summaryrefslogtreecommitdiffstats
path: root/drivers/hid/usbhid/hiddev.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/hid/usbhid/hiddev.c')
-rw-r--r--drivers/hid/usbhid/hiddev.c19
1 files changed, 17 insertions, 2 deletions
diff --git a/drivers/hid/usbhid/hiddev.c b/drivers/hid/usbhid/hiddev.c
index 433602aed46..c24d2fa3e3b 100644
--- a/drivers/hid/usbhid/hiddev.c
+++ b/drivers/hid/usbhid/hiddev.c
@@ -267,6 +267,7 @@ static int hiddev_open(struct inode *inode, struct file *file)
struct hiddev_list *list;
int res, i;
+ /* See comment in hiddev_connect() for BKL explanation */
lock_kernel();
i = iminor(inode) - HIDDEV_MINOR_BASE;
@@ -894,8 +895,22 @@ int hiddev_connect(struct hid_device *hid, unsigned int force)
hiddev->hid = hid;
hiddev->exist = 1;
- /* when lock_kernel() usage is fixed in usb_open(),
- * we could also fix it here */
+ /*
+ * BKL here is used to avoid race after usb_register_dev().
+ * Once the device node has been created, open() could happen on it.
+ * The code below will then fail, as hiddev_table hasn't been
+ * updated.
+ *
+ * The obvious fix -- introducing mutex to guard hiddev_table[]
+ * doesn't work, as usb_open() and usb_register_dev() both take
+ * minor_rwsem, thus we'll have ABBA deadlock.
+ *
+ * Before BKL pushdown, usb_open() had been acquiring it in right
+ * order, so _open() was safe to use it to protect from this race.
+ * Now the order is different, but AB-BA deadlock still doesn't occur
+ * as BKL is dropped on schedule() (i.e. while sleeping on
+ * minor_rwsem). Fugly.
+ */
lock_kernel();
retval = usb_register_dev(usbhid->intf, &hiddev_class);
if (retval) {