Big issue with IOCP vs. poll variants is that the overlapped
structures have a longer lifetime than intermediate state needed
for epoll. One cannot free overlapped structures after closing a
handle/socket, like one can do with any intermediate state when using
readiness-based IO.
This results in fewer system calls and presumably more effcient code. It
also brings the epoll (and kqueue) code more in line with how the
windows IOCP code works, incidentally.
Make more use of the built in GC code for abstracts to
be sure things are more correct. Issue before was streams could
be freed before IOCP events arrived.
Instead of setting a flag, each interrupt increments an atomic
counter. When the interrupt is finally handled, either by scheduling
code to run on the event loop or executing some out of band code, the
user must now decrement the interrupt counter with
janet_interpreter_interrupt_handled. While this counter is non-zero, the
event loop will not enter the interpreter. This changes the API a bit but
makes it possible and easy to handle signals without race conditions
or scheduler hacks, as the runtime can ensure that high priority code is
run before re-entering possibly blocking interpreter code again.
Also included is a new function janet_schedule_soon, which prepends to
the task queue instead of appending, allowing interrupt handler to skip
ahead of all other scheduled fibers.
Lastly, also update meson default options to include the
interpreter_interrupt code and raise a runtime error if os/sigaction
is used with interpreter interrupt but that build option is not enabled.
Threaded channels _can_ be marshalled, just not for communication
between threads. This is a special case since the same abstract type
is used for both threaded and non-threaded channels.
