployments and packagers of software have an interest in making it easy to deploy new capabilities in a simple way.

Priorities

Allow third parties to implement multiple plugin modules for each pluggable interface
Allow a plugin module to build as dynamic or built-in from the same source code
Allow third parties to more easily create new plugin modules
Provide a uniform method for configuring discovery of plugin modules
Improve readability of code that calls pluggable interfaces
Allow easier creation of new pluggable interfaces
Allow incremental transition of existing pluggable interfaces to the new framework

Deliverables for release 1.9

Create a plugin framework and pluggable interfaces that can support password strength and password synchronization plugin modules. These should support the capabilities of two existing extensions written by Russ Allbery -- krb5-strength and krb5-sync. The framework is subject to change in the future, so it doesn't have to accommmodate all eventualities, but we will have a goal of not painting ourselves into a corner with respect to reasonably plausible future requirements.

Requirements

thread safety
minimal configuration required

no explicit configuration needed to use built-in modules in their default modes
no explicit configuration needed to use loadable modules if the files containing them are located in the default directory for that kind of module

allow changing of default directory base for loadable plugin modules of all pluggable interfaces
any resources allocated by a plugin module must be released by that module and by no other code

Components

pluggable interface (PLIF)
- caller (consumer) interface
- provider interface
plugin manager (PLM)
plugin registry (PLR)

Pluggable interface

A pluggable interface is an interface, possibly internal to a library, that can be implemented by a third party in a modular, well-compartmentalized manner. These implementations of pluggable interfaces are called plugin modules. Pluggable interfaces allow a consumer to use the capabilities of the interface without needing to be aware of the implementation details. In particular, a pluggable interface prevents the consumer from needing to know whether the module is a built-in or a dynamically loadable module. Pluggable interfaces can be one-to-one, or one-to-many. An example of one-to-one is the DAL, and an example of one-to-many is preauth.

A pluggable interface has two parts: a consumer interface and a provider interface. Typically, library code implements the consumer interface, and application code or other library code calls the functions of the consumer interface. The provider interface is what the plugin module implements.

Caller (consumer) interface

The consumer interface isolates the consumer from implementation details of the pluggable interface. The caller does not generally need to know about whether a given module is built-in or dynamically loaded. The implementation of a consumer interface is essentially a glue layer, and can make use of domain-independent (not specific to any pluggable interface) capabilities of the plugin framework. The consumer might explicitly register a new module that it implements: this capability is part of a plugin registry.

A consumer of a pluggable interface uses an opaque handle, obtained from a loader function that is part of the pluggable interface, to call the methods of a plugin module. Each method of the consumer interface is an ordinary C function that takes the opaque handle either explicitly as its first argument or implicitly by some means such as a module name. Conceptually, these pluggable interface functions are wrapper functions that call through function pointers contained in the opaque plugin module handle object. (though this is not the only possible implementation)

A handle can represent:

the plugin module itself
a resource to which the plugin module provides access (e.g., a ccache handle)
a set of plugin modules (e.g., the set of all available preauth mechanisms)

One rationale for using wrapper functions instead of having the caller directly invoke methods through a function pointer is to make it easier for debuggers and analysis tools to recognize when a particular interface method is being called. (Function pointers might have identifier names that look nothing like the actual name of the function they point to, in addition to enabling confusing aliasing.)

The loader function is specific to the pluggable interface. One reason is for type safety: there will be a distinct opaque handle type for each pluggable interface, allowing compile-time checking to catch some sorts of programming errors. Another reason is backward compatibility: it allows a pluggable interface to support plugin modules that implement an older provider interface.

Provider interface

A plugin module is a unit of code that implements the provider interface portion of a pluggable interface. Plugin modules can be built in or dynamically loaded. Several alternatives exist for the form of the provider interface, but some have significant advantages in allowing the plugin module to use identical source code for both built-in and loadable modules.

A built-in module is a module whose executable code is located within the library shared object or executable program file, or behaves as if it were. (While separate library shared objects that the calling library depends on can contain "built-in" modules for the calling library, this can cause problems with cyclic references.) The distinguishing characteristic of a built-in module is that, as part of program startup, the operating system automatically maps the executable code of the module into the address space of the process that calls it, without any explicit action by the library or program.

A dynamically loaded module is a module whose executable code is located within a file that is distinct from the library or program that calls it. The plugin support framework uses the runtime linker (or equivalent) to explicitly map the executable code of the module into the process address space. In POSIX systems, this is typically done using dlopen().

