Re: Implement 3-level event channel routines in Linux.

[Wei Liu <Wei.Liu2@citrix.com>]

On Wed, 2013-01-02 at 21:12 +0000, Konrad Rzeszutek Wilk wrote:
> On Wed, Jan 02, 2013 at 06:46:58PM +0000, Wei Liu wrote:
> > On Wed, 2013-01-02 at 18:26 +0000, Konrad Rzeszutek Wilk wrote:
> > > On Mon, Dec 31, 2012 at 06:38:54PM +0000, Wei Liu wrote:
> > > > This patch series implements 3-level event channel routines in Linux kernel.
> > > > 
> > > > My thought is that 3-level event channel is only useful for Dom0 or driver
> > > > domain, so it is not enabled by default. Enable it with evtchn_level=3 in
> > > > kernel command line.
> > > 
> > > Could it be enabled dynamically? Say when we are close to exhausting the
> > > amount of channels? Or if the machine has some large amount of memory and
> > > hence would probably allocate many many guests?
> > > 
> > 
> > Do you mean enabling this mechanism on the fly when Dom0 / driver domain
> > is running? Or do you mean enabling it based on some metrics when
> > starting up a domain?
> 
> Either one.
> > 
> > If it is the first case, who is responsible for initiating the switching
> > process? It looks like it is Xen's responsibility to do the switch
> > because Dom0 sometimes cannot know the true capability of underlying
> > hardware. If it is Xen, then we need to add callback to poke Dom0 /
> > driver domain.
> 
> Huh? It would be as simple as Dom0 just figuring out that it is
> at the end of the available events (b/c its close to the MAX) and
> then transitioning to the 3-level one (if it can).
> 

We need to manipulate various data structures. IMHO this makes
implementation tricky and racy.

> > 
> > For both cases, what do you have in mind for the metric that used to
> > trigger the switch? How much RAM / how many CPUs is the threshold?
> 
> No idea. At what point are we running out of the events?

In practice, if you have hundreds or thousands of guests running you're
likely of running out of event channels. I presume host capable of doing
this has lots of RAM and CPUs...


Wei.