Trivial bugfix: opening a file that doesn't have ->direct_IO() with
O_DIRECT passed in flags fails (as it should) but doesn't call
->release() even though ->open() had been successful.

IOW, we have a leak there - e.g.  open() of a tty with O_DIRECT will
leak a reference to tty and tty_driver.  Besides, quite a few drivers
are _not_ happy with struct file being freed without ->release() after
->open() has returned 0.