char: xillybus: Prevent use-after-free due to race condition

The driver for XillyUSB devices maintains a kref reference count on each
xillyusb_dev structure, which represents a physical device. This reference
count reaches zero when the device has been disconnected and there are no
open file descriptors that are related to the device. When this occurs,
kref_put() calls cleanup_dev(), which clears up the device's data,
including the structure itself.

However, when xillyusb_open() is called, this reference count becomes
tricky: This function needs to obtain the xillyusb_dev structure that
relates to the inode's major and minor (as there can be several such).
xillybus_find_inode() (which is defined in xillybus_class.c) is called
for this purpose. xillybus_find_inode() holds a mutex that is global in
xillybus_class.c to protect the list of devices, and releases this
mutex before returning. As a result, nothing protects the xillyusb_dev's
reference counter from being decremented to zero before xillyusb_open()
increments it on its own behalf. Hence the structure can be freed
due to a rare race condition.

To solve this, a mutex is added. It is locked by xillyusb_open() before
the call to xillybus_find_inode() and is released only after the kref
counter has been incremented on behalf of the newly opened inode. This
protects the kref reference counters of all xillyusb_dev structs from
being decremented by xillyusb_disconnect() during this time segment, as
the call to kref_put() in this function is done with the same lock held.

There is no need to hold the lock on other calls to kref_put(), because
if xillybus_find_inode() finds a struct, xillyusb_disconnect() has not
made the call to remove it, and hence not made its call to kref_put(),
which takes place afterwards. Hence preventing xillyusb_disconnect's
call to kref_put() is enough to ensure that the reference doesn't reach
zero before it's incremented by xillyusb_open().

It would have been more natural to increment the reference count in
xillybus_find_inode() of course, however this function is also called by
Xillybus' driver for PCIe / OF, which registers a completely different
structure. Therefore, xillybus_find_inode() treats these structures as
void pointers, and accordingly can't make any changes.

Reported-by: Hyunwoo Kim <[e-mail]>
Suggested-by: Alan Stern <[e-mail]>
Signed-off-by: Eli Billauer <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Greg Kroah-Hartman <[e-mail]>

drivers/char/xillybus/xillyusb.c

···
 	struct mutex process_in_mutex; /* synchronize wakeup_all() */
 };
 
+/*
+ * kref_mutex is used in xillyusb_open() to prevent the xillyusb_dev
+ * struct from being freed during the gap between being found by
+ * xillybus_find_inode() and having its reference count incremented.
+ */
+
+static DEFINE_MUTEX(kref_mutex);
+
 /* FPGA to host opcodes */
 enum {
 	OPCODE_DATA = 0,
···
 	int rc;
 	int index;
 
+	mutex_lock(&kref_mutex);
+
 	rc = xillybus_find_inode(inode, (void **)&xdev, &index);
-	if (rc)
+	if (rc) {
+		mutex_unlock(&kref_mutex);
 		return rc;
+	}
+
+	kref_get(&xdev->kref);
+	mutex_unlock(&kref_mutex);
 
 	chan = &xdev->channels[index];
 	filp->private_data = chan;
···
 	if (((filp->f_mode & FMODE_READ) && chan->open_for_read) ||
 	    ((filp->f_mode & FMODE_WRITE) && chan->open_for_write))
 		goto unmutex_fail;
-
-	kref_get(&xdev->kref);
 
 	if (filp->f_mode & FMODE_READ)
 		chan->open_for_read = 1;
···
 	return rc;
 
 unmutex_fail:
+	kref_put(&xdev->kref, cleanup_dev);
 	mutex_unlock(&chan->lock);
 	return rc;
 }
···
 
 	xdev->dev = NULL;
 
+	mutex_lock(&kref_mutex);
 	kref_put(&xdev->kref, cleanup_dev);
+	mutex_unlock(&kref_mutex);
 }
 
 static struct usb_driver xillyusb_driver = {