Scheduler

[ioquatix]

This is a work in progress.

• Rename Fiber{} to Fiber.schedule{}.
• Simplify hooks and improve method name consistency. Format is class_method, e.g. io_wait or kernel_sleep.
• Expose scheduler callbacks as functions to aid in correctness, documentation and maintainability.
• Rework wait.c to pass IO instance, and other adjustments where possible.
• Add support for rescheduling fiber and Mutex support.
• Expose io_read as a hook for implementing io_uring (Linux) and IOCP (Windows).
• Add support for ConditionVariable.
• Add support for Queue.
• Rework usage of (some) functions which accept file descriptor arguments.

[ioquatix]

Some notes about io_uring after discussion with @dsh0416 who implemented a scheduler and has started working on a C interface for io_uring.

The most efficient (zero-copy) interface for IO#read must go directly to io_uring. We schedule the full read operation (buffer, offset, length, etc) and it goes to OS. The OS will inform us when the read operation is completed.

This is different from existing Ruby implementation which uses IO#wait_readable in a loop before calling read.

However, in practice both interfaces are critical. For example, libpq for postgres does not expose some way to hook read, just a way to do non-blocking I/O by checking IO#wait_... before calling the libpq functions.

So, we need to support both for high performance I/O.

Therefore, I introduced new optional Scheduler#io_read(io, buffer, offset, length) and Scheduler#io_write(io, buffer, offset, length). This method is experiment and we will try it with io_uring implementation. It should also fit better with Windows IOCP which up until now has been ignored.

[ioquatix]

@nobu @shyouhei can I get your feedback on this?

[ioquatix]

@eregon I have merged your changes into this branch because I want to build on them. Thanks for your effort.

[ioquatix]

@jeremyevans do you believe this code is sufficient for your Sequel pool management?

[jeremyevans]

I believe Fiber-aware Mutex is a necessary condition for Sequel connection pool management to work correctly with multiple concurrent fibers of the same thread. I'm not sure if ConditionalVariable also needs to be Fiber-aware or not. So I'm not sure if this will be sufficient by itself, but it is definitely a step in the right direction.

[ioquatix]

ConditionalVariable internally uses mutex to synchronize state, so it should be okay, but I'll check it. Thanks for your thoughts.

[ioquatix]

I need to check in thread.c the following functions:

rb_thread_sleep_deadly_allow_spurious_wakeup(void);
rb_thread_wait_for(struct timeval time);

and a few others - some are public interface and should go via the same rb_scheduler_kernel_sleep.

[ioquatix]

@nobu what do you think?