Loadable module provider interface

A loadable module exports a single function symbol, which denotes the vtable constructor function for that module. This function takes as arguments an interface version number, a pointer to a caller-allocated vtable structure for the interface, and the size of the structure (as an added precaution against version mismatches).

Although the caller (actually the plugin support code) allocates the vtable structure in the above description, one alternative is to have the module perform the allocation of the structure itself. This can cause problems if the module uses a different memory allocator than the caller.

Should the single function symbol have a prefix that depends on the name of the plugin module?

Yes: allows identical module source code to be used for the built-in and loadable module versions of a module
No: adding a prefix to the symbol adds complexity to the code that looks up the symbol

Built-in module provider interface

A built-in module provides the same interface as a loadable module. In the alternative where we use same exported function symbol for each loadable module implementing a pluggable interface, the built-in modules will still need distinct prefixes are required for each vtable retrieval function. This requires slightly different source code for a built-in module versus a loadable module.

Alternative: one symbol per method

Exporting one symbol per method in a loadable module is an alternative that might require less effort from a module author, as it does not require that a vtable be part of the provider interface. Built-in modules would probably still require a vtable (though that could be provided by the implementation of the consumer interface).

Alternative: loader with fixed module set

This approach uses a slightly different definition for "loader". A loader is a function that acts like a plugin registry that knows about a fixed, predetermined set of plugin modules. Each module in this set may implement a different pluggable interface. Typically, a built-in loader function knows about all of the built-in modules, regardless of what interface each module implements. A collection of modules (possibly implementing different pluggable interfaces) can be combined in a single dynamically loaded shared object that only exports the loader function interface. To search for a dynamically loaded plugin module, the plugin framework would load each of these shared objects and call the loader function of each one until it located the desired module.

Plugin manager

The plugin manager (PLM) provides a set of generic support capabilities that are independent of individual pluggable interfaces. It centralizes the discovery process for plugin modules. Typically, consumers of pluggable interfaces do not call it directly, but instead call a loader function of the pluggable interface, which in turn calls the plugin manager. The plugin manager works in concert with a set of plugin registries to locate plugin modules.

The krb5_init_context() functions will create a plugin manager context that will exist in the krb5_context, and add the initial plugin registries to the plugin manager context. An application may use krb5_get_plm() to retrieve the plugin manager context so it can register its own built-in plugin modules.

The plugin manager locates plugin modules using both a numeric identifer that designates a pluggable interface and a string that names a module that implements that pluggable interface. The primary way to use the plugin manager is to query it for the vtable constructor for a specified module (or a set of vtable constructors for all modules of that interface).

As a lower-level interface that supports backward compatibility for older loadable-module provider interfaces, the plugin manager can create an opaque locator handle for a specified module. The plugin manager also supports looking up sets of locator handles (for all modules implementing a specified pluggable interface).

Additional functions in the plugin manager interface allow for lookups of symbols in the plugin module (if supported by that module)

Plugin registry

A plugin registry (PLR) uses pairs of plugin interface identifiers and plugin module names to look up modules. Most of the functionality of a plugin registry is only accessible indirectly through the plugin manager, but some user-level functionality exists for each type of registry, such as an interface to create a new registry or to register a new plugin module in a registry. The primary way to use a registry is to query it for the vtable constructor for a specified module.

A plugin registry can also produce a opaque module locator handle for a specified module. It provides lookup functions so that a loader function of a pluggable interface can do the equivalent of dlsym() on the module represented by a locator handle.

Initially, the plugin framework implementation will contain two types of registry:

built-in module table registry: locates built-in plugin modules. This also allows an application to register its private built-in plugin modules.

loadable module directories registry: locates dynamically loadable plugin modules. It maps plugin interface identifiers to directories where modules implementing that interface can be found, and can thereby locate individual modules.

The krb5_init_context() functions will call various registry methods to create and configure registries for libkrb5 built-in modules and default directories for loadable modules.

Registry of built-in modules

This kind of registry keeps track of built-in modules. Typically, libkrb5 will initialize this with locators for all of the built-in modules that are linked into it. Applications can also register private built-in plugin modules using this type of registry. TO BE DETERMINED: Are libkrb5 registry tables read-only? If they are, applications will probably need to add their own registries to the plugin manager in order to register their private built-in plugin modules.

