Spunky: A kernel using Ada - Part 1: RPC
In this series of articles I'll illustrate a hobby project of mine that is trying to create a kernel for Genode written in Ada 2012. This project is not about writing a kernel from scratch but rather successively take parts from the existing base-hw kernel and translate them to Ada. Thus, the design mainly follows the approach taken with base-hw. To be able to test the already translated parts I link them together with the remaining parts from base-hw. The interfacing between the Ada and the C++ parts is done on the level of object methods. Over time, the code-base of the new kernel will become more and more Ada and less C++. Maybe later this work will also lead to some formal verification with SPARK but for now, I'm happy with Ada. So let's go!
Before thinking about any Ada translations, I had to enable the compilation and execution of Ada code in the kernel. So far, Ada was used only in user-land components where we can simply link against the spark library for run-time support. This isn't an option in the kernel at least for the fact that the library is shared. So, I wrote a little Ada "Hello World" for base-hw and tried to execute it from within the kernel. For a start, I decided to put the Ada sources into a new sub-directory of base-hw (src/core/ada) and postpone the work of creating a new repository. To make it work, I started adding Ada-run-time ingredients from the spark library and library import to the core-hw library. I had to adapt some paths in the new dependencies as they could no longer implicitly refer to their original repository (libports). Within a short time, I had it working.
As the first subject for my translation, I choose the RPC state machine of base-hw, called "IPC node". The C++ implementation of this unit has a pretty good abstraction against the rest of the kernel and its model and interface aren't too big.
As a first step I reviewed the C++ implementation and recognized several design details that I wanted to change in order to make it fit better into the Ada world. I did these adaptions beforehand in the working implementation, so, they would not add up to the error potential of the translation process.
On thing was that IPC node had virtual methods that were overridden by the thread class that inherited from IPC node. Admittedly, I don't know whether virtual methods in Ada which are overridden in C++ are impossible. But at least they would have complicated things plus, in my experience, reducing the use of inheritance isn't a bad thing. Thus, IPC node became a member of the thread class and holds a reference to its thread object in order to call the previously virtual methods.
Another annoying fact was that IPC node also did the transmission of the message content and its attached capabilities from on protection domain to another. This implied that IPC node needed to know the UTCB layout, the protection-domain interface, and the capability management. All this could be avoided without big modifications by making it a concern of the thread class and let IPC node concentrate merely on the state machine behind RPC.
The third adaption arose from the two facts that Ada procedures don't have return values and that, for better compatibility with SPARK, I don't want to have functions with writable arguments. This means, that a sub-program can't modify it's arguments and have a return value at the same time. But this was the case in the original IPC node. I solved this by replacing the original method with two new ones - one that does the modifications and a one that returns the desired value - and made clear through their names, in which order to call them.
After all this preparation, I started re-implementing the IPC node semantics in a new Ada package IPC_Node. The object layout is represented by a private record type IPC_Node.Object_Type. Constructor and destructor are represented by the procedures IPC_Node.Initialize_Object and IPC_Node.Deinitialize_Object.
The first argument of each Ada sub-program representing a C++ method is the object. When the C++ method is constant, the first argument of the Ada sub-program is "in", otherwise "in out". When the C++ method stores a reference to the object, the first argument of the Ada sub-program is a "not null access" to the object type. This avoids problems with different access- type scopes.
As IPC node internally also uses a queue of IPC nodes, I did a re-implementation of this class (the private package IPC_Node.Queue) using the same pattern as mentioned above. I tried to implement the queue also as generic package to be able to re-use it for other item types but it didn't work out yet.
Wherever IPC node calls a method on its thread object, I just skipped the call and instead left a comment in the place to remind me to add the call later. This way, I was able to skip the details of the C++-Ada interfacing completely during the re-implementation step.
One interesting detail of the re-implementation is the sub-program IPC_Node.For_Each_Helper. In C++, this is a function template that takes a lambda function and applies it to multiple IPC nodes. The lambda function, in fact, is always a method of the thread class. Surprisingly for me, this can be works just fine using sub-program access types to in Ada although the accessed sub-program is a C++ method.
Once, the Ada IPC node was compiling - which was quite an effort as I treat warnings as errors and have style warnings enabled - I wanted to integrate it into the rest of the base-hw kernel to see whether it works. To keep things as clear as possible I put the C++ glue-code in extra packages with names all starting with "CPP". For instance, the package IPC_Node is accompanied by a package CPP_IPC_Node which has the same public interface but with C++-compliant types. The sub-programs of CPP_IPC_Node translate from the C++-compliant types to types natural to Ada and then call the corresponding sub-program in IPC_Node. They can later also be used to fulfill preconditions of the IPC_Node package.
In the C++ world, I replaced the original IPC node with a place-holder that still declares the public IPC-node interface but lacks implementations. It also contains no member variables except a byte array that bloats the corresponding objects to the size needed in the Ada world. The sub-programs of CPP_IPC_Node are then exported to the symbols of the place-holder methods in C++. By doing this, Ipc_node objects can be used in the C++ world as usual while the implementation stays in Ada. In Ada, on the other hand, I don't have to deal with memory management because a reference to a sufficient object space (the byte array) is always handed in from the C++ side. All packages are therefor pure.
The last thing missing was the integration of the thread calls I postponed earlier. I added them through a new package CPP_Thread, whose public interface declares all the sub-programs called by IPC_Node but imports the implementations from the C++ thread class.
To make the whole interfacing a bit safer, I finally added an initial check assert_valid_object_size to the C++ world, which uses the Ada sub-program CPP_IPC_Node.Object_Size to ensure that the place-holder size is sufficient.
What should I say? Thanks to the almost pedantic need for correctness of the Ada compiler and the sheer endless chain of complains it kept throwing at me, the final image worked out of the box and put a big smile on my face :-)
To be continued...