Hi Adrian,

[fixed up Matsubara-sans address]

On Feb 17, 2008 4:48 AM, Adrian McMenamin
<[EMAIL PROTECTED]> wrote:
> On Sat, 2008-02-16 at 18:38 +0000, Adrian McMenamin wrote:
> > Will seek to bisect this, but I have just updated my sources to the
> > latest git and it is not booting at all on the Dreamcast.
> >
> > With early printk on, I get nothing more than this before an instant
> > reboot:
> >
> > [    0.000000] Linux version 2.6.25-rc2-10953-g52065cd
> > ([EMAIL PROTECTED]) (gcc version 3.4.6) #511 PREEMPT Sat Feb 16 18:31:43
> > GMT 2008
> > [    0.000000] console [sercon0] enabled
> > [    0.000000] Booting machvec: Sega Dreamcast
>
>
> [EMAIL PROTECTED]:~/gdrom-dev$ git bisect good
> e036eaa681a17f71b64f6d9040fe605555623919 is first bad commit
> commit e036eaa681a17f71b64f6d9040fe605555623919
> Author: Magnus Damm <[EMAIL PROTECTED]>
> Date:   Thu Feb 14 13:52:43 2008 +0900
>
>     sh: use ctrl_in/out for on chip pci access
>
>     This patch makes sure ctrl_inN/outN are used instead of inN/outN for
> on chip
>     pci registers. Without this patch addresses may be adjusted using
> the value
>     in generic_io_base. This patch makes it possible to set
> generic_io_base and
>     have pci without reading and writing all over the place.
>
>     Signed-off-by: Magnus Damm <[EMAIL PROTECTED]>
>     Acked-by: Katsuya MATSUBARA <[EMAIL PROTECTED]>
>     Signed-off-by: Paul Mundt <[EMAIL PROTECTED]>
>
> :040000 040000 17a9f8181301e3a082d8a1a2fdef9b13ff1185e4
> 87e4a2e912be56b0d12e8a92db9489d6615c31f7 M      arch

Thanks for tracking this down and sorry for breaking the dreamcast pci driver.

Most code for the dreamcast currently do direct register access using
ctrl_inN/outN or readN/writeN. Only a few places use inN/outN and
depend on the value of generic_io_base. Doing inN/outN is ok (but
outdated, use ioreadN/iowriteN instead) in a portable pci/isa driver,
but it doesn't make sense for any dreamcast specific code to rely on
generic_io_base. For the dreamcast we already know the address at
compile time so doing adjustment during runtime is just unnecessary
overhead.

It is of course possible to revert the dreamcast-specific bits of the
commit you pointed out, but I think the change for the dreamcast makes
sense since it makes the code both faster and more uniform.

However, at this point it is broken. Sorry about that. I just looked
through all upstream dreamcast-specific code I could find and irq.c is
now the only file that does inN/outN. I've attached a patch that makes
the irq code independent of generic_io_base. There is also another
patch attached that adjusts the addresses of the pci register.
Together they should solve the problem you are seeing. Please try them
on top of 2.6.25-rc2.

With these patches io ports count from 0 -> 8k-1 instead of being a
pointer. We currently rely on generic_io_base logic but that will be
changed in the future.

Please let me know the results. Thank you.

/ magnus

Attachment: linux-2.6.25-rc2-sh-dreamcast-irq-20080218.patch
Description: Binary data

Attachment: linux-2.6.25-rc2-sh-dreamcast-pci-20080218b.patch
Description: Binary data

Reply via email to