Re: [PATCH v1 3/3] ARM64 LPC: update binding doc

[Rongrong Zou <zourongrong@huawei.com>]

在 2016/1/5 20:19, Arnd Bergmann 写道:
> On Tuesday 05 January 2016 19:59:49 Rongrong Zou wrote:
>> 在 2016/1/5 0:34, Arnd Bergmann 写道:
>>> On Tuesday 05 January 2016 00:04:19 Rongrong Zou wrote:
>>>> 在 2016/1/4 19:13, Arnd Bergmann 写道:
>>>>> On Sunday 03 January 2016 20:24:14 Rongrong Zou wrote:
>>>>>> 在 2015/12/31 23:00, Rongrong Zou 写道:
>>>> Ranges property can set empty, but this means 1:1 translation. the I/O
>>>> port range is translated to MMIO address 0x00000001 00000000 to
>>>> 0x00000001 00000004, it looks wrong else. I wonder if anyone get legacy
>>>> I/O port resource from dts.
>>>
>>> As I said, nothing should really require the ranges property here, unless
>>> you have a valid IORESOURCE_MEM translation. The code that requires
>>> the ranges to be present is wrong.
>>>
>>
>> I think the openfirmware(DT) do not support for those unmapped I/O ports, because I
>> must get resource by calling of_address_to_resource(), which have to call
>> pci_address_to_pio() when resource type is IORESOURCE_IO. I'm sorry I have no
>> better idea for this now. Maybe liviu can give me some opinions.
>
> I think on x86 it works (or used to work, few people use open firmware on
> x86 these days, and it may be broken), and the pci_address_to_pio() call
> behaves differently when PCI_IOBASE is set. x86 never maps I/O ports into
> memory mapped I/O addresses, they have their own way of accessing them
> just like your platform.
>
>> /**
>>    * of_address_to_resource - Translate device tree address and return as resource
>>    *
>>    * Note that if your address is a PIO address, the conversion will fail if
>>    * the physical address can't be internally converted to an IO token with
>>    * pci_address_to_pio(), that is because it's either called to early or it
>>    * can't be matched to any host bridge IO space
>>    */
>> int of_address_to_resource(struct device_node *dev, int index,
>> 			   struct resource *r)
>
> The problem here seems to be that the code assumes that either the I/O ports
> are always mapped or they are never mapped (no PCI_IOBASE). We need to extend
> it because now we can have the combination of the two.


I am considering the following solution:

Adding unmapped isa io functions in

drivers/of/address.c,

static LIST_HEAD(legacy_io_range_list);
int isa_register_io_range(phys_addr_t addr, resource_size_t size);

/* before I call isa(LPC) bus driver, the input I/O port must be translated to phys_addr_t
(the least 16bit means port addr on bus, the second 16bit means bus id)*/

phys_addr_t isa_pio_to_bus_addr(unsigned long pio);

/* the returned PIO do not conflict with PIO get from pci_address_to_pio*/
unsigned long isa_bus_addr_to_pio(phys_addr_t address);

drivers/bus/lpc.c

lpc_bus_probe()
{
	isa_register_io_range(phys_addr_t addr, resource_size_t size);
}

inb(unsigned long port)
{
	unsigned short bus;
	phys_addr_t addr;
	/*hit isa port range*/
	if(addr = isa_pio_to_bus_addr(port))
	{
		bus = (addr >> 16) & 0xffff;

		call lpc driver with addr;
		return lpc_read_byte(bus, addr);
	}
	else /*not hit*/
	{
		return readb(PCI_IOBASE + port);
	}
}



>
>>>> For ipmi driver, I can get I/O port resource by DMI rather than dts.
>>>
>>> No, the ipmi driver uses the resource that belongs to the platform
>>> device already, you can't rely on DMI data to be present there.
>>
>> Ipmi has a lot of way to be discovered(ACPI, DMI, hardcoded, hot-add,
>> openfirmware and a few other), I think we just use one of them, not all of them.
>> It depend on vendor's hardware solution actually.
>
> I don't think we should mix multiple methods here: if the bus is described
> in DT, all its children should be there as well. Otherwise you get into problems
> e.g. if you have multiple instances of the LPC bus and the Linux I/O addresses
> for one or more of them have an offset to the bus specific addresses.
>
> The bus probe code decides what the Linux I/O port numbers are, but DMI
> and other methods have no idea of the mapping. As long as there is only
> one instance, using the first 0x1000 addresses with a 1:1 mapping saves
> us a bit of trouble, but I'd be worried about relying on that assumption
> too much.
>
> 	Arnd
>
>
> .
>

Thanks,

Rongrong