On Wed, May 23, 2018 at 01:45:30AM +0100, Al Viro wrote:

Oh, bugger...

wakeup
	removed from queue
	schedule __aio_poll_complete()

cancel
	grab ctx->lock
	remove from list
work
	aio_complete()
		check if it's in the list
		it isn't, move on to free the sucker
cancel
	call ->ki_cancel()
	BOOM

Looks like we want to call ->ki_cancel() *BEFORE* removing from the list,
as well as doing fput() after aio_complete().  The same ordering, BTW, goes
for aio_read() et.al.

Look:
CPU1:	io_cancel() grabs ->ctx_lock, finds iocb and removes it from the list.
CPU2:	aio_rw_complete() on that iocb.  Since the sucker is not in the list
anymore, we do NOT spin on ->ctx_lock and proceed to free iocb
CPU1:	pass freed iocb to ->ki_cancel().  BOOM.

BTW, it seems that the mainline is vulnerable to this one.  I might be
missing something, but...