This SDK is safe for general use.
However please keep in mind that foobar2000 v1.4 is currently in beta and some of the new features may undergo changes before the final version is out.
Components built with this SDK are compatible with foobar2000 1.3 and 1.4 series. They are not compatible with any earlier versions (will fail to load), and not guaranteed to be compatible with any future versions, though upcoming releases will aim to maintain compatibility as far as possible without crippling newly added functionality.
Note that components built with this SDK can provide foobar2000 1.4 functionality while maintaining compatibility with the 1.3 series.
While this is not recommended, you can alter the targeted foobar2000 API level, edit SDK/foobar2000.h and change the value of FOOBAR2000_TARGET_VERSION.
Currently supported values are 78 (for 1.3 series) and 79 (for 1.4 series).
API 78 components load in foobar2000 1.3 and 1.4; API 79 components load in 1.4 only.
This SDK contains project files for Visual Studio 2017.
Visual Studio versions from 2012 up produce incorrect output with default release settings on foobar2000 code - see forum thread for details. You can mitigate this issue by compiling all affected code (foobar2000 SDK code along with everything that uses foobar2000 SDK classes) with /d2notypeopt option. This is set by default on project files provided with the SDK; please make sure that you set this option on any projects of your own.
If you're aware of a better workaround - such as a source code change rather than setting an obscure compiler flag - please let us know; posting on the forum is preferred for the benefit of other users of this SDK.
By convention, foobar2000 code used to #include every single SDK and utility header in a precompiled header (PCH) file. This became an issue with certain Visual Studio setups, as the PCH payload became extremely large.
The SDK has been changed to reduce the amount of unnecessary code shoved into #included headers; errors of this type are yet to be seen with this version of the foobar2000 SDK compiled under VS2017.
If you run into this error, we recommend the following steps:
Some very old inconsistencies in the code have been cleaned up. Various bit_array classes are now in pfc namespace where they belong. Please use pfc::bit_array and so on in new code. If you have code that references bit_array classes without the pfc:: prefix, put “using pfc::bit_array” in your headers to work around it.
Each new decoder (input_entry) now provides a get_guid() and get_name() to allow user to choose the order in which they're invoked as well as disable individual installed decoders. Because of this, you are no longer supposed to walk input_entry instances in your code; however many existing components do so because this was the way things were expected to work in all past versions before 1.4.
In most cases there's nothing that you need to do about this, unless you have code that talks to input_entry instance methods directly.
The proper way to instantiate any input related classes is to call input_manager; it manages opening of decoders respecting user's decoder merit settings. Calling any of the helper methods to open decoders / read tags / etc in the SDK will call input_manager if available (v1.4) and fall back to walking input_entry services if not (v1.3).
If your component targets API level lower than 79 (this SDK comes configured for 78 by default), all your attempts to walk input_entry services return a shim service that redirects all your calls to input_manager. You cannot walk actual input_entry services.
This is to keep existing components working as intended.
If your component targets API level 79 (which means it won't load on v1.3), the shim is not installed as your component is expected to be aware of the new semantics.
Many new input methods have been added and some are mandatory for all code, in particular reporting of decoder name and GUID; existing code not providing these will not compile. However some of the methods now required to be provided by an input class are mundane and will be left blank in majority of implementations; an input_stubs class has been introduced to help with these. In past SDK versions, your input class would not derive from another class; it is now recommended to derive from input_stubs to avoid having to supply mundane methods yourself.
You can now control which level of decoder API your instance supports from your input class instead of using multi parameter input_factory classes.
The input_stubs class provides these for your convenience:
typedef input_decoder_v4 interface_decoder_t; typedef input_info_reader interface_info_reader_t; typedef input_info_writer interface_info_writer_t;
Override these in your input class to indicate supported interfaces.
For an example, if your input supports remove_tags(), indicate that you implement input_info_writer_v2:
typedef input_info_writer_v2 interface_info_writer_t;
As of version 1.4, foobar2000 is compiled with dynamic Visual C runtime and redistributes Visual C runtime libraries with the installer, putting them in the foobar2000 installation folder if necessary. The benefits of this are: * Smaller component DLLs * Increased limit of how many component DLLs can be loaded.
