On Fri, Jul 15, 2016 at 08:31:08AM +0200, Oliver Neukum wrote:
> > +static void ack_message(struct pd_sink_port *port, int msg_id)
> > +{
> > +   struct pd_msg_header *header = kzalloc(PD_MSG_HEADER_LEN, GFP_KERNEL);
> 
> This must be GFP_NOIO. We are in a cycle that can lead to deadlock.
> 
> Assume we are waiting for a request for more power to process IO
> which we need to ack.
> 
> 1. memory allocation leads to laundering, blocks on freeing memory
> 2. launderer decides to perform IO which needs more power
> 3. more power has already been requested, wait for it to be granted
> 
> 4. BANG - DEADLOCK
Agree, I'll change the GFP flag in next revision.

> > +   struct pd_msg_header *header = kzalloc(PD_MSG_HEADER_LEN +
> > +           port->nr_ps * PD_OBJ_SIZE, GFP_KERNEL);
> 
> Must be GFP_NOIO. For the same reason as above. We may be asked
> this to resolve a mismatch due to needing more power for IO.
Yes will do.

> > +static void handle_soft_reset(struct pd_sink_port *port)
> > +{
> > +   struct pd_msg_header *header = kzalloc(PD_MSG_HEADER_LEN, GFP_KERNEL);
> > +
> > +   if (!header)
> > +           return;
> > +
> > +   flush_workqueue(port->rx_wq);
> 
> That is problematic. We may be here precisely because something is wrong
> blocking progress. In particular what happens if another soft reset
> is queued?
I'm going to remove the workqueue.

> > +   struct pd_msg_header *header = kzalloc(PD_MSG_HEADER_LEN +
> > +                                   PD_OBJ_SIZE, GFP_KERNEL);
> 
> GFP_NOIO, same reasons
Yes.

> > +
> 
>       HTH
>               Oliver
Thanks for your review.

Reply via email to