iommu/vt-d: Do not use flush-queue when caching-mode is on

When an Intel IOMMU is virtualized, and a physical device is
passed-through to the VM, changes of the virtual IOMMU need to be
propagated to the physical IOMMU. The hypervisor therefore needs to
monitor PTE mappings in the IOMMU page-tables. Intel specifications
provide "caching-mode" capability that a virtual IOMMU uses to report
that the IOMMU is virtualized and a TLB flush is needed after mapping to
allow the hypervisor to propagate virtual IOMMU mappings to the physical
IOMMU. To the best of my knowledge no real physical IOMMU reports
"caching-mode" as turned on.

Synchronizing the virtual and the physical IOMMU tables is expensive if
the hypervisor is unaware which PTEs have changed, as the hypervisor is
required to walk all the virtualized tables and look for changes.
Consequently, domain flushes are much more expensive than page-specific
flushes on virtualized IOMMUs with passthrough devices. The kernel
therefore exploited the "caching-mode" indication to avoid domain
flushing and use page-specific flushing in virtualized environments. See
commit 78d5f0f500e6 ("intel-iommu: Avoid global flushes with caching
mode.")

This behavior changed after commit 13cf01744608 ("iommu/vt-d: Make use
of iova deferred flushing"). Now, when batched TLB flushing is used (the
default), full TLB domain flushes are performed frequently, requiring
the hypervisor to perform expensive synchronization between the virtual
TLB and the physical one.

Getting batched TLB flushes to use page-specific invalidations again in
such circumstances is not easy, since the TLB invalidation scheme
assumes that "full" domain TLB flushes are performed for scalability.

Disable batched TLB flushes when caching-mode is on, as the performance
benefit from using batched TLB invalidations is likely to be much
smaller than the overhead of the virtual-to-physical IOMMU page-tables
synchronization.

Fixes: 13cf01744608 ("iommu/vt-d: Make use of iova deferred flushing")
Signed-off-by: Nadav Amit <namit@vmware.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Cc: Lu Baolu <baolu.lu@linux.intel.com>
Cc: Joerg Roedel <joro@8bytes.org>
Cc: Will Deacon <will@kernel.org>
Cc: stable@vger.kernel.org
Acked-by: Lu Baolu <baolu.lu@linux.intel.com>
Link: https://lore.kernel.org/r/20210127175317.1600473-1-namit@vmware.com
Signed-off-by: Joerg Roedel <jroedel@suse.de>

diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
index f665322a09919fbaea5dd3eae59c7ac87a7fe04d..06b00b5363d8630ebe7421fc78c127f517f2074d 100644 (file)
--- a/drivers/iommu/intel/iommu.c
+++ b/drivers/iommu/intel/iommu.c
@@ -5440,6 +5440,36 @@ intel_iommu_domain_set_attr(struct iommu_domain *domain,
        return ret;
 }
 
+static bool domain_use_flush_queue(void)
+{
+       struct dmar_drhd_unit *drhd;
+       struct intel_iommu *iommu;
+       bool r = true;
+
+       if (intel_iommu_strict)
+               return false;
+
+       /*
+        * The flush queue implementation does not perform page-selective
+        * invalidations that are required for efficient TLB flushes in virtual
+        * environments. The benefit of batching is likely to be much lower than
+        * the overhead of synchronizing the virtual and physical IOMMU
+        * page-tables.
+        */
+       rcu_read_lock();
+       for_each_active_iommu(iommu, drhd) {
+               if (!cap_caching_mode(iommu->cap))
+                       continue;
+
+               pr_warn_once("IOMMU batching is disabled due to virtualization");
+               r = false;
+               break;
+       }
+       rcu_read_unlock();
+
+       return r;
+}
+
 static int
 intel_iommu_domain_get_attr(struct iommu_domain *domain,
                            enum iommu_attr attr, void *data)
@@ -5450,7 +5480,7 @@ intel_iommu_domain_get_attr(struct iommu_domain *domain,
        case IOMMU_DOMAIN_DMA:
                switch (attr) {
                case DOMAIN_ATTR_DMA_USE_FLUSH_QUEUE:
-                       *(int *)data = !intel_iommu_strict;
+                       *(int *)data = domain_use_flush_queue();
                        return 0;
                default:
                        return -ENODEV;