Re: [PATCH] Correctly flush 8250 buffers, notify ldisc of line status changes.

[Alan Cox <alan@lxorguk.ukuu.org.uk>]

On Maw, 2004-11-09 at 14:39, David Woodhouse wrote:
> Actually I needed it to respond to CTS going away, and it provides a
> notification *before* the data are *sent*. Which lets me know that the
> first bytes of my packet were dropped by the automatic contention
> detection circuitry and I need to flush the rest of the packet from the
> FIFO rather than letting the hardware driver wait for CTS to come back
> then then send a corrupt half-packet.

But you can't flush the fifo from that callback, you don't have any
locking on it. What are you locking semantics ? Define them, versus
open, versus close, versus other I/O.

> That solves a different problem, and isn't quite as useful to me. I want
> to be able to respond to CTS going low as quickly as possible, by
> flushing the rest of the characters from the outgoing queue. I was happy
> enough with using a tasklet to actually call the flush method, to avoid
> the deadlock you pointed out without changing the locking of all the
> hardware drivers). 

So you want to replace a tasklet that responds to the event with a
tasklet which is called by the event ? Tell me how they differ when low
latency is set on the tty - I don't see any difference in performance

> > Andrew - please reject the patch.
> 
> I'll submit the bit which makes the flush_buffer method work on its own.
> Alan, would you care to offer a viable alternative which solves the
> problem I'm interested in?

If you can't define the locking semantics, its not viable anyway. I see
two things we can do usefully here. The first is to teach
tty_flip_buffer_push more about urgency - since the new tty code I'm
running/crashing/debugging has a real packet queue not just flip buffers
we can queue a tiny message to the queue front and we can probably begin
to cheat a bit more on low_latency v immediate.