[RFC PATCH 09/14] iommu: lazy-populate iommu_group reserved_regions/type attrs

[Pavol Sakac]

iommu_group_alloc() eagerly creates two attribute files -
reserved_regions and type - that are read only on the VFIO /
iommufd ioctl path (VFIO_IOMMU_GET_INFO, iommufd type query).
On systems with thousands of SR-IOV VFs each VF typically gets
its own iommu_group, so deferring these two attrs eliminates two
kernfs nodes per group at boot.

iommu_group_ktype is the first non-device_ktype consumer of the
populate mechanism: wire iommu_group_populate_one /
iommu_group_populate_all and call kernfs_set_lazy() on the
group's kernfs node immediately after kobject_init_and_add().
See the in-diff comment at the kernfs_set_lazy() site for the
ordering rationale (must precede kobject_create_and_add("devices")
to interact correctly with kernfs_inc_rev()).

Both callbacks fast-path on iommu_group_sysfs_populated() so
re-entry after the directory is fully populated is cheap and
cannot fire sysfs_warn_dup().

The name attribute is created on demand by iommu_group_set_name()
on a separate API path and is not handled by these callbacks;
groups without a set name return -ENOENT on lookup of name.

No userspace ABI change.  A selftest for this deferral is added
later in the series under tools/testing/selftests/sysfs-lazy/.

Cc: Joerg Roedel <joro@xxxxxxxxxx>
Cc: Will Deacon <will@xxxxxxxxxx>
Cc: Robin Murphy <robin.murphy@xxxxxxx>
Cc: Jason Gunthorpe <jgg@xxxxxxxx>
Cc: Kevin Tian <kevin.tian@xxxxxxxxx>
Cc: Shuah Khan <shuah@xxxxxxxxxx>
Cc: iommu@xxxxxxxxxxxxxxx
Cc: linux-kselftest@xxxxxxxxxxxxxxx
Cc: linux-kernel@xxxxxxxxxxxxxxx
Cc: linux-api@xxxxxxxxxxxxxxx
Cc: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
Cc: Danilo Krummrich <dakr@xxxxxxxxxx>
Cc: driver-core@xxxxxxxxxxxxxxx
Assisted-by: Claude:claude-opus-4.7
Signed-off-by: Pavol Sakac <sakacpav@xxxxxxxxx>
---
 drivers/iommu/iommu.c | 132 +++++++++++++++++++++++++++++++++++++-----
 1 file changed, 119 insertions(+), 13 deletions(-)

diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
index 61c12ba782066..7a8dd43f0a09a 100644
--- a/drivers/iommu/iommu.c
+++ b/drivers/iommu/iommu.c
@@ -21,6 +21,7 @@
 #include <linux/iommufd.h>
 #include <linux/idr.h>
 #include <linux/err.h>
+#include <linux/kernfs.h>
 #include <linux/pci.h>
 #include <linux/pci-ats.h>
 #include <linux/bitops.h>
@@ -71,8 +72,22 @@ struct iommu_group {
 	struct list_head entry;
 	unsigned int owner_cnt;
 	void *owner;
+	/* Embedded lazy-sysfs state (every iommu_group is lazy). */
+	struct sysfs_lazy_state sysfs_lazy;
 };
 
+static inline bool iommu_group_sysfs_populated(const struct iommu_group *group)
+{
+	return READ_ONCE(group->sysfs_lazy.populated);
+}
+
+static inline void
+iommu_group_sysfs_set_populated(struct iommu_group *group)
+{
+	lockdep_assert_held(&group->sysfs_lazy.lock);
+	WRITE_ONCE(group->sysfs_lazy.populated, true);
+}
+
 struct group_device {
 	struct list_head list;
 	struct device *dev;
@@ -1033,9 +1048,99 @@ static void iommu_group_release(struct kobject *kobj)
 	kfree(group);
 }
 
