Re: About commit "io: change inX() to have their own IO barrier overrides"

[John Garry <john.garry@huawei.com>]

On 03/03/2020 16:40, Arnd Bergmann wrote:
> On Tue, Mar 3, 2020 at 2:18 PM John Garry <john.garry@huawei.com> wrote:
>>
>> + linux-arch
>>
>> For background, see
>> https://lore.kernel.org/lkml/2e80d7bc-32a0-cc40-00a9-8a383a1966c2@huawei.com/
>>
>>>>
>>>> So today only ARM64 uses it for this relevant code, above. But maybe
>>>> others in future will want to use it - any arch without native IO port
>>>> access is a candidate.
>>>
>>> I'm looking at Arnd here for help.
>>>
>>>>
>>>>>
>>>>> As long as the expectations are set, I see no reason why it shouldn't
>>>>> but, I'll let Arnd comment on it too.
>>>>
>>>> ok, so it looks reasonable consider replicating your change for ***, above.
>>
>> To be clear, I would make this change in lib/logic_pio.c since
>> __io_pbr() can be overridden per-arch:
>>
>>    #define BUILD_LOGIC_IO(bw, type)
>>    type logic_in##bw(unsigned long addr)
>>    {
>>         type ret = (type)~0;
>>         if (addr < MMIO_UPPER_LIMIT) {
>> -          ret = read##bw(PCI_IOBASE + addr);
>> +          __io_pbr();
>> +          ret = __raw_read##bw(PCI_IOBASE + addr);
>> +          __io_pbr();
> 
> __io_par();
> 
>>         } else if (addr >= MMIO_UPPER_LIMIT && addr < IO_SPACE_LIMIT) {
>>             struct logic_pio_hwaddr *entry = find_io_range(addr);
>>
>> ...
>>
>> (forgetting leX_to_cpu for the moment)
> 
> Yes, I suppose this is required to get consistent behavior on arm64,
> which overrides __io_par() but not __io_ar(), with the current code
> the barrier after read is weaker when LOGIC_PIO is enabled than it
> is otherwise.

Ok.

Apart from that, this code is somewhat hidden. I mean, most people would 
consider generic IO port accessors come from asm-generic/io.h only, 
which is not the case here. Maybe this can be better integrated into 
asm-generic/io.h, the only hint today being the logic_pio.h include half 
way through the file.

> 
> For other architectures, I suppose we would need another indirection
> level, as those can also override the default inb() itself to do something
> other than readb(PCI_IOBASE + addr), and that is not handled
> here either. We can do that if we need LOGIC_PIO on a second
> architecture.

Jiaxun Yang did mention that MIPS may want to move away from its own IO 
space management.

Thanks,
John