On Wed, Jun 18, 2025 at 04:09:12PM +0200, Benjamin Gaignard wrote:

+config VSI_IOMMU
+	bool "Verisilicon IOMMU Support"
+	depends on ARM64
+	select IOMMU_API
+	select ARM_DMA_USE_IOMMU

ARM_DMA_USE_IOMMU is only used by ARM32, you don't need it if you
depends on ARM64

+static void vsi_iommu_release_device(struct device *dev)
+{
+	struct vsi_iommu *iommu = dev_iommu_priv_get(dev);
+
+	device_link_remove(dev, iommu->dev);
+}

This does not seem right, release is supposed to reprogram the HW to
stop walking any page table.

You should implement a static blocked (or identity?) domain that idles
the hardware and use that as the blocked and release_domain in the
ops.

The logic around vsi_iommu_detach_device() and
vsi_iommu_attach_device() is also not quite right. The attach can
happen while iommu->domain is already set and doesn't deal with
removing the iommu from the old domain's list.

I would probably change vsi_iommu_enable() into vsi_iommu_set_paging()
and then presumably vsi_iommu_disable() is vsi_iommu_set_blocking() ?

vsi_iommu_detach_device() should be deleted and integrated into the
blocked domain and attach error unwind.

+static int vsi_iommu_of_xlate(struct device *dev,
+			      const struct of_phandle_args *args)
+{
+	struct platform_device *iommu_dev;
+
+	if (!dev_iommu_priv_get(dev)) {
+		iommu_dev = of_find_device_by_node(args->np);
+		if (WARN_ON(!iommu_dev))
+			return -EINVAL;
+
+		dev_iommu_priv_set(dev, platform_get_drvdata(iommu_dev));
+	}

The driver should ideally not be calling dev_iommu_priv_set/get here,
and this leads the reference doesn't it? Do what ARM did to locate
the iommu_dev.

I would also add a comment here:

+static int vsi_iommu_map(struct iommu_domain *domain, unsigned long _iova,
+                        phys_addr_t paddr, size_t size, size_t count,
+                        int prot, gfp_t gfp, size_t *mapped)
+{
+       struct vsi_iommu_domain *vsi_domain = to_vsi_domain(domain);
+       unsigned long flags;
+       dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
+       u32 *page_table, *pte_addr;
+       u32 dte, pte_index;
+       int ret;

/*
 * IOMMU drivers are not supposed to lock the page table, however this
 * driver does not safely handle the cache flushing or table
 * installation across concurrent threads so locking is used as a simple
 * solution.
 */

+       spin_lock_irqsave(&vsi_domain->dt_lock, flags);

Jason