On Tue, Jul 16, 2019 at 06:04:47PM +0800, Herbert Xu wrote: > On Mon, Jul 15, 2019 at 12:10:46PM -0400, Daniel Jordan wrote: > > > > I've been wrong before plenty of times, and there's nothing preventing this > > from being one of those times :) , but in this case I believe what I'm showing > > is correct. > > > > The padata_do_serial call for a given job ensures padata_reorder runs on the > > CPU that the job hashed to in padata_do_parallel, which is not necessarily the > > same CPU as the one that padata_do_parallel itself ran on. > > You're right. I was taking the comment in the code at face value, > never trust comments :) > > While looking at the code in question, I think it is seriously > broken. For instance, padata_replace does not deal with async > crypto at all. It would fail miserably if the underlying async > crypto held onto references to the old pd. Hm, yes looks like that. padata_replace should not call padata_free_pd() as long as the refcount is not zero. Currenlty padata_flush_queues() will BUG if there are references left. Maybe we can fix it if we call padata_free_pd() from padata_serial_worker() when it sent out the last object.
