Linux inotify poller wastes an OS thread

[WGH-]

Every Watcher instance does unix.EpollWait in a loop inside a goroutine. Since this is direct syscall, which circumvents Go runtime poller, it wastes entire OS thread. If you create N watchers, N OS threads will be doing nothing except waiting in epoll_wait syscall.

This can be theoretically fixed, because both inotify fd and even epoll fd are pollable.

There's even convenient method in internal/poll (https://golang.org/pkg/internal/poll/#FD.RawRead) which does what we want. However, internal packages are not acessible from third-party code.

I'm not sure it's even fixable in current Go. There're many related issues in the Go bug tracker which are not yet resolved (probably not a complete list): golang/go#15021 golang/go#24331 golang/go#22939

[WGH-]

It's already fixable in modern Go, by the way.

[horahoradev]

I'm working on this. I currently have a basic implementation.

Essentially, I:

• deleted all of the epoll/poller code
• set inotify's FD to nonblocking, added it to the runtime IO poller with os.NewFile
• swapped out the poller.Wait() call for .Read(), which will return early if the file is closed

I'll run through this again tomorrow to see if either I have any flaws in my logic or anything can be improved.

[arp242]

Fixed with #434