On Sat,  3 Aug 2019 07:01:53 -0700 Tejun Heo [off-list ref] wrote:

There currently is no way to universally identify and lookup a bdi
without holding a reference and pointer to it.  This patch adds an
non-recycling bdi->id and implements bdi_get_by_id() which looks up
bdis by their ids.  This will be used by memcg foreign inode flushing.

Why is the id non-recycling?  Presumably to address some
lifetime/lookup issues, but what are they?

Why was the IDR code not used?

I left bdi_list alone for simplicity and because while rb_tree does
support rcu assignment it doesn't seem to guarantee lossless walk when
walk is racing aginst tree rebalance operations.

...

+/**
+ * bdi_get_by_id - lookup and get bdi from its id
+ * @id: bdi id to lookup
+ *
+ * Find bdi matching @id and get it.  Returns NULL if the matching bdi
+ * doesn't exist or is already unregistered.
+ */
+struct backing_dev_info *bdi_get_by_id(u64 id)
+{
+	struct backing_dev_info *bdi = NULL;
+	struct rb_node **p;
+
+	spin_lock_irq(&bdi_lock);

Why irq-safe?  Everywhere else uses spin_lock_bh(&bdi_lock).

+	p = bdi_lookup_rb_node(id, NULL);
+	if (*p) {
+		bdi = rb_entry(*p, struct backing_dev_info, rb_node);
+		bdi_get(bdi);
+	}
+	spin_unlock_irq(&bdi_lock);
+
+	return bdi;
+}
+

...