[RFC] [PATCH] hvc_console: improve tty/console put_chars handling

[Hendrik Brueckner]

Hi folks,

On Tue, Jul 05, 2011 at 04:22:43PM +1000, Benjamin Herrenschmidt wrote:
> On Tue, 2011-07-05 at 04:17 +0000, Tabi Timur-B04825 wrote:
> > On Mon, Jul 4, 2011 at 9:24 AM, Hendrik Brueckner
> > <brueckner at linux.vnet.ibm.com> wrote:
> > 
> > I started that thread.  After much soul searching, we came to the
> > conclusion that HVC is not compatible with hypervisors that return
> > BUSY on writes. 
> 
> That is a fun conclusion considering that hvc has been written for the
> pseries hypervisor which ... can return BUSY on writes :-)
> 
> We just need to fix HVC properly.

So I took initiative and looked again into this issue.  Below you can
find a patch that is based on Ben's -EAGAIN idea.  The hvc console layer
takes care of retrying depending on the backend's return code.

However, the main issue is that from a backend perspective, there is no
difference between tty and console output.  Because consoles, especially
the preferred console, behave different than a simple tty.  Blocked write
to a preferred console can stop the system.

So with the patch below, the backend can now indirectly control the way
console output is handled for it.  I still have to think if this solution
is ok or if it is better to introduce a new callback to console output only
(and might provide a default implemenatation similar to the patch below).

NOTE: I did not yet test this patch but will do.. I just want to share it
      early to get feedback from you.

-->8---------------------------------------------------------------------

Currently, the hvc_console_print() function drops console output if the
hvc backend's put_chars() returns 0.  This patch changes this behavior
to allow a retry through returning -EAGAIN.

This change also affects the hvc_push() function.  Both functions are
changed to handle -EAGAIN and to retry the put_chars() operation.

If a hvc backend returns -EAGAIN, the retry handling differs:

  - hvc_console_print() spins to write the complete console output.
  - hvc_push() behaves the same way as for returning 0.

Now hvc backends can indirectly control the way how console output is
handled through the hvc console layer.

Signed-off-by: Hendrik Brueckner <brueckner at linux.vnet.ibm.com>
---
 drivers/tty/hvc/hvc_console.c |    8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

--- a/drivers/tty/hvc/hvc_console.c
+++ b/drivers/tty/hvc/hvc_console.c
@@ -163,8 +163,10 @@ static void hvc_console_print(struct con
 		} else {
 			r = cons_ops[index]->put_chars(vtermnos[index], c, i);
 			if (r <= 0) {
-				/* throw away chars on error */
-				i = 0;
+				/* throw away characters on error
+				 * but spin in case of -EAGAIN */
+				if (r != -EAGAIN)
+					i = 0;
 			} else if (r > 0) {
 				i -= r;
 				if (i > 0)
@@ -448,7 +450,7 @@ static int hvc_push(struct hvc_struct *h
 
 	n = hp->ops->put_chars(hp->vtermno, hp->outbuf, hp->n_outbuf);
 	if (n <= 0) {
-		if (n == 0) {
+		if (n == 0 || n == -EAGAIN) {
 			hp->do_wakeup = 1;
 			return 0;
 		}