Event driven network I/O

The basic idea is that it should be the read/write operations that registers for select/poll/whatever notification, and not the other way around.

To do this, all I/O functions needs to be turned into asyncronous operations with completetion callbacks.

Benefits

• Much cleaner I/O event loop
• Less CPU overhead for networking I/O
• The gory details of non-blocking I/O centralized and hidden away from the protocols -> less code dupliaction

Notes

• A cbdata type should be used instead of a unix FD. This struct contains all the gory details about current handlers, arguments, pending I/O and such.
• The global table only contains references to the current cbdata registered struct for a UNIX fd (i.e. lookup from UNIX fd to "filehandle" struct)
• User supplied callback data MUST be cbdata registered
• I/O data MUST be refecence counted buffers

The reference counted buffers needs to keep

• reference count, starts at 1 when the buffer is allocated, and the buffer is freed when the reference count reaches zero.
• amount of data currently in the buffer
• actual (allocated) size of the buffer
• the actual buffer itself

Why a cbdata type is required for the filehandle:

1. if there is pending callbacks when a handle is closed
2. to 100% be able to detect if the caller reuses a closed handle, either directly or indirectly via a pending callback.
3. fresh start for each new handle