This SDK comes configured for static runtime by default, for compatibility with foobar2000 version 1.3. If your component is for foobar2000 series 1.4 only, you can switch to using dynamic runtime instead - make sure to change the setting for all projects in your workspace.
tl;dr if you don't know what this is about you probably don't care and your component isn't in any way affected by this.
The way service_query() is implemented has been redesigned for performance reasons - the old way ( implemented per interface class via FB2K_MAKE_SERVICE_INTERFACE macro ) resulted in every single class_guid from the entire SDK being included in your DLL, due to strange behaviors of Microsoft linker.
The service_query() function itself is now provided by service_impl_* classes, so it's materialized only for services that your code implements and spawns.
The default implementation calls a newly added static method of handle_service_query(), implemented in service_base to check against all supported interfaces by walking derived class chain.
If you wish to override it, create a function with a matching signature in your class and it will be called instead:
class myClass : public service_base { public: static bool handle_service_query(service_ptr & out, const GUID & guid, myClass * in) { if ( guid == myClassGUID ) { out = in; return true; } return false; } };
The above change to service_query() implementation rules obviously breaks any existing code where one class inherits from multiple service classes and supplies a custom service_query().
While using multi inheritance is not recommended and very rarely done, a new template has been added to help with such cases: service_multi_inherit<class1, class2>.
You can use it to avoid having to supply service_query() code yourself and possibly change it if service_query() semantics change again in the future.
foobar2000 version 1.3 uses different metadb behavior semantics than older versions, to greatly improve the performance of multithreaded operation by reducing the time spent within global metadb locks.
Any methods that:
.. are now marked deprecated and implemented only for backwards compatibility; they should not be used in any new code.
It is recommended that you change your existing code using these to obtain track information using new get_info_ref() style methods for much better performance as these methods have minimal overhead and require no special care when used in multiple concurrent threads.
Each component must link against:
Optionally, components can use helper libraries with various non-critical code that is commonly reused across various foobar2000 components:
Foobar2000_SDK, foobar2000_component_client and PFC are included in sourcecode form; you can link against them by adding them to your workspace and using dependencies. To link against shared.dll, you must add “shared.lib” to linker input manually.
Component code should include the following header files:
A component is a DLL that implements one or more entrypoint services and interacts with services provided by other components.
A service type is an interface class, deriving directly or indirectly from service_base
class. A service type class must not have any data members; it can only provide virtual methods (to be overridden by service implementation), helper non-virtual methods built around virtual methods, static helper methods, and constants / enums. Each service interface class must have a static class_guid
member, used for identification when enumerating services or querying for supported functionality. A service type declaration should declare a class with public virtual/helper/static methods, and use FB2K_MAKE_SERVICE_INTERFACE()
/ FB2K_MAKE_SERVICE_INTERFACE_ENTRYPOINT()
macro to implement standard service behaviors for the class; additionally, class_guid
needs to be defined outside class declaration (e.g. const GUID someclass::class_guid = {….};
). Note that most of components will not declare their own service types, they will only implement existing ones declared in the SDK.
A service implementation is a class derived from relevant service type class, implementing virtual methods declared by service type class. Note that service implementation class does not implement virtual methods declared by service_base; those are implemented by service type declaration framework (service_query
) or by instantiation framework (service_add_ref
/ service_release
). Service implementation classes are instantiated using service_factory
templates in case of entrypoint services (see below), or using service_impl_t
template and operator new
: ”myserviceptr = new service_impl_t<myservice_impl>(params);
”.
Each service object provides reference counter features and (service_add_ref()
and service_release()
methods) as well as a method to query for extended functionality (service_query()
method). Those methods are implemented by service framework and should be never overridden by service implementations. These methods should also never be called directly - reference counter methods are managed by relevant autopointer templates, service_query_t
function template should be used instead of calling service_query
directly, to ensure type safety and correct type conversions.
An entrypoint service type is a special case of a service type that can be registered using service_factory
templates, and then accessed from any point of service system (excluding DLL startup/shutdown code, such as code inside static object constructors/destructors). An entrypoint service type class must directly derive from service_base
.
Registered entrypoint services can be accessed using:
service_enum_t<T>
template, service_class_helper_t<T>
template, etc, e.g. service_enum_t<someclass> e; service_ptr_t<someclass> ptr; while(e.next(ptr)) ptr->dosomething();
playlist_manager
- using someclass::get()
, e.g.: auto api = someclass::get(); api->dosomething(); api->dosomethingelse();
someclass::g_dosomething()
- those use relevant service enumeration methods internally.
