Re: [Qemu-devel] [PATCH v3] pci: Introduce helper to retrieve a PCI device's DMA address space

["Michael S. Tsirkin" <mst@redhat.com>]

On Sat, Aug 10, 2013 at 01:09:08AM +1000, Alexey Kardashevskiy wrote:
> A PCI device's DMA address space (possibly an IOMMU) is returned by a
> method on the PCIBus.  At the moment that only has one caller, so the
> method is simply open coded.  We'll need another caller for VFIO, so
> this patch introduces a helper/wrapper function.
> 
> If IOMMU is not set, the pci_device_iommu_address_space() function
> returns the parent's IOMMU skipping the "bus master" address space as
> otherwise proper emulation would require more effort for no benefit.
> 
> Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
> [aik: added inheritance from parent if iommu is not set for the current bus]
> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> 
> ---
> Changes:
> v3:
> * added comment about ignoring bus master address space
> 
> v2:
> * added inheritance, needed for a pci-bridge on spapr-ppc64
> * pci_iommu_as renamed to pci_device_iommu_address_space
> ---
>  hw/pci/pci.c         | 24 ++++++++++++++++++------
>  include/hw/pci/pci.h |  1 +
>  2 files changed, 19 insertions(+), 6 deletions(-)
> 
> diff --git a/hw/pci/pci.c b/hw/pci/pci.c
> index 4c004f5..dbfa395 100644
> --- a/hw/pci/pci.c
> +++ b/hw/pci/pci.c
> @@ -812,12 +812,7 @@ static PCIDevice *do_pci_register_device(PCIDevice *pci_dev, PCIBus *bus,
>      }
>  
>      pci_dev->bus = bus;
> -    if (bus->iommu_fn) {
> -        dma_as = bus->iommu_fn(bus, bus->iommu_opaque, devfn);
> -    } else {
> -        /* FIXME: inherit memory region from bus creator */
> -        dma_as = &address_space_memory;
> -    }
> +    dma_as = pci_device_iommu_address_space(pci_dev);
>  
>      memory_region_init_alias(&pci_dev->bus_master_enable_region,
>                               OBJECT(pci_dev), "bus master",
> @@ -2239,6 +2234,23 @@ static void pci_device_class_init(ObjectClass *klass, void *data)
>      k->props = pci_props;
>  }
>  
> +AddressSpace *pci_device_iommu_address_space(PCIDevice *dev)
> +{
> +    PCIBus *bus = PCI_BUS(dev->bus);
> +
> +    if (bus->iommu_fn) {
> +        return bus->iommu_fn(bus, bus->iommu_opaque, dev->devfn);
> +    }
> +
> +    if (bus->parent_dev) {
> +        /** We are ignoring the bus master DMA bit of the bridge
> +         *  as it would complicate things such as VFIO for no good reason */

        /*
         * Always
         * like
         * this
         */

        /** Never
         * like this */

The comment should be improved I think.
I would put it like this:
        /*
         * Note: this does not check bus master enable bit on device or
         * any of the pci to pci bridges above it, it's up to the caller to
         * check that before initiating the transaction.
         *
         * TODO: design a mechanism for callers to do this without
         * doing bus scans on data path.
         */

Would you like me to queue this on the pci tree? If yes I can
tweak the comment myself, no need to repost.

> +        return pci_device_iommu_address_space(bus->parent_dev);
> +    }
> +
> +    return &address_space_memory;
> +}
> +
>  void pci_setup_iommu(PCIBus *bus, PCIIOMMUFunc fn, void *opaque)
>  {
>      bus->iommu_fn = fn;
> diff --git a/include/hw/pci/pci.h b/include/hw/pci/pci.h
> index ccec2ba..2374aa9 100644
> --- a/include/hw/pci/pci.h
> +++ b/include/hw/pci/pci.h
> @@ -405,6 +405,7 @@ void pci_device_deassert_intx(PCIDevice *dev);
>  
>  typedef AddressSpace *(*PCIIOMMUFunc)(PCIBus *, void *, int);
>  
> +AddressSpace *pci_device_iommu_address_space(PCIDevice *dev);
>  void pci_setup_iommu(PCIBus *bus, PCIIOMMUFunc fn, void *opaque);
>  
>  static inline void
> -- 
> 1.8.3.2