[PATCH 2/5] iommu/arm-smmu: add support for PCI master devices

[Alex Williamson]

On Wed, 2014-07-09 at 14:26 +0100, Will Deacon wrote:
> [Adding Alex; question below]
> 
> On Thu, Jul 03, 2014 at 03:22:37PM +0100, Varun Sethi wrote:
> > > +static int __arm_smmu_get_pci_sid(struct pci_dev *pdev, u16 alias, void
> > > +*data) {
> > > +     *((u16 *)data) = alias;
> > > +     return 0; /* Continue walking */
> > > +}
> 
> [...]
> 
> > > @@ -1598,15 +1642,36 @@ static int arm_smmu_add_device(struct device
> > > *dev)
> > >               return PTR_ERR(group);
> > >       }
> > >
> > > +     if (dev_is_pci(dev)) {
> > > +             struct arm_smmu_master_cfg *cfg;
> > > +             struct pci_dev *pdev = to_pci_dev(dev);
> > > +
> > > +             cfg = kzalloc(sizeof(*cfg), GFP_KERNEL);
> > > +             if (!cfg) {
> > > +                     ret = -ENOMEM;
> > > +                     goto out_put_group;
> > > +             }
> > > +
> > > +             cfg->num_streamids = 1;
> > > +             pci_for_each_dma_alias(pdev, __arm_smmu_get_pci_sid,
> > > +                                    &cfg->streamids[0]);
> > [Sethi Varun-B16395] We need to look for upstream DMA device. We should be
> > using pci_find_dma_isolation_root here. Also, this would also imply that
> > there could be multiple devices sharing the same stream ID. So, we should
> > check if a particular stream ID value has already been configured in the
> > SMR registers.
> 
> pci_find_dma_isolation_root doesn't exist in any of the trees I have. Alex,
> is this queued anywhere and do I actually need it?

That function was in the v3 DMA alias series, in v4 it got replaced.
iommu_group_get_for_pci_dev() is the common function for finding or
creating a group for a device and pci_for_each_dma_alias() is the
interface to walk each alias in a PCI topology.  The pci_ interface is
in 3.16-rc and the iommu_ pieces are current in next via the iommu tree.

> The purpose of this code is to find the requester ID of a device as it
> appears at the host controller. At this point, we can map it (via firmware
> tables that are TBD) to a Stream ID for the SMMU. It looks to me like
> pci_for_each_dma_alias walks over non-transparent PCI bridges correctly, so
> the callback I provide just updates the alias until the walk has completed.

Yep, that's the intended usage.  There are cases in VT-d where it wants
to add context entries for every alias and cases elsewhere that we just
want the final alias.  pci_for_each_dma_alias() is meant to fit both use
cases.  Thanks,

Alex