Hi.

Le vendredi 17 mai 2013 20:47:32, vous avez écrit :
> On Fri, May 17, 2013 at 07:20:10PM +0200, Vincent Pelletier wrote:
> > If not, would it be possible to have a rw sysfs pseudofile per-device
> > (...per port ?) to enable DMADIR ?
> 
> Yeap, that sounds like the best we can do at this point.  Care to
> write up a patch?

First attempt attached (2/2).

Now that I got it working, I'm thining it would be better to automatically 
fallback to enabling DMADIR per-device level during initialisation (just like 
current fallbacks to 1.5Gbps and UDMA/33:PIO3, as visible in 1/2 commit 
message).
I believe it will slow down boot when such device is plugged in, though, any 
idea on how this can be avoided ?

> While better, please go into more details.  The problem here is that
> the bridge requires DMADIR and while libata makes use of DMADIR for
> regular commands, it doesn't do that for internal commands which are
> used during EH, right?

Just to be clear about EH: the timeout happens during device  initialisation 
(both at boot or on hotplug), not during error handling (or, maybe it happens 
in both places, but for the same reason).
I'm not comfortable with calling device discovery/initialisation as "error 
handling", but maybe it's unambiguous when used to libata.

I assume the failure to clear UNIT ATTENTION is just an error-on-error 
which gets fixed both because it wouldn't fail to clear (I didn't check) and 
because there is no error state to clear.

> Please go into full details of what's going on and be verbose.

3rd try attached (1/2).

Regards,
-- 
Vincent Pelletier