On Tue, 14 Jan 2020, Dan Carpenter wrote: > I'm trying to take a stab at diagnosing these bugs. (But I'm seldom > smart enough to actually fix anything). These hiddev_disconnect() bugs > are a race condition: > > devel/drivers/hid/usbhid/hiddev.c > 924 void hiddev_disconnect(struct hid_device *hid) > 925 { > 926 struct hiddev *hiddev = hid->hiddev; > 927 struct usbhid_device *usbhid = hid->driver_data; > 928 > 929 usb_deregister_dev(usbhid->intf, &hiddev_class); > 930 > 931 mutex_lock(&hiddev->existancelock); > 932 hiddev->exist = 0; > ^^^^^^^^^^^^^^^^^^ > We set "exist = 0;" > > 933 > 934 if (hiddev->open) { > 935 mutex_unlock(&hiddev->existancelock); > ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ > Then we drop the lock. > > 936 hid_hw_close(hiddev->hid); > 937 wake_up_interruptible(&hiddev->wait); > ^^^^^^ > The other thread frees hiddev and it crashes here. > > 938 } else { > 939 mutex_unlock(&hiddev->existancelock); > 940 kfree(hiddev); > 941 } > 942 } > > The other thread is doing: > > drivers/hid/usbhid/hiddev.c > 216 static int hiddev_release(struct inode * inode, struct file * file) > 217 { > 218 struct hiddev_list *list = file->private_data; > 219 unsigned long flags; > 220 > 221 spin_lock_irqsave(&list->hiddev->list_lock, flags); > 222 list_del(&list->node); > 223 spin_unlock_irqrestore(&list->hiddev->list_lock, flags); > 224 > 225 mutex_lock(&list->hiddev->existancelock); > ^^^^^^^^^^^^^^^^^^^^^^^^^^^ > Takes the lock. > > 226 if (!--list->hiddev->open) { > ^^^^^^^^^^^^^^^^^^^^ > Decrements open. > > 227 if (list->hiddev->exist) { > ^^^^^^^^^^^^^^^^^^^ > This is false. > > 228 hid_hw_close(list->hiddev->hid); > 229 hid_hw_power(list->hiddev->hid, PM_HINT_NORMAL); > 230 } else { > 231 mutex_unlock(&list->hiddev->existancelock); > 232 kfree(list->hiddev); > ^^^^^^^^^^^^ > Freed here. > > 233 vfree(list); > 234 return 0; > 235 } > 236 } > 237 > 238 mutex_unlock(&list->hiddev->existancelock); > 239 vfree(list); > 240 > 241 return 0; > 242 } > > I'm not sure what the fix should be though. Good sleuthing! It certainly looks as though hiddev_disconnect() should not try to reference hiddev after dropping the existance_lock -- at least, not in the case that the device file is still open. So I suggest reordering the statements in that routine: First call hid_hw_close() and wake_up_interruptible(), then drop the mutex. Too bad syzbot doesn't have a reproducer to test this, though. Alan Stern