Avoid consuming CPU when waiting for input.

[sunfishcode]

When reading input, start by calling event::read, which efficiently blocks until input arrives, rather than polling for input every 100 milliseconds.

This preserves the existing POLL_WAIT polling behavior by moving the poll after the first read call returns.

Fixes #521.

[codecov]

Codecov Report

Merging #651 (0745f64) into main (973dbb5) will increase coverage by 0.00%.
The diff coverage is 0.00%.

@@           Coverage Diff           @@
##             main     #651   +/-   ##
=======================================
  Coverage   49.17%   49.17%           
=======================================
  Files          42       42           
  Lines        7921     7920    -1     
=======================================
  Hits         3895     3895           
+ Misses       4026     4025    -1     

Files	Coverage Δ
src/engine.rs	5.14% <0.00%> (+<0.01%)	⬆️

[fdncred]

I'm excited to see this. Thanks! Is there some way that once can confirm this is working? I've never see excess CPU so I'm not sure of the procedure exactly.

[sunfishcode]

On Linux, for example, build the basic example, install strace, and then run strace target/debug/examples/basic. Before this PR, you'll see it calling epoll_wait 10 times a second, once it starts up. With this PR, it'll start up and then enter a call to epoll_wait that waits until a key is pressed.

[sunfishcode]

Another way to see this on Linux is the powertop program. Before this PR, powertop shows a running basic example process having about 10 events/sec., which doesn't put it at the top of the list on my system, but it is on the list. With the PR, it's not on the list at all.

[fdncred]

The strace trick works as described. I can also see it in powertop as you suggest.

Before

After - i'm guessing these 0.25 events per second are me doing ls twice

I didn't notice anything abnormal with the basic example. It seemed to work as expected.

[sholderbach]

The changes look good to me!

Also saw the reduction in epolls via strace.

I like how the change removes one layer of the loop making things a bit more readable. And having the wait condition directly between the reading and processing of the events makes intuitive sense.