On 2025/9/9 21:06:33, Will Deacon [off-list ref] wrote:

On Thu, Aug 14, 2025 at 05:30:04PM +0800, Qinxin Xia wrote:

quoted

+/**
+ * iova_info_dump - dump iova alloced
+ * @s - file structure used to generate serialized output
+ * @iovad: - iova domain in question.
+ */
+static int iommu_iova_info_dump(struct seq_file *s, struct iommu_domain *domain)
+{
+	struct iova_domain *iovad;
+	unsigned long long pfn;
+	unsigned long i_shift;
+	struct rb_node *node;
+	unsigned long flags;
+	size_t prot_size;
+
+	iovad = iommu_domain_to_iovad(domain);
+	if (!iovad)
+		return -ENOMEM;
+
+	i_shift = iova_shift(iovad);
+
+	/* Take the lock so that no other thread is manipulating the rbtree */
+	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
+	assert_spin_locked(&iovad->iova_rbtree_lock);
+
+	for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
+		struct iova *iova = rb_entry(node, struct iova, node);
+
+		if (iova->pfn_hi <= iova->pfn_lo)
+			continue;
+
+		for (pfn = iova->pfn_lo; pfn <= iova->pfn_hi; ) {
+			prot_size = domain->ops->dump_iova_prot(s, domain, pfn << i_shift);
+			pfn = ((pfn << i_shift) + prot_size) >> i_shift;
+		}
+	}
+
+	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);

Why is the IOVA rbtree lock sufficient for serialising the page-table
accesses made by ->dump_iova_prot()? I don't see anything here that
prevents the walker walking into page-table pages that are e.g. being
freed or manipulated concurrently.

Will

  Thank you for catching this critical race condition.I will fix this in
  next version. And,Jason suggests putting io_ptdump on top of iommu pt.
  What do you think?