iommu/vt-d: Global PASID name space

index 582fd01..2a81936 100644
@@ -518,6 +518,58 @@ static int iommu_identity_mapping;
 int intel_iommu_gfx_mapped;
 EXPORT_SYMBOL_GPL(intel_iommu_gfx_mapped);
 
+/*
+ * Intel IOMMU global PASID pool:
+ *
+ * Eventually I'm promised we will get a multi-level PASID table
+ * and it won't have to be physically contiguous. Until then,
+ * limit the size because 8MiB contiguous allocations can be hard
+ * to come by. The limit of 0x20000, which is 1MiB for each of
+ * the PASID and PASID-state tables, is somewhat arbitrary.
+ *
+ * PASID 0 is reserved in caching mode (virtualised IOMMU).
+ */
+#define PASID_MIN	0x1
+#define PASID_MAX	0x20000
+static DEFINE_SPINLOCK(pasid_lock);
+u32 intel_iommu_pasid_max = PASID_MAX;
+static DEFINE_IDR(pasid_idr);
+
+int intel_iommu_alloc_pasid(void *ptr, int start, int end, gfp_t gfp)
+{
+	int ret, min, max;
+
+	min = max_t(int, start, PASID_MIN);
+	max = min_t(int, end, intel_iommu_pasid_max);
+
+	WARN_ON(in_interrupt());
+	idr_preload(gfp);
+	spin_lock(&pasid_lock);
+	ret = idr_alloc(&pasid_idr, ptr, min, max, GFP_ATOMIC);
+	spin_unlock(&pasid_lock);
+	idr_preload_end();
+
+	return ret;
+}
+
+void intel_iommu_free_pasid(int pasid)
+{
+	spin_lock(&pasid_lock);
+	idr_remove(&pasid_idr, pasid);
+	spin_unlock(&pasid_lock);
+}
+
+void *intel_iommu_lookup_pasid(int pasid)
+{
+	void *p;
+
+	spin_lock(&pasid_lock);
+	p = idr_find(&pasid_idr, pasid);
+	spin_unlock(&pasid_lock);
+
+	return p;
+}
+
 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
 static DEFINE_SPINLOCK(device_domain_lock);
 static LIST_HEAD(device_domain_list);

@@ -3263,6 +3315,18 @@ static int __init init_dmars(void)
 	}
 
 	for_each_active_iommu(iommu, drhd) {
+		/*
+		 * Find the max pasid size of all IOMMU's in the system.
+		 * we need to ensure the system pasid table is no bigger
+		 * than the smallest supported.
+		 */
+		if (pasid_enabled(iommu)) {
+			u32 temp = 2 << ecap_pss(iommu->ecap);
+
+			intel_iommu_pasid_max = min_t(u32, temp,
+						      intel_iommu_pasid_max);
+		}
+
 		g_iommus[iommu->seq_id] = iommu;
 
 		intel_iommu_init_qi(iommu);

index 8dad3dd..bf83073 100644
@@ -494,4 +494,9 @@ extern struct intel_iommu *intel_svm_device_to_iommu(struct device *dev);
 
 extern const struct attribute_group *intel_iommu_groups[];
 
+extern u32 intel_iommu_pasid_max;
+int intel_iommu_alloc_pasid(void *ptr, int start, int end, gfp_t gfp);
+void intel_iommu_free_pasid(int pasid);
+void *intel_iommu_lookup_pasid(int pasid);
+
 #endif

index 99bc9bd..f3b7394 100644
@@ -298,6 +298,7 @@ static const struct mmu_notifier_ops intel_mmuops = {
 };
 
 static DEFINE_MUTEX(pasid_mutex);
+static LIST_HEAD(global_svm_list);
 
 int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_ops *ops)
 {

@@ -329,13 +330,13 @@ int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_
 
 	mutex_lock(&pasid_mutex);
 	if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
-		int i;
+		struct intel_svm *t;
 
-		idr_for_each_entry(&iommu->pasid_idr, svm, i) {
-			if (svm->mm != mm ||
-			    (svm->flags & SVM_FLAG_PRIVATE_PASID))
+		list_for_each_entry(t, &global_svm_list, list) {
+			if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
 				continue;
 
+			svm = t;
 			if (svm->pasid >= pasid_max) {
 				dev_warn(dev,
 					 "Limited PASID width. Cannot use existing PASID %d\n",

@@ -403,6 +404,7 @@ int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_
 		svm->mm = mm;
 		svm->flags = flags;
 		INIT_LIST_HEAD_RCU(&svm->devs);
+		INIT_LIST_HEAD(&svm->list);
 		ret = -ENOMEM;
 		if (mm) {
 			ret = mmu_notifier_register(&svm->notifier, mm);

