[arch-projects] [netcfg] [PATCH 3/4] Lower latency in timeout_wait

[Henrik Hallberg]

On Tue, Jul 03, 2012 at 01:08:34AM +0200, Jouke Witteveen wrote:
> On Mon, Jul 2, 2012 at 11:07 PM, Henrik Hallberg <henrik at k2h.se>
> wrote:
> > --- src/globals | 3 ++- 1 file changed, 2 insertions(+), 1
> > deletion(-)
> >
> > diff --git a/src/globals b/src/globals index 436334a..62cce3e 100644
> > --- a/src/globals +++ b/src/globals @@ -100,10 +100,11 @@ function
> > checkyesno() { # $2...: condition command function timeout_wait() {
> > local timeout="$1" +    (( timeout *= 10 )) shift while ! eval "$*";
> > do (( timeout-- > 0 )) || return 1 -        sleep 1 +        sleep
> > 0.1 done return 0 } -- 1.7.11.1
> >
> 
> I don't think this is such a good idea. This saves less than .45
> seconds on average per timeout_wait (old expected waste: .5s, new
> expected waste: .05s, added polling overhead).

If you assume uniform distribution, yes. However, this is not the case.
For the cases which I fixed in patch 4, the first eval (no sleep at all)
was sufficient. (Then, of course, the sleep time does not matter). If
not, the second always was.

> If you want to do it > this way, you should probably decouple ((
> timeout-- > 0 )) into (( timeout -= .1 )) and (( timeout > 0 )).

I could have sworn that I tested floating point calculations in bash,
failed, and found texts claiming that it was not possible without e.g.
`bc`. TIL.

> The gain is usually a lot less than the wait (saving .45s on something
> that takes 3s isn't all too exciting) and not seldom, netcfg is an
> interactive script so that half a second doesn't really matter.

Before patch 2, there were two calls each to {start,stop}_wpa. With this
change alone, and the uneven distribution of required time, I typically
saved four seconds -- i.e. the lions share of `netcfg -u` wall time.

> Why would we try to make a bash script super fast?

Why keep unnescessary waiting times in? I'm happier to boot and have my net
connected by the time my shell is. than waiting a little longer. With
this and DHCP_OPTIONS='--noarp', it takes four seconds instead of
twelve.

> Note that even with the 1s sleeps, it can be beneficial to call
> `timeout_wait 1 'whatever'`, as it might return in 0s. This means the
> other patches still make some sense without this one.

They do. I divided this set of patches like I did so you could pick and
choose.

/ H