by Bjorn Helgaas

[permalink] [raw]

Subject: Re: [Resend PATCH 5/5] IA64/PCI/ACPI: Rework PCI root bridge ACPI resource conversion

On Thu, Oct 31, 2013 at 10:26:03AM +0800, Lan Tianyu wrote:
> On 2013年10月31日 00:23, Bjorn Helgaas wrote:
> > On Wed, Oct 30, 2013 at 2:34 AM, Lan Tianyu <[email protected]> wrote:
> >> On 2013年10月29日 01:32, Bjorn Helgaas wrote:
> >>>
> >>> On Sat, Oct 26, 2013 at 10:53 AM, Lan Tianyu <[email protected]> wrote:
> >>>>
> >>>> On 10/24/2013 06:39 AM, Bjorn Helgaas wrote:
> >>>>>
> >>>>> On Fri, Oct 18, 2013 at 08:44:12PM +0800, Lan Tianyu wrote:
> >>>>>>
> >>>>>>
> >>>>>> On 10/18/2013 04:33 AM, Bjorn Helgaas wrote:
> >>>
> >>>
> >>>>>>> I wonder if it would make sense to make
> >>>>>>> acpi_dev_resource_address_space() ignore addr.translation_offset for
> >>>>>>> IO resources. Or maybe ignore it if the _TTP (type translation) bit
> >>>>>>> is set?
> >>>>>>
> >>>>>>
> >>>>>>
> >>>>>> I wonder why current code doesn't check _TTP? The code in the
> >>>>>> add_io_space() seems to think _TTP is always set, right?
> >>>>>
> >>>>>
> >>>>> I think it's an oversight, and you should fix it. I suggest that you
> >>>>> ignore the _TRA value when _TTP is set. Obviously this only applies
> >>>>> to I/O port resources, since _TTP is only defined in the I/O Resource
> >>>>> Flag (Table 6-185 in ACPI 5.0 spec).
> >>>>
> >>>>
> >>>> _TTP is also defined in the Memory Resource flag, Please have a look at
> >>>> Table 6-184 in the ACPI 5.0 Spec.
> >>>
> >>>
> >>> Yes, you're right. That would be for a host bridge that converts I/O
> >>> on the primary (upstream) side of the bridge to memory on the PCI
> >>> side. I've never seen such a bridge, and I can't really imagine why
> >>> anybody would do that. But I guess you should be able to safely
> >>> ignore _TRA when _TTP is set in either a MEM or IO descriptor, because
> >>> the same reasoning should apply to both.
> >>>
> >>>> I am not sure how to deal with _TTP unsetting io resource? _TTP unsetting
> >>>> mean the resource is IO on the primary side and also IO on the secondary
> >>>> side.
> >>>
> >>>
> >>> If _TTP is not set, I guess you would apply _TRA. That's what you
> >>> already do for MEM descriptors, and think you should just do the same
> >>> for IO descriptors. I would guess that having _TTP = 0 and _TRA != 0
> >>> is rare for IO descriptors, but I suppose it could happen.
> >>
> >>
> >> Yes, my concern is for the IO resource case of _TTP=0 and _TRA !=0. The
> >> only reason for this case I think of is that the IO resource offsets on
> >> the prime bus and second bus are different. In this case, we still need
> >> to pass _TRA to new_space() and the finial resource->start still should be
> >> acpi_resource->min + offset returned by add_io_space(), right?
> >
> > No, I don't think so. If the "phys_base" argument to new_space() is
> > non-zero, it is the base of an MMIO region that needs to be
> > ioremapped. This is handling the _TTP=1 case, where the MMIO region
> > is translated by the bridge into an IO region on PCI.
> >
> > If _TTP=0, the region is IO on both the upstream and downstream sides
> > of the host bridge, and we don't want to ioremap a new MMIO region for
> > it. It might be part of the "legacy I/O port space," but that's
> > already covered elsewhere.
> >
> > I don't think we need to add special handling for the _TTP=0 and _TRA
> > != 0 case because I don't think it exists in the field. If and when
> > it *does* exist, we'll know what to do. In the meantime, it should
> > look just like the MEM path.
>
>
> OK. I get it. acpi_dev_resource_address_space() will only apply _TRA to
> resource ->start and ->end for both mem and io resource when _TTP=0. In
> the add_window(), the offset returned by add_io_space() will be added
> directly to ->start and ->end.
>
> add_window() {
> ...
> if (resource->flags & IORESOURCE_MEM) {
> root = &iomem_resource;
> offset = addr.translation_offset;

I can wait for your patch to see the whole thing, but I would expect
"offset = 0" here. For MEM resources, the arch code should not need to
look inside "addr" at all.

> } else if (resource->flags & IORESOURCE_IO) {
> root = &ioport_resource;
>
> offset = add_io_space(info, &addr);
> if (offset == ~0)
> return AE_OK;
>
> resource->start += offset;
> resource->end += offset;
> } else
> return AE_OK;
>
> ...
> }
>
> >
> >> If yes, I think _TRA can't be applied to IO resource in the
> >> acpi_dev_resource_address_space() regardless of the value of _TTP.
> >>
> >> BTW, Translation Sparse(_TRS) is only meaningful if _TTP is set.(Table
> >> 6-185). The add_io_space() doesn't check _TTP when set sparse. So this
> >> should be corrected?
> >
> > Sure, I'm OK with this. It's possible we could trip over a BIOS bug
> > where _TRS=1 but _TTP=0, but I think the risk is low because only
> > large ia64 boxes would use this, and there aren't very many of those.
> >
>
> Ok. I will add a check for _TTP before setting sparse. Something likes this.
>
> add_io_space()
> {
> ...
> if (addr->info.io.translation == ACPI_TYPE_TRANSLATION &&
> addr->info.io.translation_type == ACPI_SPARSE_TRANSLATION)
> sparse = 1;
> ...
> }
>
>
>
> > Bjorn
> >
>
>
> --
> Best regards
> Tianyu Lan