Registry of loadable modules

This kind of registry keeps track of a few additional items needed for locating loadable modules. This includes a base directory pathname to directory tree of plugin modules. Each subdirectory of the base directory has a name that corresponds to the pluggable interface identifier for the modules that intended to be in that subdirectory.

This registry type has functions to configure the mapping of pluggable interface identifiers to subdirectory names, as well as functions to configure the mapping of individual loadable module files to module names / interface identifier pairs. The krb5_init_context() functions will do this for the pluggable interfaces that are part of libkrb5.

NEEDS MORE DETAIL

Details

Deallocator functions, etc. omitted for conciseness.

NEEDS MORE DETAIL

Plugin manager details

The following functions are meant to be used by a consumer of pluggable interfaces:

k5_plm_init: initialize a plugin manager context
k5_plm_fini: shut down a plugin manager context
k5_plm_add_reg: add a new registry to a plugin manager context

The following functions are meant to be used by a loader function of a pluggable interface:

k5_plm_get_vtfn: get a vtable constructor
k5_plm_get_vtfn_set: get a set of vtable constructors

The following functions are meant to be used by a loader function that needs to do lower-level operations (particularly for backward compatibility with older loadable modules). They are typically not implemented by registries that only track built-in modules:

k5_plm_get_ploc: retrieve a module locator handle
k5_plm_get_ploc_set: retrieve a set of module locator handles
k5_ploc_get_vtfn: get a vtable constructor function from a module locator
k5_ploc_get_data: get a data symbol from a module locator
k5_ploc_get_func: get a function symbol from a module locator
k5_ploc_get_vtfn_set: get a set of vtable constructor functions from a module locator set
k5_ploc_get_data_set: get a set of data symbol from a module locator set
k5_ploc_get_func_set: get a set of function symbol from a module locator set

Old background

for reference

Current plugins

We currently have the following plugin frameworks:

Preauth: All shared objects from profile-specified or installation directory are loaded. Two vtables are read from the shared objects, one for libkrb5 and one for the KDC. The preauth framework iterates over the module list invoking functions to generate or handle preauth data. Preauth vtable functions receive a callback function and data object which allow it to request information such as the expected enctype or FAST armor key for the request.

Authdata: Very similar to the preauth framework.

KDB: The profile specifies a database library name for each realm. Shared objects matching the library name are loaded from a profile-specified and installation directory; the first matching object with an appropriately-named vtable data object is used, and the rest are ignored. libkdb5 contains wrappers which invoke functions in the library's vtable, or (for some optional functions) default implementations if the vtable left the function pointer as NULL.

KDC location: All shared objects from an installation directory are located. A vtable is read from the shared objects. The KDC location framework iterates over each vtable and invokes a lookup function; modules can return success with a location, an error (which halts the location process), or a distinguished error code which passes control along to the next module or the built-in location mechanisms.

GSSAPI: The file /etc/gss/mechs can specify a list of mechanism OIDs and shared object filenames; filenames are taken as relative to an installation directory. Shared objects implementing mechanisms can export either a function returning a vtable, or can export each GSSAPI interface individually.

The following areas of functionality are virtualized but have no exposed plugin framework:

Serialization: Serialization table entries can be registered with krb5_register_serializer. Data objects are matched to table entries by magic number. The registration function is exported by libkrb5 and is named with the krb5_ prefix, but it and its associated structure are declared in k5-int.h rather than krb5.h. It is not used outside of libkrb5.

ccache: Very similar to serialization, except that ccache implementations are selected using a URL-style prefix in the ccache name.

keytab: Very similar to ccache, except that the keytab registration function is used outside of libkrb5 to register a "KDB keytab", which is used by kadmind to serve GSSRPC without requiring a keytab file containing the kadmin keys.

Replay cache: Very similar to ccache, except that the replay cache registration function is not used anywhere (even inside libkrb5).

Plugin frameworks which are "not exposed" may still be productively used by vendor forks of the krb5 tree.

