On Thu, 2009-10-15 at 13:57 -0500, Scott Wood wrote:

I'd say the dropping approach is quite undesirable (significant 
potential for output loss unless the buffer is huge), unless there's 
simply no way to safely spin.  Hopefully there are no such backends, but 
if there are perhaps we can have them return some special code to 
indicate that.

Should never spin. Best is to keep a copy in the upper layer of the
pending data and throttle (not accept further data from tty layer) until
we have managed to flush out that "pending" buffer.

quoted

If we just busy loop, it actually does not matter how we let hvc_console react 
on 0, as long as we adopt all backends to use that interface consistent.

On the other hand, backends might want to do special magic on congestion so I 
personally tend to let the backend loop instead of hvc_console. But I am really  
not sure.

Doing it in the backend requires the backend to know whether it's being 
called for printk or for user I/O.  In the latter case, we don't want to 
spin, but rather wait for an IRQ (or poll with a timer if there's no IRQ). 

Ben.