Re: [PATCH v3 3/5] iommu: Add Allwinner H6 IOMMU driver

[Joerg Roedel <joro@8bytes.org>]

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?

> +static size_t sun50i_iommu_unmap(struct iommu_domain *domain, unsigned long iova,
> +				 size_t size, struct iommu_iotlb_gather *gather)
> +{
> +	struct sun50i_iommu_domain *sun50i_domain = to_sun50i_domain(domain);
> +	struct sun50i_iommu *iommu = sun50i_domain->iommu;
> +	phys_addr_t pt_phys;
> +	dma_addr_t pte_dma;
> +	u32 *pte_addr;
> +	u32 dte;
> +
> +	dte = sun50i_domain->dt[sun50i_iova_get_dte_index(iova)];
> +	if (!sun50i_dte_is_pt_valid(dte))
> +		return 0;
> +
> +	pt_phys = sun50i_dte_get_pt_address(dte);
> +	pte_addr = (u32 *)phys_to_virt(pt_phys) + sun50i_iova_get_pte_index(iova);
> +	pte_dma = pt_phys + sun50i_iova_get_pte_index(iova) * PT_ENTRY_SIZE;
> +
> +	if (!sun50i_pte_is_page_valid(*pte_addr))
> +		return 0;
> +
> +	memset(pte_addr, 0, sizeof(*pte_addr));
> +	sun50i_table_flush(sun50i_domain, pte_addr, 1);
> +
> +	spin_lock(&iommu->iommu_lock);
> +	sun50i_iommu_ptw_invalidate(iommu, iova);
> +	spin_unlock(&iommu->iommu_lock);

And is that also needed? You clear a PTE here and not a top-level DT
entry. All these spinlocks in the map/unmap paths will truly hurt
performance.

And if it is really needed you can defer it into the iotlb_sync()
call-back.

> +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!

Thanks,

	Joerg