An entrypoint service type must use FB2K_MAKE_SERVICE_INTERFACE_ENTRYPOINT()
macro to implement standard entrypoint service behaviors, as opposed to all other service types that use FB2K_MAKE_SERVICE_INTERFACE()
macro instead.
You can register your own entrypoint service implementations using either service_factory_t
or service_factory_single_t
template - the difference between the two is that the former instantiates the class on demand, while the latter keeps a single static instance and returns references to it when requested; the latter is faster but usable only for things that have no per-instance member data to maintain. Each service_factory_t
/ service_factory_single_t
instance should be a static variable, such as: ”static service_factory_t<myclass> g_myclass_factory;
”.
Certain service types require custom service_factory
helper templates to be used instead of standard service_factory_t
/ service_factory_single_t
templates; see documentation of specific service type for exact info about registering.
A typical entrypoint service implementation looks like this:
class myservice_impl : public myservice { public: void dosomething() {....}; void dosomethingelse(int meh) {...}; }; static service_factory_single_t<myservice_impl> g_myservice_impl_factory;
Additional methods can be added to any service type, by declaring a new service type class deriving from service type class you want to extend. For example:
class myservice : public service_base { public: virtual void dosomething() = 0; FB2K_MAKE_SERVICE_INTERFACE_ENTRYPOINT(myservice); }; class myservice_v2 : public myservice { public: virtual void dosomethingelse() = 0; FB2K_MAKE_SERVICE_INTERFACE(myservice_v2, myservice); };
In such scenario, to query whether a myservice instance is a myservice_v2
and to retrieve myservice_v2
pointer, use service_query
functionality:
service_ptr_t<myservice> ptr; (...) service_ptr_t<myservice_v2> ptr_ex; if (ptr->service_query_t(ptr_ex)) { /* ptr_ex is a valid pointer to myservice_v2 interface of our myservice instance */ (...) } else {/* this myservice implementation does not implement myservice_v2 */ (...) }
When performing most kinds of service operations, service_ptr_t<T>
template should be used rather than working with service pointers directly; it automatically manages reference counter calls, ensuring that the service object is deleted when it is no longer referenced.
For convenience, all service classes have myclass::ptr
typedef'd to service_ptr_t<myclass>
.
When working with pointers to core fb2k services, just use C++11 auto
keyword and someclass::get()
, e.g. auto myAPI = playlist_manager::get();
Most of API functions use C++ exceptions to signal failure conditions. All used exception classes must derive from std::exception
(which pfc::exception
is typedef'd to); this design allows various instances of code to use single catch()
line to get human-readable description of the problem to display.
Additionally, special subclasses of exceptions are defined for use in specific conditions, such as exception_io
for I/O failures. As a result, you must provide an exception handler whenever you invoke any kind of I/O code that may fail, unless in specific case calling context already handles exceptions (e.g. input implementation code - any exceptions should be forwarded to calling context, since exceptions are a part of input API).
Implementations of global callback services such as playlist_callback
, playback_callback
or library_callback
must not throw unhandled exceptions; behaviors in case they do are undefined (app termination is to be expected).
In order to create your entries in the configuration file, you must instantiate some objects that derive from cfg_var
class. Those can be either predefined classes (cfg_int
, cfg_string
, etc) or your own classes implementing relevant methods.
Each cfg_var
instance has a GUID assigned, to identify its configuration file entry. The GUID is passed to its constructor (which implementations must take care of, typically by providing a constructor that takes a GUID and forwards it to cfg_var
constructor).
Note that cfg_var
objects can only be instantiated statically (either directly as static objects, or as members of other static objects). Additionally, you can create configuration data objects that can be accessed by other components, by implementing config_object
service. Some standard configuration variables can be also accessed using config_object
interface.
Multiple service classes presented by the SDK allow your component to receive notifications about various events:
All of global callbacks operate only within main app thread, allowing easy cooperation with windows GUI - for an example, you perform playlist view window repainting directly from your playlist_callback implementation.
There are restrictions on things that are legal to call from within global callbacks. For an example, you can't modify a playlist from inside a playlist callback, because there are other registered callbacks tracking playlist changes that haven't been notified about the change being currently processed yet.
