iommu/vt-d: Implement SVM_FLAG_SUPERVISOR_MODE for kernel access

This is only usable for the static 1:1 mapping of physical memory.

Any access to vmalloc or module regions will require some way of doing
an IOTLB flush. It's theoretically possible to hook into the
tlb_flush_kernel_range() function, but that seems like overkill — most
of the addresses accessed through a kernel PASID *will* be in the 1:1
mapping.

If we really need to allow access to more interesting kernel regions,
then the answer will probably be an explicit IOTLB flush call after use,
akin to the DMA API's unmap function.

In fact, it might be worth introducing that sooner rather than later, and
making it just BUG() if the address isn't in the static 1:1 mapping.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>

drivers/iommu/intel-svm.c

@@ -269,11 +269,10 @@ int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_
 	struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
 	struct intel_svm_dev *sdev;
 	struct intel_svm *svm = NULL;
+	struct mm_struct *mm = NULL;
 	int pasid_max;
 	int ret;
 
-	BUG_ON(pasid && !current->mm);
-
 	if (WARN_ON(!iommu))
 		return -EINVAL;
 
@@ -284,12 +283,20 @@ int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_
 	} else
 		pasid_max = 1 << 20;
 
+	if ((flags & SVM_FLAG_SUPERVISOR_MODE)) {
+		if (!ecap_srs(iommu->ecap))
+			return -EINVAL;
+	} else if (pasid) {
+		mm = get_task_mm(current);
+		BUG_ON(!mm);
+	}
+
 	mutex_lock(&pasid_mutex);
 	if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
 		int i;
 
 		idr_for_each_entry(&iommu->pasid_idr, svm, i) {
-			if (svm->mm != current->mm ||
+			if (svm->mm != mm ||
 			    (svm->flags & SVM_FLAG_PRIVATE_PASID))
 				continue;
 
@@ -355,17 +362,22 @@ int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_
 		}
 		svm->pasid = ret;
 		svm->notifier.ops = &intel_mmuops;
-		svm->mm = get_task_mm(current);
+		svm->mm = mm;
 		svm->flags = flags;
 		INIT_LIST_HEAD_RCU(&svm->devs);
 		ret = -ENOMEM;
-		if (!svm->mm || (ret = mmu_notifier_register(&svm->notifier, svm->mm))) {
-			idr_remove(&svm->iommu->pasid_idr, svm->pasid);
-			kfree(svm);
-			kfree(sdev);
-			goto out;
-		}
-		iommu->pasid_table[svm->pasid].val = (u64)__pa(svm->mm->pgd) | 1;
+		if (mm) {
+			ret = mmu_notifier_register(&svm->notifier, mm);
+			if (ret) {
+				idr_remove(&svm->iommu->pasid_idr, svm->pasid);
+				kfree(svm);
+				kfree(sdev);
+				goto out;
+			}
+			iommu->pasid_table[svm->pasid].val = (u64)__pa(mm->pgd) | 1;
+			mm = NULL;
+		} else
+			iommu->pasid_table[svm->pasid].val = (u64)__pa(init_mm.pgd) | 1 | (1ULL << 11);
 		wmb();
 	}
 	list_add_rcu(&sdev->list, &svm->devs);
@@ -375,6 +387,8 @@ int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_
 	ret = 0;
  out:
 	mutex_unlock(&pasid_mutex);
+	if (mm)
+		mmput(mm);
 	return ret;
 }
 EXPORT_SYMBOL_GPL(intel_svm_bind_mm);
@@ -416,7 +430,8 @@ int intel_svm_unbind_mm(struct device *dev, int pasid)
 					mmu_notifier_unregister(&svm->notifier, svm->mm);
 
 					idr_remove(&svm->iommu->pasid_idr, svm->pasid);
-					mmput(svm->mm);
+					if (svm->mm)
+						mmput(svm->mm);
 					/* We mandate that no page faults may be outstanding
 					 * for the PASID when intel_svm_unbind_mm() is called.
 					 * If that is not obeyed, subtle errors will happen.
@@ -500,6 +515,10 @@ static irqreturn_t prq_event_thread(int irq, void *d)
 		}
 
 		result = QI_RESP_INVALID;
+		/* Since we're using init_mm.pgd directly, we should never take
+		 * any faults on kernel addresses. */
+		if (!svm->mm)
+			goto bad_req;
 		down_read(&svm->mm->mmap_sem);
 		vma = find_extend_vma(svm->mm, address);
 		if (!vma || address < vma->vm_start)

include/linux/intel-svm.h

@@ -40,7 +40,20 @@ struct svm_dev_ops {
  * disambiguate between multiple device contexts which access the same MM,
  * if there is no other way to do so. It should be used sparingly, if at all.
  */
-#define SVM_FLAG_PRIVATE_PASID	(1<<0)
+#define SVM_FLAG_PRIVATE_PASID		(1<<0)
+
+/*
+ * The SVM_FLAG_SUPERVISOR_MODE flag requests a PASID which can be used only
+ * for access to kernel addresses. No IOTLB flushes are automatically done
+ * for kernel mappings; it is valid only for access to the kernel's static
+ * 1:1 mapping of physical memory — not to vmalloc or even module mappings.
+ * A future API addition may permit the use of such ranges, by means of an
+ * explicit IOTLB flush call (akin to the DMA API's unmap method).
+ *
+ * It is unlikely that we will ever hook into flush_tlb_kernel_range() to
+ * do such IOTLB flushes automatically.
+ */
+#define SVM_FLAG_SUPERVISOR_MODE	(1<<1)
 
 /**
  * intel_svm_bind_mm() - Bind the current process to a PASID