Re: [RFC PATCH v1] Replace tasklets with per-cpu implementation.

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Thu, Aug 28, 2014 at 01:39:54PM +0100, Jan Beulich wrote:
> >>> On 27.08.14 at 19:58, <konrad.wilk@oracle.com> wrote:
> > The 'hvm_do_IRQ_dpci' is the on that is most often scheduled
> > and run. The performance bottleneck comes from the fact that
> > we take the same spinlock three times: tasklet_schedule,
> > when we are about to execute the tasklet, and when we are
> > done executing the tasklet.
> 
> Before starting all the work here, did you investigate alternatives
> to this specific use of a tasklet? E.g., it being a softirq one, making
> it have its own softirq?

If I understand you right, you mean implement an tasklet API that
would only been be used by the hvm_do_IRQ_dpci? Its own spinlock,
list, and an seperate tasklet_schedule?

I did think about it a bit a the start, but discarded it since
I figured it would be a no-go upstream - as it is an one-off and seems
hackish.

I can certainly prototype one up and see if it matches the performance
of this implementation if you would like?