[Linux-ia64] heads up: i/o mmu related changes

[Linux-ia64] heads up: i/o mmu related changes

[David Mosberger]

I justed checked in a changeset
(davidm@tiger.hpl.hp.com|ChangeSet|20030610070955|57077) to make the
kernel work better for machines with I/O MMU.  This should fix the
long-standing bug where IDE disks on machines with memory above 4GB
would MCA if the machine has an I/O MMU.

The trick was to make PCI_DMA_BUS_IS_PHYS a variable which is set at
runtime.  Machines with I/O MMUs or machines with PCI buses which live
in separate address spaces MUST clear this variable.  I updated the
code for the zx1 I/O MMU.  I didn't touch the SGI I/O MMU code,
though, as I wouldn't be able to test it.  Also, it would be good if
someone could test whether IDE disks work fine on machines with
Intel-chipsets and memory above 4GB.

	--david

Re: [Linux-ia64] heads up: i/o mmu related changes

[David Mosberger]

>>>>> On Tue, 10 Jun 2003 01:07:18 -0700, David Mosberger <davidm@linux.hpl.hp.com> said:

  David> I justed checked in a changeset
  David> (davidm@tiger.hpl.hp.com|ChangeSet|20030610070955|57077) to
  David> make the kernel work better for machines with I/O MMU.  This
  David> should fix the long-standing bug where IDE disks on machines
  David> with memory above 4GB would MCA if the machine has an I/O
  David> MMU.

  David> The trick was to make PCI_DMA_BUS_IS_PHYS a variable which is
  David> set at runtime.  Machines with I/O MMUs or machines with PCI
  David> buses which live in separate address spaces MUST clear this
  David> variable.  I updated the code for the zx1 I/O MMU.  I didn't
  David> touch the SGI I/O MMU code, though, as I wouldn't be able to
  David> test it.

Actually, you may want to hold off on this one.  I think I'll need to
fine-tune this a bit more.  My current thinking is that we need a
variable ia64_max_iommu_page_size or something like that, which will
tell us what to set BIO_VMERGE_BOUNDARY to.

	--david

Re: [Linux-ia64] heads up: i/o mmu related changes

[Jesse Barnes]

On Tue, Jun 10, 2003 at 10:36:57AM -0700, David Mosberger wrote:
> >>>>> On Tue, 10 Jun 2003 01:07:18 -0700, David Mosberger <davidm@linux.hpl.hp.com> said:
> 
>   David> I justed checked in a changeset
>   David> (davidm@tiger.hpl.hp.com|ChangeSet|20030610070955|57077) to
>   David> make the kernel work better for machines with I/O MMU.  This
>   David> should fix the long-standing bug where IDE disks on machines
>   David> with memory above 4GB would MCA if the machine has an I/O
>   David> MMU.
> 
>   David> The trick was to make PCI_DMA_BUS_IS_PHYS a variable which is
>   David> set at runtime.  Machines with I/O MMUs or machines with PCI
>   David> buses which live in separate address spaces MUST clear this
>   David> variable.  I updated the code for the zx1 I/O MMU.  I didn't
>   David> touch the SGI I/O MMU code, though, as I wouldn't be able to
>   David> test it.
> 
> Actually, you may want to hold off on this one.  I think I'll need to
> fine-tune this a bit more.  My current thinking is that we need a
> variable ia64_max_iommu_page_size or something like that, which will
> tell us what to set BIO_VMERGE_BOUNDARY to.

Ok, the changes for sn2 look pretty straightforward.  FYI, our IOMMU
supports 4k and 16k pages.

Thanks,
Jesse