[PATCH][RFC] fix a race between bsg_open() and bsg_unregister_queue()

[Al Viro <viro@xxxxxxxxxxxxxxxxxx>]

	Consider the following scenario:

task A: open() on /dev/bsg/<something>
	calls chrdev_open()
	finds and grabs a reference to bsg_device.cdev in inode->i_cdev
	refcount on that cdev is 2 now (1 from creation + 1 we'd just grabbed)
	calls bsg_open().
	fetches to_bsg_device(inode)->queue - that would be ->queue in the
	same bsg_device instance.
	gets preempted away and loses CPU before it gets to calling blk_get_queue().

task B: calls bsg_unregister_queue() on the same queue, which calls
	cdev_device_del(), which makes cdev impossible to look up and
	drops the reference to that cdev; refcount is 1 now, so nothing gets
	freed yet.
	caller of bsg_unregister_queue() proceeds to destroy the queue and
	free it, allowing reuse of memory that used to contain it.

task A: regains CPU
	calls blk_get_queue() on something that no longer points to
	a request_queue instance.  In particular, "dying" flag is no longer
	guaranteed to be there, so we proceed to increment what we think is
	a queue refcount, corrupting whatever lives in that memory now.

Usually we'll end up with memory not reused yet, and blk_get_queue() will
fail without buggering anything up.  Not guaranteed, though...

AFAICS, the fact that request_queue freeing is RCU-delayed means that
it can be fixed by the following:
	* mark bsg_device on bsg_unregister_queue() as goner
	* have bsg_open() do rcu_read_lock(), then check that flag and do
blk_get_queue() only if the flag hadn't been set yet.  If we did not observe
the flag after rcu_read_lock(), we know that queue have been freed yet
- RCU delay couldn't have run out.

Comments?

Signed-off-by: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
---
diff --git a/block/bsg.c b/block/bsg.c
index 2ab1351eb082..643641087691 100644
--- a/block/bsg.c
+++ b/block/bsg.c
@@ -28,6 +28,7 @@ struct bsg_device {
 	unsigned int timeout;
 	unsigned int reserved_size;
 	bsg_sg_io_fn *sg_io_fn;
+	bool goner;
 };
 
 static inline struct bsg_device *to_bsg_device(struct inode *inode)
@@ -71,9 +72,14 @@ static int bsg_sg_io(struct bsg_device *bd, fmode_t mode, void __user *uarg)
 
 static int bsg_open(struct inode *inode, struct file *file)
 {
-	if (!blk_get_queue(to_bsg_device(inode)->queue))
-		return -ENXIO;
-	return 0;
+	struct bsg_device *bd = to_bsg_device(inode);
+	int err = 0;
+
+	rcu_read_lock();
+	if (bd->goner || !blk_get_queue(bd->queue))
+		err = -ENXIO;
+	rcu_read_unlock();
+	return err;
 }
 
 static int bsg_release(struct inode *inode, struct file *file)
@@ -175,6 +181,7 @@ static void bsg_device_release(struct device *dev)
 
 void bsg_unregister_queue(struct bsg_device *bd)
 {
+	bd->goner = true;
 	if (bd->queue->kobj.sd)
 		sysfs_remove_link(&bd->queue->kobj, "bsg");
 	cdev_device_del(&bd->cdev, &bd->device);