[PATCH] Fix drain logic for small number of connections

[Maxim Dounin]

Hello!

On Fri, Dec 30, 2016 at 12:13:01PM -0600, Joel Cunningham wrote:

> > On Dec 29, 2016, at 1:13 PM, Maxim Dounin <mdounin at mdounin.ru> wrote:

[...]

> > I agree though that if connection shortage is occasional the 
> > current behaviour might not be optimal with worker_connections set 
> > to a small value, or there are only a few connections can be 
> > reused.  Instead of gradually reducing keepalive (and post accept) 
> > timeout to a smaller value it closes all keepalive connections 
> > from time to time.
> > 
> > Probably optimal solution would be to close no more than some 
> > fraction of reusable connections at once.  Patch series below.
> 
> I tested out the patch series and I did run into a problem.  
> Looks like the reusable_connections_n increment and decrement 
> are swapped.  The first call to ngx_reusable_connection causes 
> unsigned integer wrap around.  I swapped the increment/decrement 
> and the patch gives the expected behavior of only freeing 1/8 
> (or at least 1) of the reusable connections.

Thanks for noticing this, fixed.

-- 
Maxim Dounin
http://nginx.org/