Re: [RFC PATCH 1/2] pci: don't assume pref memio are 64bit -v3

[Yinghai Lu <yinghai@kernel.org>]

Ivan Kokshaysky wrote:
> On Wed, Apr 22, 2009 at 03:37:04PM -0700, Yinghai Lu wrote:
>> one system with 4g installed ( there is 1g hole)
>>
>> when 4G installed.
>> BIOS put ACPI etc need the hole
>> [    0.000000] BIOS-provided physical RAM map:
>> [    0.000000]  BIOS-e820: 0000000000000000 - 000000000009bc00 (usable)
>> [    0.000000]  BIOS-e820: 000000000009bc00 - 00000000000a0000 (reserved)
>> [    0.000000]  BIOS-e820: 00000000000e3000 - 0000000000100000 (reserved)
>> [    0.000000]  BIOS-e820: 0000000000100000 - 00000000bffa0000 (usable)
>> [    0.000000]  BIOS-e820: 00000000bffa0000 - 00000000bffae000 (ACPI data)
>> [    0.000000]  BIOS-e820: 00000000bffae000 - 00000000bfff0000 (ACPI NVS)
>> [    0.000000]  BIOS-e820: 00000000bfff0000 - 00000000c0000000 (reserved)
>> [    0.000000]  BIOS-e820: 00000000fee00000 - 00000000fee01000 (reserved)
>> [    0.000000]  BIOS-e820: 00000000ffb00000 - 0000000100000000 (reserved)
>> [    0.000000]  BIOS-e820: 0000000100000000 - 0000000140000000 (usable)
>> so in kernel resource will be reserved for 0xbffa0000 - 0xbfff0000 for ACPI
>> 0x100000 -  0xbffa0000 for RAM...
>>
>> and BIOS set
>> [    0.240007] pci 0000:00:01.0: bridge 64bit mmio pref: [0xbdf00000-0xddefffff]
>> [    0.237102] pci 0000:01:00.0: reg 10 32bit mmio: [0xc0000000-0xcfffffff]
>> that is conflict with reserved res. so it can not be reserved Kernel.
>>
>> then Kernel try to get range from 0x140000000 ( above the RAM, 5G and above 4g)
>> and set let the bridge to use it, and ATI cards to use it.
>>
>> but the problem is that ATI only support 32bit ...
> 
> Yinghai, you are trying to get a quick fix for quite a complex problem
> that cannot be solved with a quick fix. Even more, there is no rush on
> a quick fix because it's not a critical bug at all. 32-bit stuff ends
> up above 4G *only* when there is no free space left on the 32-bit
> PCI bus, and it can be considered as very effective (though rather ugly)
> way of disabling the BAR that we failed to allocate.
> In this particular case it was simply a side effect of the "pci_mem_start"
> issue (which was indeed critical, but hopefully fixed now).
> 

i agreed that that is not crital bug at all. pci_mem_start patch should fix that allocation alone.
actually "pci: don't assume pref memio are 64bit" just make kernel give customer surprise.

> 
>> +/* need to insert those two under iomem */
>> +struct resource iomem32_resource = {
>> +	.name	= "PCI mem 32bit",
>> +	.start	= 0,
>> +	.end	= 0xffffffff,
>> +	.flags	= IORESOURCE_MEM,
>> +};
>> +struct resource iomem64_resource = {
>> +	.name	= "PCI mem 64bit",
>> +	.start	= 1ULL<<32,
>> +	.end	= -1,
>> +	.flags	= IORESOURCE_MEM | IORESOURCE_MEM_64,
>> +};
>> +
> 
> This only works on x86 and similar systems with 1:1 CPU address to bus
> address mapping. There is a lot of machines with multiple 32-bit PCI
> bus spaces (4G per PCI domain).

need to move that code to arch code for x86?

YH