On Sun, Nov 13, 2022 at 11:03:20AM +0200, Eli Billauer wrote: > On 13/11/2022 10:47, Hyunwoo Kim wrote: > > And, even if the mutex_unlock(&unit_mutex); of xillybus_find_inode() > > is finally moved, xdev may be released before kref_get() is executed > > if xillyusb_disconnect() ends just before the function returns. > > (Of course, this is an extremely rare case.) > > > > So, in xillyusb_open() we need to move kref_get() above xillybus_find_inode(). > > First of all, that's impossible. kref_get() is called on a member of a > specific @xdev's struct, and it's xillybus_find_inode()'s job to find it. So > before the call to xillybus_find_inode(), we don't know which @xdev it is. > That's the tricky part of all this. > > The solution of this submitted patch was a lock that briefly prevents the > kref_put() of all @xdevs. The way it works is that if an @xdev is found by > xillybus_find_inode(), it necessarily means that xillyusb_disconnect()'s > call to xillybus_cleanup_chrdev() hasn't returned (yet). Therefore, holding > @kref_mutex guarantees that the kref_put() call, which is later on, isn't > reached for the @xdev that has been found. > > If you've found a flaw in this mechanism, please be more specific about it. you're right. It seems that you only need to move the location of unit_mutex in xillybus_find_inode(). Regards, Hyunwoo Kim