@@ -433,6 +435,8 @@ int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_
 		 * anyway. Surely that can be left entirely to the guest? */
 		if (cap_caching_mode(iommu->cap))
 			intel_flush_pasid_dev(svm, sdev, 0);
+
+		list_add_tail(&svm->list, &global_svm_list);
 	}
 	list_add_rcu(&sdev->list, &svm->devs);
 

@@ -488,6 +492,8 @@ int intel_svm_unbind_mm(struct device *dev, int pasid)
 					if (svm->mm)
 						mmu_notifier_unregister(&svm->notifier, svm->mm);
 
+					list_del(&svm->list);
+
 					/* 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.

index bf83073..a2013dd 100644
@@ -486,6 +486,7 @@ struct intel_svm {
 	int flags;
 	int pasid;
 	struct list_head devs;
+	struct list_head list;
 };
 
 extern int intel_iommu_enable_pasid(struct intel_iommu *iommu, struct intel_svm_dev *sdev);

index f3b7394..1c45f75 100644
@@ -85,8 +85,6 @@ int intel_svm_alloc_pasid_tables(struct intel_iommu *iommu)
 				iommu->name);
 	}
 
-	idr_init(&iommu->pasid_idr);
-
 	return 0;
 }
 

@@ -102,7 +100,7 @@ int intel_svm_free_pasid_tables(struct intel_iommu *iommu)
 		free_pages((unsigned long)iommu->pasid_state_table, order);
 		iommu->pasid_state_table = NULL;
 	}
-	idr_destroy(&iommu->pasid_idr);
+
 	return 0;
 }
 

@@ -392,9 +390,9 @@ int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_
 			pasid_max = iommu->pasid_max;
 
 		/* Do not use PASID 0 in caching mode (virtualised IOMMU) */
-		ret = idr_alloc(&iommu->pasid_idr, svm,
-				!!cap_caching_mode(iommu->cap),
-				pasid_max - 1, GFP_KERNEL);
+		ret = intel_iommu_alloc_pasid(svm,
+					      !!cap_caching_mode(iommu->cap),
+					      pasid_max - 1, GFP_KERNEL);
 		if (ret < 0) {
 			kfree(svm);
 			goto out;

@@ -409,7 +407,7 @@ int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_
 		if (mm) {
 			ret = mmu_notifier_register(&svm->notifier, mm);
 			if (ret) {
-				idr_remove(&svm->iommu->pasid_idr, svm->pasid);
+				intel_iommu_free_pasid(svm->pasid);
 				kfree(svm);
 				kfree(sdev);
 				goto out;

@@ -463,7 +461,7 @@ int intel_svm_unbind_mm(struct device *dev, int pasid)
 	if (!iommu || !iommu->pasid_table)
 		goto out;
 
-	svm = idr_find(&iommu->pasid_idr, pasid);
+	svm = intel_iommu_lookup_pasid(pasid);
 	if (!svm)
 		goto out;
 

@@ -488,7 +486,7 @@ int intel_svm_unbind_mm(struct device *dev, int pasid)
 					svm->iommu->pasid_table[svm->pasid].val = 0;
 					wmb();
 
-					idr_remove(&svm->iommu->pasid_idr, svm->pasid);
+					intel_iommu_free_pasid(svm->pasid);
 					if (svm->mm)
 						mmu_notifier_unregister(&svm->notifier, svm->mm);
 

@@ -523,7 +521,7 @@ int intel_svm_is_pasid_valid(struct device *dev, int pasid)
 	if (!iommu || !iommu->pasid_table)
 		goto out;
 
-	svm = idr_find(&iommu->pasid_idr, pasid);
+	svm = intel_iommu_lookup_pasid(pasid);
 	if (!svm)
 		goto out;
 

@@ -621,7 +619,7 @@ static irqreturn_t prq_event_thread(int irq, void *d)
 
 		if (!svm || svm->pasid != req->pasid) {
 			rcu_read_lock();
-			svm = idr_find(&iommu->pasid_idr, req->pasid);
+			svm = intel_iommu_lookup_pasid(req->pasid);
 			/* It *can't* go away, because the driver is not permitted
 			 * to unbind the mm while any page faults are outstanding.
 			 * So we only need RCU to protect the internal idr code. */

index a2013dd..e50bef8 100644
@@ -418,7 +418,6 @@ struct intel_iommu {
 	struct pasid_state_entry *pasid_state_table;
 	struct page_req_dsc *prq;
 	unsigned char prq_name[16];    /* Name for PRQ interrupt */
-	struct idr pasid_idr;
 	u32 pasid_max;
 #endif
 	struct q_inval  *qi;            /* Queued invalidation info */