tree 01f2d9eb9a23bd14575ace882b3b3cc8b60f62e6
parent da308e8da700637d6271dddda08128094691b980
author Alan Stern <stern@rowland.harvard.edu> 1151805901 -0400
committer Greg Kroah-Hartman <gregkh@suse.de> 1159383529 -0700

usbfs: private mutex for open, release, and remove

The usbfs code doesn't provide sufficient mutual exclusion among open,
release, and remove.  Release vs. remove is okay because they both
acquire the device lock, but open is not exclusive with either one.  All
three routines modify the udev->filelist linked list, so they must not
run concurrently.

Apparently someone gave this a minimum amount of thought in the past by
explicitly acquiring the BKL at the start of the usbdev_open routine.
Oddly enough, there's a comment pointing out that locking is unnecessary
because chrdev_open already has acquired the BKL.

But this ignores the point that the files in /proc/bus/usb/* are not
char device files; they are regular files and so they don't get any
special locking.  Furthermore it's necessary to acquire the same lock in
the release and remove routines, which the code does not do.

Yet another problem arises because the same file_operations structure is
accessible through both the /proc/bus/usb/* and /dev/usb/usbdev* file
nodes.  Even when one of them has been removed, it's still possible for
userspace to open the other.  So simple locking around the individual
remove routines is insufficient; we need to lock the entire
usb_notify_remove_device notifier chain.

Rather than rely on the BKL, this patch (as723) introduces a new private
mutex for the purpose.  Holding the BKL while invoking a notifier chain
doesn't seem like a good idea.


Signed-off-by: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>