Future plugins

The following areas are candidates for future plugin support:

PRNG
profile / configuration
DNS / host-realm mapping
password quality policy
lockout
audit
password synchronization

Current support infrastructure

In libkrb5support, we have functions to facilitate loading plugins from shared objects. There is a set of functions to load individual plugins from named files and mechglue; these are currently used by the HDB bridge and GSS mechglue:

krb5int_open_plugin - Create a plugin handle from a filename
krb5int_close_plugin - Close a plugin handle
krb5int_get_plugin_data - Retrieve a data object from a plugin handle by symbol name
krb5int_get_plugin_func - Retrieve a function object from a plugin handle by symbol name

There is another set of functions to scan a list of directories for plugins:

krb5int_open_plugin_dirs - Create a plugin dir handle from a list of directories and (optionally) filebases
krb5int_close_plugin_dirs - Close a plugin dir handle
krb5int_get_plugin_dir_data - Retrieve a list of data objects from a plugin dir handle by symbol name
krb5int_get_plugin_dir_func - Retrieve a list of function objects from a plugin dir handle by symbol name
krb5int_free_plugin_dir_data - Free a list of data objects returned by krb5int_get_plugin_dir_data
krb5int_free_plugin_dir_func - Free a list of function objects returned by krb5int_get_plugin_dir_func

Problem areas

Every caller of krb5int_open_plugin_dirs specifies either no filebases (e.g. preauth plugins) or a single filebase (KDB plugins). Accepting and processing a list of filebases is probably needless complexity.

Callers of krb5int_open_plugin_dirs have to know what directories to supply, which means they need to know the krb5 install root as well as the magic plugin area for OS X, and they need logic for reading a profile variable to determine the alternate plugin directory for the test suite (currently only implemented for KDB and preauth plugins).

In most uses of plugins, we read a data object containing a list of function pointers. This makes it mostly impossible to supply a plugin which works with multiple versions of krb5. If we instead read a function object which we invoked with a version number to retrieve the vtable, it would be possible (though perhaps awkward) to create a shared object which works with multiple versions.

We are somewhat schizophrenic about how plugins can access krb5 library functionality, and in particular internal symbols. Sometimes we call functions directly, sometimes we make use of a vtable passed into the plugin (e.g. the preauth_get_client_data_proc function), sometimes we use the accessor to invoke internal functions, and sometimes we call APIs or internal functions directly. Ideally we should have a consistent policy with a sound justification.

When measuring code coverage with gcov, we cannot use shared libraries; this means we need to link in-tree plugins statically into the libraries or programs which load them. We have an ad-hoc method to do this with KDB plugins, but not with other plugin types.

Administrators have an easier time writing scripts than creating linkable shared objects. In some cases it might yield a better administrator experience to create plugin interfaces via subprocesses than loading shared objects, although in many cases this might not be feasible.

In some scenarios such as embedded environments, it may be more useful to allow applications to supply plugin vtables via an API (as we do for keytabs and ccaches, though those APIs are not public) than to load them from shared objects in the filesystem.

Definitions

pluggable interface: an (internal) interface that can be implemented by a third party. These can be one-to-one, or one-to-many. An example of one-to-one is the DAL, and an example of one-to-many is preauth.

module

a unit of code that implements a pluggable interface. It can be built in, or it can be dynamically loadable.

built-in: a module whose executable code is located within the library shared object or executable program file, or behaves as if it were. (While separate library shared objects that the calling library depends on can contain "built-in" modules for the calling library, this can cause problems with cyclic references.) The distinguishing characteristic of a built-in module is that, as part of program startup, the operating system automatically maps the executable code of the module into the address space of the process that calls it, without any explicit action by the library or program.
dynamically loaded: a module whose executable code is located within a file that is distinct from the library or program that calls it. The plugin support framework uses the runtime linker (or equivalent) to explicitly map the executable code of the module into the process address space. In POSIX systems, this is typically done using dlopen().

discovery

process of enumerating what modules are available for a pluggable interface. Includes possible filtering of the raw discovered set.

compiled-in
directory scan
explicit inclusion by configuration
explicit exclusion by configuration

loading

the process of making modules available for calling. This can involve dynamically loading a module using the runtime linker, or it can involve registering a vtable provided by an application.

built-in
dynamic loading
application-registered

selection: the process of a caller invoking one specific module from the set of loaded modules that implement an interface.

consumer interface: the interface that a caller uses to access the services of a pluggable interface. Typically, but not always, the krb5 library implements the consumer interface.

provider interface: the interface that a module author implements

Links

Some possibly useful ideas at Dr Dobb's. It's rather biased toward C++, and we may disagree with some of the details. Among other things, it advocates separating the domain-specific interface of a plugin module from its domain-independent interface: plugin modules get passed a handle for the "registry services" of the plugin manager, and the modules must call those services to register their objects or methods.