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

From: Bjorn Helgaas
Date: Thu Oct 17 2013 - 16:34:14 EST


[+cc Mike]

On Thu, Oct 17, 2013 at 12:09 AM, Lan Tianyu <tianyu.lan@xxxxxxxxx> wrote:
> On 2013å10æ17æ 07:55, Bjorn Helgaas wrote:
>> On Fri, Oct 11, 2013 at 08:19:01PM +0800, tianyu.lan@xxxxxxxxx wrote:
>>> From: Lan Tianyu <tianyu.lan@xxxxxxxxx>
>>>
>>> Using ACPI resource functions to convert ACPI resource to generic resource
>>>
>>> Signed-off-by: Lan Tianyu <tianyu.lan@xxxxxxxxx>
>>> ---
>>> This patch just passes through compilation test due to no ia64 machine on hand.
>>>
>>> arch/ia64/pci/pci.c | 38 +++++++++++++++++++++-----------------
>>> 1 file changed, 21 insertions(+), 17 deletions(-)
>>>
>>> diff --git a/arch/ia64/pci/pci.c b/arch/ia64/pci/pci.c
>>> index 2326790..14fa175 100644
>>> --- a/arch/ia64/pci/pci.c
>>> +++ b/arch/ia64/pci/pci.c
>>> @@ -232,8 +232,9 @@ out:
>>> return ~0;
>>> }
>>>
>>> -static acpi_status resource_to_window(struct acpi_resource *resource,
>>> - struct acpi_resource_address64 *addr)
>>> +static acpi_status resource_to_window(struct acpi_resource *ares,
>>> + struct acpi_resource_address64 *addr,
>>> + struct resource *res)
>>> {
>>> acpi_status status;
>>>
>>> @@ -244,7 +245,7 @@ static acpi_status resource_to_window(struct acpi_resource *resource,
>>> * - producers, i.e., the address space is routed downstream,
>>> * not consumed by the bridge itself
>>> */
>>> - status = acpi_resource_to_address64(resource, addr);
>>> + status = acpi_dev_resource_address_space_full(ares, addr, res);
>>> if (ACPI_SUCCESS(status) &&
>>> (addr->resource_type == ACPI_MEMORY_RANGE ||
>>> addr->resource_type == ACPI_IO_RANGE) &&
>>> @@ -255,51 +256,54 @@ static acpi_status resource_to_window(struct acpi_resource *resource,
>>> return AE_ERROR;
>>> }
>>>
>>> -static acpi_status count_window(struct acpi_resource *resource, void *data)
>>> +static acpi_status count_window(struct acpi_resource *ares, void *data)
>>> {
>>> unsigned int *windows = (unsigned int *) data;
>>> struct acpi_resource_address64 addr;
>>> + struct resource res;
>>> acpi_status status;
>>>
>>> - status = resource_to_window(resource, &addr);
>>> + status = resource_to_window(ares, &addr, &res);
>>> if (ACPI_SUCCESS(status))
>>> (*windows)++;
>>>
>>> return AE_OK;
>>> }
>>>
>>> -static acpi_status add_window(struct acpi_resource *res, void *data)
>>> +static acpi_status add_window(struct acpi_resource *ares, void *data)
>>> {
>>> struct pci_root_info *info = data;
>>> - struct resource *resource;
>>> + struct resource *resource = &info->res[info->res_num];
>>> struct acpi_resource_address64 addr;
>>> acpi_status status;
>>> - unsigned long flags, offset = 0;
>>> + unsigned long offset = 0;
>>> struct resource *root;
>>>
>>> /* Return AE_OK for non-window resources to keep scanning for more */
>>> - status = resource_to_window(res, &addr);
>>> + status = resource_to_window(ares, &addr, resource);
>>> if (!ACPI_SUCCESS(status))
>>> return AE_OK;
>>>
>>> - if (addr.resource_type == ACPI_MEMORY_RANGE) {
>>> - flags = IORESOURCE_MEM;
>>> + if (resource->flags & IORESOURCE_MEM) {
>>> root = &iomem_resource;
>>> offset = addr.translation_offset;
>>> - } else if (addr.resource_type == ACPI_IO_RANGE) {
>>> - flags = IORESOURCE_IO;
>>> + } else if (resource->flags & IORESOURCE_IO) {
>>> root = &ioport_resource;
>>> offset = add_io_space(info, &addr);
>>> if (offset == ~0)
>>> return AE_OK;
>>> +
>>> + /*
>>> + * io space address translation offset depends
>>> + * on the return value of add_io_space(). So
>>> + * Repopulate resource->start and end here.
>>
>> "Repopulate" makes it sound like "resource->start" got clobbered
>> somewhere. But it didn't. I think it's just that each bridge can
>> support its own I/O port range, and the PCI port numbers reported
>> in the acpi_resource may not be distinct, and add_io_space() adds
>> an offset so all the I/O port ranges fit into one global I/O port
>> space.
>>
>> For example, I think these two bridges have the same port numbers
>> (0x0-0xfff) in their acpi_resource, but the second has an offset
>> of 0x1000000 in the system I/O port space, and I think this offset
>> is what add_io_space() returns:
>>
>> HWP0002:00: host bridge window [io 0x0000-0x0fff] (PCI [0x0-0xfff])
>> HWP0002:09: host bridge window [io 0x1000000-0x1000fff] (PCI [0x0-0xfff])
>>
>>> + */
>>> + resource->start = addr.minimum + offset;
>>> + resource->end = resource->start + addr.address_length - 1;
>>
>> Can't we use this:
>>
>> resource->start += offset;
>> resource->end += offset;
>>
>> to avoid breaking the encapsulation of struct acpi_resource_address64?
>
> resource->start has been populated with "addr.minimum +
> addr.translation_offset" in the acpi_dev_resource_address_space().

