Hi Jörg,
On Wed, May 13, 2020 at 11:53:04AM +0200, Joerg Roedel wrote:
> On Tue, May 05, 2020 at 12:09:32PM +0200, Maxime Ripard wrote:
> > +static u32 *sun50i_dte_get_page_table(struct sun50i_iommu_domain *sun50i_domain,
> > + dma_addr_t iova, gfp_t gfp)
> > +{
> > + struct sun50i_iommu *iommu = sun50i_domain->iommu;
> > + unsigned long flags;
> > + u32 *page_table;
> > + u32 *dte_addr;
> > + u32 old_dte;
> > + u32 dte;
> > +
> > + dte_addr = &sun50i_domain->dt[sun50i_iova_get_dte_index(iova)];
> > + dte = *dte_addr;
> > + if (sun50i_dte_is_pt_valid(dte)) {
> > + phys_addr_t pt_phys = sun50i_dte_get_pt_address(dte);
> > + return (u32 *)phys_to_virt(pt_phys);
> > + }
> > +
> > + page_table = sun50i_iommu_alloc_page_table(iommu, gfp);
> > + if (IS_ERR(page_table))
> > + return page_table;
> > +
> > + dte = sun50i_mk_dte(virt_to_phys(page_table));
> > + old_dte = cmpxchg(dte_addr, 0, dte);
> > + if (old_dte) {
> > + phys_addr_t installed_pt_phys =
> > + sun50i_dte_get_pt_address(old_dte);
> > + u32 *installed_pt = phys_to_virt(installed_pt_phys);
> > + u32 *drop_pt = page_table;
> > +
> > + page_table = installed_pt;
> > + dte = old_dte;
> > + sun50i_iommu_free_page_table(iommu, drop_pt);
> > + }
> > +
> > + sun50i_table_flush(sun50i_domain, page_table, PT_SIZE);
> > + sun50i_table_flush(sun50i_domain, dte_addr, 1);
> > +
> > + spin_lock_irqsave(&iommu->iommu_lock, flags);
> > + sun50i_iommu_ptw_invalidate(iommu, iova);
> > + spin_unlock_irqrestore(&iommu->iommu_lock, flags);
>
> Why is that needed, does the PTW also cache non-present entries?
The documentation is pretty sparse, so I was conservative again. I've removed it
and it seems to be working fine in all the cases I could test, so I guess we're
fine without it :)
> > +static int sun50i_iommu_add_device(struct device *dev)
> > +{
> > + struct sun50i_iommu *iommu;
> > + struct iommu_group *group;
> > +
> > + iommu = sun50i_iommu_from_dev(dev);
> > + if (!iommu)
> > + return -ENODEV;
> > +
> > + group = iommu_group_get_for_dev(dev);
> > + if (IS_ERR(group))
> > + return PTR_ERR(group);
> > +
> > + iommu_group_put(group);
> > +
> > + return 0;
> > +}
> > +
> > +static void sun50i_iommu_remove_device(struct device *dev)
> > +{
> > + iommu_group_remove_device(dev);
> > +}
>
>
> These two call-backs have been renamed in the iommu-tree to
> probe_device() and release_device() with slightly different semantics
> and function signatures. I think for this driver they should look like
> this:
>
> static struct iommu_device *sun50i_iommu_probe_device(struct device *dev)
> {
> struct sun50i_iommu *iommu;
>
> iommu = sun50i_iommu_from_dev(dev);
> if (!iommu)
> return ERR_PTR(-ENODEV);
>
> return &iommu->iommu;
> }
>
> static void sun50i_iommu_release_device(struct device *dev)
> {
> }
>
> Can you pleas rebase these patches to the 'core' branch of the
> iommu-tree and use these new call-backs?
>
> The rest of your driver looks good to me. Good work!
I've rebased it on next, and it indeed works with your suggestion. I'll resend a
new version in a short while.
Thanks!
Maxime
Attachment:
signature.asc
Description: PGP signature
