Checking for I/O events

An I/O event in ERTS is any event triggered by a file descriptor on Unix or any OBJECT HANDLE that can be passed to WaitForMultipleObjects on Windows. The check I/O infrastructure is used by linked-in drivers through driver_select and by NIFs through enif_select.

The main user of the check I/O subsystem is network communication through gen_tcp, gen_udp, gen_sctp and socket on Unix (on Windows socket used its own internal check I/O implementation based on completion ports). It is also used by various other parts, such as when doing os:cmd/1 or reading from the terminal.

This document gives an overview of how the check I/O subsystem works.

The check I/O subsystem consists of a platform specific (erl_poll) and platform agnostic part (check_io).

erl_poll is the basic mechanisms for checking if any events have been signalled and allows waiting for a timeout if needed. The implementation of polling is very platform specific and lives in erts/emulator/sys/common/erl_poll.c for Unix and erts/emulator/sys/win32/erl_poll.c for Windows.

check_io is the cross-platform part of the check I/O subsystem that makes sure that erl_poll has the correct state and dispatches events to the correct entity. The implementation can be found in erts/emulator/sys/common/erl_check_io.c.

check_io is then used by ports and NIFs to listen to events. Ports are communicated to through port signals and are delivered through the ready_input and ready_output callbacks. NIFs get an Erlang message whenever an event is triggered.

Polling

The polling subsystem basically has two API functions; erts_poll_control and erts_poll_wait. erts_poll_control is used to update a ErtsPollSet and erts_poll_wait is used to wait for an event in the pollset to be triggered or a timeout to happen (the timeout can be 0 if just checking). Only a single thread usually calls erts_poll_wait at the same time, but multiple threads calls erts_poll_control at any time.

The implementation of Unix and Windows are completely different as Windows does not really have a concept of "polling" for an event.

Polling on Unix

The poll implemention on Unix support a variety of different polling mechanisms. At the writing of this document they are: epoll (Linux), kqueue (MacOS + *Bsd), /dev/poll (Solaris), poll and select. epoll+kqueue+/dev/poll are referred to as "kernel polling" methods, as the information about which FDs are currently monitored lives in the OS kernel, while poll+select are "user polling" methods as the caller needs to supply all FDs of interest to the kernel everything erts_poll_wait is called.

By default all Unix'es use a "kernel polling" method, but has a fallback pollset that uses "user polling" for FDs that the "kernel polling" mechanism does not support (for example the stdin FD on Linux cannot be monitored by epoll).

As the kernel polling methods have their monitoring information in the kernel it is possible to update these in parallel and without waking the thread that is currently waiting for events. For user polling a queue of needed updates is managed by each pollset and whenever an update is done to that queue the thread waiting on events is woken up to update the set of file descriptors it is waiting on.

When using kernel polling it is possible to have multiple poll threads (using the +IOt flag) that read events from the same pollset. This can be useful for very busy systems with many many FDs that become active alot. If the kernel you are using is not very good at allowing multiple threads to check in the same pollset (this primarily applied to old versions of Linux), then it is also possible to configure erts to use separate pollsets for each pollthread ((using the +IOp flag)).

When an event is triggered it is removed from the pollset and needs to be re-enabled before any new events are triggered. If ONESHOT or equivalent is available then kernel polling uses that flag, otherwise erl_poll will update the pollset as the event is triggered.

Polling on Windows

Polling on Windows uses similar mechanism to "user polling" on Unix, except that because WaitForMultipleObjects is limited to wait for 64 handles it also needs to manage a thread pool. New threads are created as needed, so if the system only ever listens for events on less then 64 handles only 1 thread will be created, but as the number of concurrent handles grow more and more threads will be created.

The thread pool is never shrunk, so if the system at any point uses 1000 handles, there will forever be 16 threads in the thread pool.

Check I/O

Checking for I/O is done by dedicated polling threads. By default, one thread will always be waiting for I/O events using default polling method and the "aux" thread will be waiting in the fallback pollset if such exists.

When an event is triggered it is dispatched to the correct port or process depending on whether it is a driver or nif that has requested the event. As the pollsets use ONESHOT, the event is disabled until the port/NIF registers a new interest in the event.

When you do a driver_select in a linked-in driver, that select will be active until it is disabled. Because of this we need to insert the FD back into the pollset when a driver_select event has been handled. This is done by the port re-inserting the FD in the pollset after a ready_input/ready_output event is called. For NIFs you need to call enif_select for each event that you want, so no such mechanism needs to exist for NIFs.

Scheduler pollset

For very active FDs the fact that we need to re-insert events each time they trigger can lead to quite a lot of overhead. Because of this there is an optimization that places FDs that are never deselected into a special pollset managed that is not checked by the poll threads, but instead checked by the normal schedulers. In this pollset, the FDs no longer use the ONESHOT mechanism, instead they trigger as soon as there is data. For this to work, and not re-trigger on FDs before the port/nif has handled the event, there is a global counter called erts_port_task_outstanding_io_tasks that is incremented for each FD that is dispatched from the scheduler pollset. That counter is then decremented as the FDs are handled by the ports/processes that have subscribed to the event. When it reaches 0, we know that it is safe to check for new events. This increases the latency for how quickly we check for new events by a bit, but drastically reduces the CPU usage for very active FDs.