Question regarding the behavior when O_NONBLOCK is absent

[Peter Hutterer]

On Sat, Jun 04, 2022 at 10:59:07AM +0900, Nayan Deshmukh wrote:
> Hello,
> 
> The libevdev requires that the fd should be opened with the O_NONBLOCK
> flag. What kind of behaviour can we expect if we try to create a
> libevdev device without the O_NONBLOCK flag? The main question is
> whether it will be blocked or it could be undefined behaviour as well.
> 
> A little background, I have written libevdev bindings in rust [1] and
> I wanted to document the kind of behaviour people should expect and
> whether I should mark the function as unsafe if it were to exhibit
> undefined behaviour.

no worries there, it's well-defined behaviour, but you should use the
LIBEVDEV_READ_FLAG_BLOCKING  flag in libevdev_next_event(). Reason being
that libevdev tries to read more events from the fd to cache them whenever
you call that function [1], which means if the fd may block even when more
events are already cached and ready to be returned. With that flag you
disable that part, so reading only blocks when there are no more events
cached and we really need to read from the fd.

IOW, the code is (very) roughly like this:

	def libevdev_next_event(self, flags):
	   if self.queue.is_empty() || flags != READ_FLAG_BLOCKING:
		   self.queue.append_events_from_fd()

	   return self.queue.pop(0)

Cheers,
   Peter

[1] goal is to avoid SYN_DROPPED even when processing of any specific event
takes a bit longer