Al Viro root-caused a race where the IOCB_CMD_POLL handling of
fget/fput() could cause us to access the file pointer after it had
already been freed:

 "In more details - normally IOCB_CMD_POLL handling looks so:

   1) io_submit(2) allocates aio_kiocb instance and passes it to
      aio_poll()

   2) aio_poll() resolves the descriptor to struct file by req->file =
      fget(iocb->aio_fildes)

   3) aio_poll() sets ->woken to false and raises ->ki_refcnt of that
      aio_kiocb to 2 (bumps by 1, that is).

   4) aio_poll() calls vfs_poll(). After sanity checks (basically,
      "poll_wait() had been called and only once") it locks the queue.
      That's what the extra reference to iocb had been for - we know we
      can safely access it.

   5) With queue locked, we check if ->woken has already been set to
      true (by aio_poll_wake()) and, if it had been, we unlock the
      queue, drop a reference to aio_kiocb and bugger off - at that
      point it's a responsibility to aio_poll_wake() and the stuff
      called/scheduled by it. That code will drop the reference to file
      in req->file, along with the other reference to our aio_kiocb.

   6) otherwise, we see whether we need to wait. If we do, we unlock the
      queue, drop one reference to aio_kiocb and go away - eventual
      wakeup (or cancel) will deal with the reference to file and with
      the other reference to aio_kiocb

   7) otherwise we remove ourselves from waitqueue (still under the
      queue lock), so that wakeup won't get us. No async activity will
      be happening, so we can safely drop req->file and iocb ourselves.

  If wakeup happens while we are in vfs_poll(), we are fine - aio_kiocb
  won't get freed under us, so we can do all the checks and locking
  safely. And we don't touch ->file if we detect that case.

  However, vfs_poll() most certainly *does* touch the file it had been
  given. So wakeup coming while we are still in ->poll() might end up
  doing fput() on that file. That case is not too rare, and usually we
  are saved by the still present reference from descriptor table - that
  fput() is not the final one.

  But if another thread closes that descriptor right after our fget()
  and wakeup does happen before ->poll() returns, we are in trouble -
  final fput() done while we are in the middle of a method:

Al also wrote a patch to take an extra reference to the file descriptor
to fix this, but I instead suggested we just streamline the whole file
pointer handling by submit_io() so that the generic aio submission code
simply keeps the file pointer around until the aio has completed.

Fixes: bfe4037e ("aio: implement IOCB_CMD_POLL")
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Reported-by: syzbot+503d4cc169fcec1cb18c@syzkaller.appspotmail.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>