On Tue, Jan 28, 2014 at 06:25:39AM +0000, Raghu Gandham wrote:

Date:   Tue, 28 Jan 2014 06:25:39 +0000
From: Raghu Gandham <redacted>
To: Aaro Koskinen <aaro.koskinen@iki.fi>, Dmitry Torokhov
 [off-list ref]
CC: "linux-input@vger.kernel.org" <redacted>,
 "linux-mips@linux-mips.org" [off-list ref],
 "linux-kernel@vger.kernel.org" [off-list ref]
Subject: RE: [PATCH] Input: i8042-io - Exclude mips platforms when
 allocating/deallocating IO regions.
Content-Type: text/plain; charset="us-ascii"

Hi Aaro,

quoted

On Sun, Jan 26, 2014 at 10:56:38PM -0800, Dmitry Torokhov wrote:

quoted

On Mon, Jan 27, 2014 at 12:32:36AM +0000, Raghu Gandham wrote:

quoted

quoted

On Sat, Jan 25, 2014 at 11:01:54AM -0800, Raghu Gandham wrote:

quoted

The standard IO regions are already reserved by the platform
code on most MIPS devices(malta, cobalt, sni). The Commit
197a1e96c8be5b6005145af3a4c0e45e2d651444
("Input: i8042-io - fix up region handling on MIPS") introduced
a bug on these MIPS platforms causing i8042 driver to fail when
trying to reserve IO ports.
Prior to the above mentioned commit request_region is skipped on
MIPS but release_region is called.

This patch reverts commit
197a1e96c8be5b6005145af3a4c0e45e2d651444
and also avoids calling release_region for MIPS.

The problem is that IO regions are reserved on _most_, but not
_all_ devices.
MIPS should figure out what they want to do with i8042 registers
and be consistent on all devices.

Please examine the attached patch which went upstream in April of 2004.
Since then it had been a convention not to call request_region
routine in
i8042 for MIPS. The attached patch had a glitch that it guarded
request_region in i8042-io.h but skipped guarding release_region in
i8042-io.h. I believe that the issue Aaro saw was due to this
glitch. Below is the error quoted in Aaro's commit message.

    [    2.112000] Trying to free nonexistent resource <0000000000000060-

000000000000006f>

quoted

quoted

My patch reinstates the convention followed on MIPS devices along
with fixing Aaro's issue.

I assume that Aaro did test his patch and on his box request_region()
succeeds. That would indicate that various MIPS sub-arches still not
settled on the topic.

request_region() succeeds on Loongson and OCTEON.

This would mean that before your patch in oct of 2012, Loongson and Octeon 
were not reserving the IO space for i8042.

quoted

On OCTEONs without PCI, request_region() will fail which is correct as there
is no I/O space.

I wasn't aware of that 2004 patch (it pre-dates GIT history of mainline Linux).
Why the regions are already reserved by the platform code?

The only information I have is the comment before request_region in i8042-io.h that
touching data register on some platforms is flaky.  If your patch was primarily aimed at
addressing the error message from release_region, the current patch I uploaded should
also take care of it too. 

I think the patch (http://patchwork.linux-mips.org/patch/6419/) should be
applied.  The argumentation for reserving ports in the platform code are
the same as on x86 - touch the registers and bad things may happen.  This
is because a fair number of older MIPS platforms were based on x86 chipsets
or at least are using very similar designs.

The fact that on certain platforms such as Loongson, some Octeon systems
and others the request_region() call to reserve the keyboard call succeeds
is by accident not design.

I wish i8042.c was a real platform driver, not using platform_create_bundle.
That would leave a natural place in the arch/platform code to deal with
I/O port allocation and platform_device creation, as necessary for a
platform.

  Ralf