+/* Lazy iommu_group attrs (excludes "name", managed separately). */
+static const struct {
+	const char			*name;
+	struct iommu_group_attribute	*attr;
+} iommu_group_lazy_attrs[] = {
+	{ "reserved_regions", &iommu_group_attr_reserved_regions },
+	{ "type",             &iommu_group_attr_type },
+};
+
+static int iommu_group_populate_one(struct kobject *kobj, const char *name)
+{
+	struct iommu_group *group = to_iommu_group(kobj);
+	size_t i;
+	int ret = -ENOENT;
+	bool name_present = false;
+
+	/* Fast path: directory fully populated; kernfs has authoritative state. */
+	if (iommu_group_sysfs_populated(group))
+		return -ENOENT;
+
+	mutex_lock(&group->sysfs_lazy.lock);
+	/* Re-check under the lock against a concurrent populate_all. */
+	if (iommu_group_sysfs_populated(group)) {
+		ret = -ENOENT;
+		goto out;
+	}
+
+	/* Materialise the whole table (2 entries); cheaper than per-name dispatch. */
+	for (i = 0; i < ARRAY_SIZE(iommu_group_lazy_attrs); i++) {
+		struct iommu_group_attribute *attr =
+			iommu_group_lazy_attrs[i].attr;
+		int rc;
+
+		if (sysfs_kn_exists(&group->kobj, attr->attr.name))
+			rc = 0;
+		else
+			rc = iommu_group_create_file(group, attr);
+
+		if (!strcmp(name, iommu_group_lazy_attrs[i].name)) {
+			name_present = true;
+			ret = rc;
+		} else if (rc) {
+			pr_warn("group %d: lazy-create %s failed: %d\n",
+				group->id,
+				iommu_group_lazy_attrs[i].name, rc);
+		}
+	}
+	iommu_group_sysfs_set_populated(group);
+
+out:
+	mutex_unlock(&group->sysfs_lazy.lock);
+	return name_present ? ret : -ENOENT;
+}
+
+static void iommu_group_populate_all(struct kobject *kobj)
+{
+	struct iommu_group *group = to_iommu_group(kobj);
+	size_t i;
+	int ret;
+
+	/* Fast path: directory already fully populated. */
+	if (iommu_group_sysfs_populated(group))
+		return;
+
+	mutex_lock(&group->sysfs_lazy.lock);
+	if (iommu_group_sysfs_populated(group))
+		goto out;
+
+	for (i = 0; i < ARRAY_SIZE(iommu_group_lazy_attrs); i++) {
+		struct iommu_group_attribute *attr =
+			iommu_group_lazy_attrs[i].attr;
+
+		/* Existence check absorbs the populate_one race. */
+		if (sysfs_kn_exists(&group->kobj, attr->attr.name))
+			continue;
+
+		ret = iommu_group_create_file(group, attr);
+		if (ret)
+			pr_warn("group %d: lazy-create %s failed: %d\n",
+				group->id,
+				iommu_group_lazy_attrs[i].name, ret);
+	}
+
+	iommu_group_sysfs_set_populated(group);
+out:
+	mutex_unlock(&group->sysfs_lazy.lock);
+}
+
 static const struct kobj_type iommu_group_ktype = {
 	.sysfs_ops = &iommu_group_sysfs_ops,
 	.release = iommu_group_release,
+	.populate     = iommu_group_populate_one,
+	.populate_all = iommu_group_populate_all,
 };
 
 /**
@@ -1060,6 +1165,7 @@ struct iommu_group *iommu_group_alloc(void)
 
 	group->kobj.kset = iommu_group_kset;
 	mutex_init(&group->mutex);
+	mutex_init(&group->sysfs_lazy.lock);
 	INIT_LIST_HEAD(&group->devices);
 	INIT_LIST_HEAD(&group->entry);
 	xa_init(&group->pasid_array);
@@ -1078,6 +1184,19 @@ struct iommu_group *iommu_group_alloc(void)
 		return ERR_PTR(ret);
 	}
 
+	/*
+	 * Defer reserved_regions and type - both are read only on
+	 * VFIO/iommufd ioctl paths, so let the populate callbacks
+	 * materialise them on demand.
+	 *
+	 * KERNFS_LAZY MUST be set before kobject_create_and_add("devices"),
+	 * which calls kernfs_inc_rev() on this kobject's directory and
+	 * would otherwise leave a stale negative dentry cached for a
+	 * missing reserved_regions/type child.  WARN_ON because
+	 * iommu_group_alloc() controls all group kobjects directly.
+	 */
+	WARN_ON(kernfs_set_lazy(group->kobj.sd));
+
 	group->devices_kobj = kobject_create_and_add("devices", &group->kobj);
 	if (!group->devices_kobj) {
 		kobject_put(&group->kobj); /* triggers .release & free */
@@ -1091,19 +1210,6 @@ struct iommu_group *iommu_group_alloc(void)
 	 */
 	kobject_put(&group->kobj);
 
-	ret = iommu_group_create_file(group,
-				      &iommu_group_attr_reserved_regions);
-	if (ret) {
-		kobject_put(group->devices_kobj);
-		return ERR_PTR(ret);
-	}
-
-	ret = iommu_group_create_file(group, &iommu_group_attr_type);
-	if (ret) {
-		kobject_put(group->devices_kobj);
-		return ERR_PTR(ret);
-	}
-
 	pr_debug("Allocated group %d\n", group->id);
 
 	return group;
-- 
2.47.3




Amazon Web Services Development Center Germany GmbH
Tamara-Danz-Str. 13
10243 Berlin
Geschaeftsfuehrung: Christof Hellmis, Andreas Stieger
Eingetragen am Amtsgericht Charlottenburg unter HRB 257764 B
Sitz: Berlin
Ust-ID: DE 365 538 597