Re: [PATCH v3 11/18] dmaengine: idxd: ims setup for the vdcm
From: Raj, Ashok
Date: Fri Oct 09 2020 - 09:40:11 EST
On Fri, Oct 09, 2020 at 10:12:18AM -0300, Jason Gunthorpe wrote:
> On Fri, Oct 09, 2020 at 06:02:09AM -0700, Raj, Ashok wrote:
> > On Fri, Oct 09, 2020 at 09:49:45AM -0300, Jason Gunthorpe wrote:
> > > On Fri, Oct 09, 2020 at 05:43:07AM -0700, Raj, Ashok wrote:
> > > > On Fri, Oct 09, 2020 at 08:57:37AM -0300, Jason Gunthorpe wrote:
> > > > > On Thu, Oct 08, 2020 at 06:22:31PM -0700, Raj, Ashok wrote:
> > > > >
> > > > > > Not randomly put there Jason :-).. There is a good reason for it.
> > > > >
> > > > > Sure the PASID value being associated with the IRQ make sense, but
> > > > > combining that register with the interrupt mask is just a compltely
> > > > > random thing to do.
> > > >
> > > > Hummm... Not sure what you are complaining.. but in any case giving
> > > > hardware a more efficient way to store interrupt entries breaking any
> > > > boundaries that maybe implied by the spec is why IMS was defined.
> > >
> > > I'm saying this PASID stuff is just some HW detail of IDXD and nothing
> > > that the core irqchip code should concern itself with
> >
> > Ok, so you are saying this is device specific why is generic framework
> > having to worry about the PASID stuff?
> >
> > I thought we are consolidating code that otherwise similar drivers would
> > require anyway. I thought that's what Thomas was accomplishing with the new
> > framework.
>
> My point is why would another driver combine PASID and the IRQ mask in
> one register? There is no spec saying to do this, no common design
IMS is a concept. How a device organizes its interrupt data is completely
hardware specific. Some vendor could keep them organized like how MSIx is
done today, and put PASID's in a separate offset. Or put all interrupt
related entries all in one place like how idxd handles it today.
Cheers,
Ashok