Re: [PATCH v2 2/4] iommu: Add Allwinner H6 IOMMU driver

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

Hi Maxime,

On Thu, Feb 20, 2020 at 07:15:14PM +0100, Maxime Ripard wrote:
> +struct sun50i_iommu_domain {
> +	struct iommu_domain domain;
> +
> +	/* Number of devices attached to the domain */
> +	refcount_t refcnt;
> +
> +	/* Lock to modify the Directory Table */
> +	spinlock_t dt_lock;

I suggest you make page-table updates lock-less. Otherwise this lock
will become a bottle-neck when using the IOMMU through DMA-API.

> +
> +static int sun50i_iommu_map(struct iommu_domain *domain, unsigned long iova,
> +			    phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
> +{
> +	struct sun50i_iommu_domain *sun50i_domain = to_sun50i_domain(domain);
> +	struct sun50i_iommu *iommu = sun50i_domain->iommu;
> +	u32 pte_index;
> +	u32 *page_table, *pte_addr;
> +	unsigned long flags;
> +	int ret = 0;
> +
> +	spin_lock_irqsave(&sun50i_domain->dt_lock, flags);
> +	page_table = sun50i_dte_get_page_table(sun50i_domain, iova, gfp);
> +	if (IS_ERR(page_table)) {
> +		ret = PTR_ERR(page_table);
> +		goto out;
> +	}
> +
> +	pte_index = sun50i_iova_get_pte_index(iova);
> +	pte_addr = &page_table[pte_index];
> +	if (sun50i_pte_is_page_valid(*pte_addr)) {

You can use unlikely() here.

> +		phys_addr_t page_phys = sun50i_pte_get_page_address(*pte_addr);
> +		dev_err(iommu->dev,
> +			"iova %pad already mapped to %pa cannot remap to %pa prot: %#x\n",
> +			&iova, &page_phys, &paddr, prot);
> +		ret = -EBUSY;
> +		goto out;
> +	}
> +
> +	*pte_addr = sun50i_mk_pte(paddr, prot);
> +	sun50i_table_flush(sun50i_domain, pte_addr, 1);

This maps only one page, right? But the function needs to map up to
'size' as given in the parameter list.

> +
> +	spin_lock_irqsave(&iommu->iommu_lock, flags);
> +	sun50i_iommu_tlb_invalidate(iommu, iova);
> +	spin_unlock_irqrestore(&iommu->iommu_lock, flags);

Why is there a need to flush the TLB here? The IOMMU-API provides
call-backs so that the user of the API can decide when it wants
to flush the IO/TLB. Such flushes are usually expensive and doing them
on every map and unmap will cost significant performance.

> +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;
> +	unsigned long flags;
> +	phys_addr_t pt_phys;
> +	dma_addr_t pte_dma;
> +	u32 *pte_addr;
> +	u32 dte;
> +
> +	spin_lock_irqsave(&sun50i_domain->dt_lock, flags);
> +
> +	dte = sun50i_domain->dt[sun50i_iova_get_dte_index(iova)];
> +	if (!sun50i_dte_is_pt_valid(dte)) {
> +		spin_unlock_irqrestore(&sun50i_domain->dt_lock, flags);
> +		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)) {
> +		spin_unlock_irqrestore(&sun50i_domain->dt_lock, flags);
> +		return 0;
> +	}
> +
> +	memset(pte_addr, 0, sizeof(*pte_addr));
> +	sun50i_table_flush(sun50i_domain, pte_addr, 1);
> +
> +	spin_lock(&iommu->iommu_lock);
> +	sun50i_iommu_tlb_invalidate(iommu, iova);
> +	sun50i_iommu_ptw_invalidate(iommu, iova);
> +	spin_unlock(&iommu->iommu_lock);

Same objections as in the map function. This only unmaps one page, and
is the IO/TLB flush really needed here?

> +static struct iommu_domain *sun50i_iommu_domain_alloc(unsigned type)
> +{
> +	struct sun50i_iommu_domain *sun50i_domain;
> +
> +	if (type != IOMMU_DOMAIN_DMA && type != IOMMU_DOMAIN_UNMANAGED)
> +		return NULL;

I think you should at least also support identity domains here. The
iommu-core code might allocate those for default domains.


Regards,

	Joerg