I have been thinking now about IPC and such between processes and system calls. Most processes will need to interact with the kernel process by calling command and also potentially will need to read and write data in a memory mapped means. Basically I want it to be fast. The input stream and output stream and the communication buffer works, however there is latency and it requires that the data be serialized back and forth from the processes. This works for the most part, but it can get a bit iffy in a way. Most of the overhead would be spent deserializing and such. I think what would be faster on the kernel end is if there was a system call interface based on specific indexes, perhaps an enumeration of system calls. There of course would need to be a way to reverse the index of a system call so it is known what there is. This can be done with all services for the most part. Basically instead of some way over the top abstract service system it can instead be simplified to just being method calls for the most part. The main issue though is that system calls would have to be done with interfaces and connectors for those interfaces. Pretty much it will be similar to the socket code in a way however, it will be mostly static. So for instance say there is something that is implemented as a remote system call that provides an interface. It could go for any interface really, although it would be nice if there were default interfaces I could use. But anyway I can use any interface in which case: List. Of course it would work with generics in a way but the object type would need to be an interface. Basically the system call could only pass interfaces, boxed types, arrays, and primitive types. It sort of acts like a gigantic set of proxy methods and objects for example. It cannot work with concrete classes though, which is the main thing. But, there could be a special handle interfaces and additionally factories which can be used to initialize interfaces. So really the client can do what I already similarly have: client side factories. Something that knows how to create the interfaces for the client. Each interface though for a unique object would need a handle so that way it is known what maps to what though, which could become complicated.


But my goal would be to just have like a single method for the most part:

SystemCall.call(int __id, Object... __args);

I suppose it would be simple like POSIX for the most part. Do I really need super complex interfaces and service protocols to do things? There are permissions and such to consider too though. But essentially I want to make it easy. Also there could be an abstract SystemCall class but then one which exposes all of the system calls with the correct arguments too. But the primary goal of this is to make it so that the code is simple and there are no complex things to consider. The streaming stuff is nice, but it is just going to be too slow and have tons of overhead for the most part anyway.


Also, when it comes to sending byte arrays, they could be transformed into a special kind of ByteBuffer but not like the one there, more like a ByteArray that way it can get and set accordingly. At least when sending things on remote ends to the kernel, it will map accordingly.


There would be services on the kernel side, but they would just associate themselves with system calls and then they would be forwarded the call information as needed (which task made the call for example). When it comes to Java SE, this could still be wrapped in the input and output streams but it would only really do system calls. On a real compiled environment it would be more smart and allow for a more direct means of sharing data.


At least with this system call code, I can have it where the client can perform some special things without needing to interact with the kernel. So for example to get the identity hash code of an object. Instead of it going through the kernel or some special interface it can just access internal things directly. Well the identity hash code is a bad example, but say I wanted to get the raw length of a given array, it does not have to go through the kernel. So really it is a special call. But basically, I will have it where there can be local overrides for a system call which are never passed to the kernel. This would make the Java SE and other interfaces easier to implement without requiring magic at all.


I really want to work on the JIT now and write to some binary format. What if for speed I just assume that the classes are fully verified and the imports and such are correct for the most part. Well thinking about volatiles the main thing I always worried about is that the base class could change and the volatiles would be broken. But the thing is if the JIT has cached libraries that were recompiled then it would detect (I would hope anyway) that the dependency has changed and thus recompile it. And I am going to enforce that dependencies not be replaced as such. If the program unique ID changes then the program will need to be recompiled anyway.


So I really just need an interface which stores source class information. But for the JIT I can just have direct access to the source class files. I could write some compact exported JIT fields, or I can just store the original class files which potentially have been stripped. Even then I could pack200 the JARs anyway and cache them. I do not really need to store the original JAR in any way only the classes and resources matter.


Pack200 is quite a squashed and complex format however, but it is a good way to store JAR files in a packed way to reduce disk space until the JARs themselves are needed for classes for JIT purposes.


But, my own Pack200 could just operate on the class file that I have myself. I could also have an export for the class file format too.


But definitely for now, if there is a JIT I will just assume that there is the original JAR or something that contains all the class information for JIT purposes. That will simplify everything and I do not have to worry about constants or volatiles. I just need a means to uniquely identify classes but at least in the initial stage I need not worry about that.


Actually when it comes to the library manager, there will need to be a basic one which is rather minimal. Then there will need to be one which can perform JIT operations.