Channels v2

[ryanolson]

• BufferedChannel
  • Writes will yield if the channel is at capacity
  • Reads will yield if the channel is empty
  • Writes will preferentially resume awaiting readers
  • In the case a reader is resumed from a writer’s execution context, the writers will be rescheduled on scheduler of the caller’s coroutine or the default scheduler
  • Back pressure is managed by transfer the execution context downstream; upstream is only rescheduled for execution when space exists in the buffer.
  • Enabled pipeline concurrency proprotional allowing upstream and downstream to be co-scheduled
• ImmediateChannel
  • Writes will suspend if there are no awaiting Readers
  • Reads will suspend if there are no awaiting Writers
  • Awaiting writers holding data are always processed first
  • Suspended writers are put in a FIFO resume linked-list after their data has been consumed
  • If no incoming (writer) data is available, writers are resumed in FIFO ordering from the resume queue
  • Back pressure is managed by transferring the execution context downstream, pausing upstream progress
  • Does not enable pipeline concurrency as the execution context is transferred.
• RecentChannel
  • Writes will never yield
  • Writes will preferentially resume awaiting readers
  • If the Channel is at capacity, the oldest data at the front of the queue is popped and the newest data is pushed to the back of the queue.
  • Back pressure is managed by dropping oldest data then transferring the execution context downstream with the upstream getting rescheduled
• NullChannel
  • Writes never yield, immediate drops the data sent to it
  • Read never yield, always returns a channel closed

#262 adds mrc::core::concepts::not_void - all channels should use not_void at minimum

#264 adds the first concrete v2 channel based on coroutines

• note: type-erasing concrete implementations of channels requires wrapping them in a generic task which has a generic return type
  • this will add some unknown amount of overhead
  • extend Add ImmediateChannel #264's benchmarks to measure this overhead