That's true, but this is a change from previous behavior. Previously,
x86 applied addr.translation_offset to both MEM and IO resources (in
setup_resource()), but ia64 applied it only to MEM resources (in
add_window()). With your patch, we apply it to both types in
acpi_dev_resource_address_space(), which is a change for ia64.

I know translation_offset is used on some HP ia64 boxes, but I'm not
aware of it being used for IO resources on any x86 boxes. On those
ia64 boxes, the bridge also does type translation (the resource is MEM
on the primary side but IO on the secondary side). In that case, I'm
not sure it makes sense to add the translation_offset to an IO address
and expect the result to be a MEM address.

On these HP ia64 boxes, the firmware puts the CPU physical address of
the MEM resource in the translation_offset (see the call to
new_space()). The bridge then translates that MEM resource to IO on
the secondary side. It's awfully hard for me to extract this usage
from the ACPI spec, so possibly this is just a quirk of the way HP
encoded these IO resources. But it *is* a precedent, and I'm not
aware of anybody doing anything that conflicts with it.

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 think the main intent of translation_offset (_TRA) is to map a
smaller address space into part of a larger space of the same type,
e.g., a 32-bit PCI memory space into a 40+ bit CPU memory space. That
doesn't apply directly to IO ports, because I don't think any CPU has
a native IO port address space larger than 16 bits, so there's no
extra space to map into.

Mike, is there any chance you could collect an acpidump from an rx7620
or similar ia64 system? In particular, I want to see a multi-node
system where we have several PCI domains, and whether it sets the _TTP
bits.

> continuing to add the offset to resource->start seems not right.
>
> The add_io_space() accepts translation_offset and then ioremap it to
> mmio address. Add the result to io_space array and assign a space
> number. Left shift the space number 24 bits as the return offset of
> add_io_space().
>
> When one io port address is accessed, __ia64_mk_io_addr() will do
> reverse operations and find associated mmio address.

Yep, got it. I wrote all that code originally :) Obviously it hasn't
turned out to be particularly easy to understand.

Bjorn
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/