You must not enter modal message loops from inside global callbacks, as those allow any unrelated code (queued messages, user input, etc.) to be executed, without being aware that a global callback is being processed. Certain global API methods such as metadb_io::load_info_multi or threaded_process::run_modal enter modal loops when called. Use main_thread_callback service to avoid this problem and delay execution of problematic code.
You should also avoid firing a cross-thread SendMessage() inside global callbacks as well as performing any operations that dispatch global callbacks when handling a message that was sent through a cross-thread SendMessage(). Doing so may result in rare unwanted recursions - SendMessage() call will block the calling thread and immediately process any incoming cross-thread SendMessage() messages. If you're handling a cross-thread SendMessage() and need to perform such operation, delay it using PostMessage() or main_thread_callback.
This chapter describes things you should keep on your mind when designing your own service type classes. Since 99% of components will only implement existing service types rather than adding their own cross-DLL-communication protocols, you can probably skip reading this chapter.
It is important that all function parameters used by virtual methods of services are cross-DLL safe (do not depend on compiler-specific or runtime-specific behaviors, so no unexpected behaviors occur when calling code is built with different compiler/runtime than callee). To achieve this, any classes passed around must be either simple objects with no structure that could possibly vary with different compilers/runtimes (i.e. make sure that any memory blocks are freed on the side that allocated them); easiest way to achieve this is to reduce all complex data objects or classes passed around to interfaces with virtual methods, with implementation details hidden from callee. For an example, use pfc::string_base&
as parameter to a function that is meant to return variable-length strings.
When designing an entrypoint service interface meant to have multiple different implementations, you should consider making it possible for all its implementations to use service_factory_single_t
(i.e. no per-instance member data); by e.g. moving functionality that needs multi-instance operation to a separate service type class that is created on-demand by one of entrypoint service methods. For example:
class myservice : public service_base { public: //this method accesses per-instance member data of the implementation class virtual void workerfunction(const void * p_databuffer,t_size p_buffersize) = 0; //this method is used to determine which implementation can be used to process specific data stream. virtual bool queryfunction(const char * p_dataformat) = 0; FB2K_MAKE_SERVICE_INTERFACE_ENTRYPOINT(myservice); }; (...) service_ptr_t<myservice> findservice(const char * p_dataformat) { service_enum_t<myservice> e; service_ptr_t<myservice> ptr; //BOTTLENECK, this dynamically instantiates the service for each query. while(e.next(ptr)) { if (ptr->queryfunction(p_dataformat)) return ptr; } throw exception_io_data(); }
.. should be changed to:
//no longer an entrypoint service - use myservice::instantiate to get an instance instead. class myservice_instance : public service_base { public: virtual void workerfunction(const void * p_databuffer,t_size p_buffersize) = 0; FB2K_MAKE_SERVICE_INTERFACE(myservice_instance,service_base); }; class myservice : public service_base { public: //this method is used to determine which implementation can be used to process specific data stream. virtual bool queryfunction(const char * p_dataformat) = 0; virtual service_ptr_t<myservice_instance> instantiate() = 0; FB2K_MAKE_SERVICE_INTERFACE_ENTRYPOINT(myservice); }; template<typename t_myservice_instance_impl> class myservice_impl_t : public myservice { public: //implementation of myservice_instance must provide static bool g_queryformatfunction(const char*); bool queryfunction(const char * p_dataformat) {return t_myservice_instance_impl::g_queryfunction(p_dataformat);} service_ptr_t<myservice_instance> instantiate() {return new service_impl_t<t_myservice_instance_impl>();} }; template<typename t_myservice_instance_impl> class myservice_factory_t : public service_factory_single_t<myservice_impl_t<t_myservice_instance_impl> > {}; //usage: static myservice_factory_t<myclass> g_myclass_factory; (...) service_ptr_t<myservice_instance> findservice(const char * p_dataformat) { service_enum_t<myservice> e; service_ptr_t<myservice> ptr; //no more bottleneck, enumerated service does not perform inefficient operations when requesting an instance. while(e.next(ptr)) { //"inefficient" part is used only once, with implementation that actually supports what we request. if (ptr->queryfunction(p_dataformat)) return ptr->instantiate(); } throw exception_io_data(); }