Re: [Qemu-devel] [0/10] Preliminary work for IOMMU emulation support; the easy bits

[David Gibson]

On Thu, Sep 01, 2011 at 02:33:41PM +0300, Eduard - Gabriel Munteanu wrote:
> On Thu, Sep 01, 2011 at 03:00:53PM +1000, David Gibson wrote:
> > A while back, Eduard - Gabriel Munteanu send a series of patches
> > implementing support for emulating the AMD IOMMU in conjunction with
> > qemu emulated PCI devices.  A revised patch series added support for
> > the Intel IOMMU, and I also send a revised version of this series
> > which added support for the hypervisor mediated IOMMU on the pseries
> > machine.
> > 
> > Richard Henderson also weighed in on the discussion, and there's still
> > a fair bit to be thrashed out in terms of exactly how to set up an
> > IOMMU / DMA translation subsystem.
> > 
> > However, really only 2 or 3 patches in any of these series have
> > contained anything interesting.  The rest of the series has been
> > converting existing PCI emulated devices to use the new DMA interface
> > which worked through the IOMMU translation, whatever it was.  While we
> > keep working out what we want for the guts of the IOMMU support, these
> > device conversion patches keep bitrotting against updates to the
> > various device implementations themselves.
> 
> Hi,
> 
> This sounds like a good idea. We should be able to agree on these bits,
> at least, and get them merged.
> 
> > Really, regardless of whether we're actually implementing IOMMU
> > translation, it makes sense that qemu code should distinguish between
> > when it is really operating in CPU physical addresses and when it is
> > operating in bus or DMA addresses which might have some kind of
> > translation into physical addresses.
> > 
> > This series, therefore, begins the conversion of existing PCI device
> > emulation code to use new (stub) pci dma access functions.  These are
> > for now, just defined to be untranslated cpu physical memory accesses,
> > as before, but has two advantages:
> > 
> >    * It becomes obvious where the code is working with dma addresses,
> >      so it's easier to grep for what might be affected by an IOMMU or
> >      other bus address translation.
> > 
> >    * The new stubs take the PCIDevice *, from which any of the various
> >      suggested IOMMU interfaces should be able to locate the correct
> >      IOMMU translation context.
> > 
> > This series only converts the easy cases so far.  That is simple
> > direct DMA access from device code:
> > cpu_physical_memory_{read,write}(), ld*_phys() and st*_phys().  It
> > doesn't handle devices which use the scatter/gather code (just ide and
> > UHCI, so far).  I plan to address that later, but I have some details
> > still to work out.
> 
> I think AHCI belongs on that list too, my patches didn't handle it
> either.

Ah, yes.  I was lumping that in with IDE.

> Somewhere down the road we could try to poison the old cpu_*
> interfaces, as suggested by others before.
> 
> > Anthony, please apply.
> 
> It's nice to see this topic got revived. I'll try to post my latest
> SeaBIOS patches these days. The reason I didn't last time was I didn't
> get much review/acks/naks on the IOMMU patches from core devs. Hopefully
> things will go smoother this time.
> 
> 
>       Cheers,
>       Eduard
> 

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson