Re: [PATCH 01/12] iommu/vt-d: Use iommu_get_domain_for_dev() in debugfs

From: Baolu Lu
Date: Wed Jun 01 2022 - 01:33:46 EST


Hi Robin,

Thank you for the comments.

On 2022/5/31 21:52, Robin Murphy wrote:
On 2022-05-31 04:02, Baolu Lu wrote:
On 2022/5/30 20:14, Jason Gunthorpe wrote:
On Sun, May 29, 2022 at 01:14:46PM +0800, Baolu Lu wrote:

[--snip--]

diff --git a/drivers/iommu/intel/debugfs.c b/drivers/iommu/intel/debugfs.c
index d927ef10641b..e6f4835b8d9f 100644
--- a/drivers/iommu/intel/debugfs.c
+++ b/drivers/iommu/intel/debugfs.c
@@ -333,25 +333,28 @@ static void pgtable_walk_level(struct seq_file *m, struct dma_pte *pde,
              continue;

          path[level] = pde->val;
-        if (dma_pte_superpage(pde) || level == 1)
+        if (dma_pte_superpage(pde) || level == 1) {
              dump_page_info(m, start, path);
-        else
-            pgtable_walk_level(m, phys_to_virt(dma_pte_addr(pde)),
+        } else {
+            unsigned long phys_addr;
+
+            phys_addr = (unsigned long)dma_pte_addr(pde);
+            if (!pfn_valid(__phys_to_pfn(phys_addr)))

Given that pte_present(pde) passed just above, it was almost certainly a valid entry, so it seems unlikely that the physical address it pointed to could have disappeared in the meantime. If you're worried about the potential case where we've been preempted during this walk for long enough that the page has already been freed by an unmap, reallocated,

Yes. This is exactly what I am worried about and what this patch wants
to solve.

and filled with someone else's data that happens to look like valid PTEs, this still isn't enough, since that data could just as well happen to look like valid physical addresses too.
I imagine that if you want to safely walk pagetables concurrently with them potentially being freed, you'd probably need to get RCU involved.

I don't want to make the map/unmap interface more complex or inefficient
because of a debugfs feature. I hope that the debugfs and map/unmap
interfaces are orthogonal, just like the IOMMU hardware traversing the
page tables, as long as the accessed physical address is valid and
accessible. Otherwise, stop the traversal immediately. If we can't
achieve this, I'd rather stop supporting this debugfs node.


+                break;
+            pgtable_walk_level(m, phys_to_virt(phys_addr),

Also, obligatory reminder that pfn_valid() only means that pfn_to_page() gets you a valid struct page. Whether that page is direct-mapped kernel memory or not is a different matter.

Perhaps I can check this from the page flags?


                         level - 1, start, path);
+        }
          path[level] = 0;
      }
  }

-static int show_device_domain_translation(struct device *dev, void *data)
+static int __show_device_domain_translation(struct device *dev, void *data)
  {
      struct device_domain_info *info = dev_iommu_priv_get(dev);
      struct dmar_domain *domain = info->domain;
      struct seq_file *m = data;
      u64 path[6] = { 0 };

-    if (!domain)
-        return 0;
-
      seq_printf(m, "Device %s @0x%llx\n", dev_name(dev),
             (u64)virt_to_phys(domain->pgd));
      seq_puts(m, "IOVA_PFN\t\tPML5E\t\t\tPML4E\t\t\tPDPE\t\t\tPDE\t\t\tPTE\n");
@@ -359,20 +362,27 @@ static int show_device_domain_translation(struct device *dev, void *data)
      pgtable_walk_level(m, domain->pgd, domain->agaw + 2, 0, path);
      seq_putc(m, '\n');

-    return 0;
+    return 1;
  }

-static int domain_translation_struct_show(struct seq_file *m, void *unused)
+static int show_device_domain_translation(struct device *dev, void *data)
  {
-    unsigned long flags;
-    int ret;
+    struct iommu_group *group;

-    spin_lock_irqsave(&device_domain_lock, flags);
-    ret = bus_for_each_dev(&pci_bus_type, NULL, m,
-                   show_device_domain_translation);
-    spin_unlock_irqrestore(&device_domain_lock, flags);
+    group = iommu_group_get(dev);
+    if (group) {
+        iommu_group_for_each_dev(group, data,
+                     __show_device_domain_translation);

Why group_for_each_dev?

This will hold the group mutex when the callback is invoked. With the
group mutex hold, the domain could not get changed.

If there *are* multiple devices in the group then by definition they should be attached to the same domain, so dumping that domain's mappings more than once seems pointless. Especially given that the outer bus_for_each_dev iteration will already visit each individual device anyway, so this would only make the redundancy even worse than it already is.

__show_device_domain_translation() only dumps mappings once as it always
returns 1.

Best regards,
baolu