AACI broken with commit 29a4f2d3

[Takashi Iwai]

At Mon, 29 Mar 2010 09:57:57 +0200,
I wrote:
> 
> At Mon, 29 Mar 2010 13:15:22 +0530,
> Philby John wrote:
> > 
> > On 03/27/2010 02:41 AM, Takashi Iwai wrote:
> > > At Fri, 26 Mar 2010 21:37:51 +0530,
> > > Philby John wrote:
> > >>
> > >> On Fri, 2010-03-26 at 14:12 +0000, Catalin Marinas wrote:
> > >>> On Fri, 2010-03-26 at 14:08 +0000, Mark Brown wrote:
> > >>>> On Fri, Mar 26, 2010 at 01:54:45PM +0000, Catalin Marinas wrote:
> > >>>>
> > >>>>> But the above says "the power down control and status register (0x26) of
> > >>>>> the CODEC". So this refers to the AC97 registers rather than the AACI
> > >>>>> registers. Your patch reads from the AACI registers. The AC97 registers
> > >>>>> I think are access with aaci_ac97_(read|write) functions.
> > >>>>
> > >>>> Yes, they are - but note that some CODECs will power up in low power
> > >>>> mode and therefore attempts to read immediately after the controller
> > >>>> probe function starts executing may fail until the controller has issued
> > >>>> a warm reset.
> > >>>
> > >>> Yes, possibly. But my point is that accessing offset 0x26 in the AACI
> > >>> register space has nothing to do with the AC97 power register. At offset
> > >>> 0x26 in the AACI register space you find the top part of the AACIIE2
> > >>> register (if you can even read this as a half-word).
> > >>>
> > >> > From b411099000bbbb9b076168ee98742a36018a67ac Mon Sep 17 00:00:00 2001
> > >> From: Philby John<pjohn at in.mvista.com>
> > >> Date: Fri, 26 Mar 2010 16:41:06 +0530
> > >> Subject: [PATCH] Fix alignment faults on ARM Cortex introduced by commit 29a4f2d3
> > >>
> > >> The commit 29a4f2d3 used writel() at offset 0x26 which is
> > >> half-word aligned causing unaligned exceptions on a
> > >> Cortex-A8. The original patch solved the "aaci-pl041 fpga:04:
> > >> ac97 read back fail" issue on a soft reset. Reading from any
> > >> arbitrary aaci register seems to solve this issue.
> > >
> > > Then, isn't this a generic problem like PCI write-posting?
> > >
> > 
> > Yes, it does look like. Takashi, do you know of a better way than 
> > reading from an aaci register?
> 
> Well, in the case of PCI, you'd need to read the just-written register
> value again before any time-sensitive operation such as reset via
> re-asserting the register bit.
> 
> But, I'm really not sure whether this is the case, too...

This issue is still pending.
Can anyone give a proper patch with a rational explanation?

Philby's last patch looks OK to apply wrt coding (at least better than
the original one), but it doesn't clarify exactly why.


thanks,

Takashi