Re: [Qemu-devel] VFIO use of HOST_PAGE_ALIGN

[Alexey Kardashevskiy <aik@ozlabs.ru>]

On 06/04/2015 03:53 AM, Peter Crosthwaite wrote:
> On Wed, Jun 3, 2015 at 4:16 AM, Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
>> On 06/01/2015 04:27 AM, Peter Crosthwaite wrote:
>>>
>>> On Sun, May 31, 2015 at 6:34 AM, Alexey Kardashevskiy <aik@ozlabs.ru>
>>> wrote:
>>>>
>>>> On 05/27/2015 01:22 AM, Alex Williamson wrote:
>>>>>
>>>>>
>>>>> [cc +alexey]
>>>>>
>>>>> On Mon, 2015-05-25 at 00:48 -0700, Peter Crosthwaite wrote:
>>>>>>
>>>>>>
>>>>>> Hi Alex and all,
>>>>>>
>>>>>> I am working on a patch series to enable multiple CPU architectures to
>>>>>> run at once. It's a long story, but I have hit a snag in hw/vfio/pci.c
>>>>>> which AFAICS is the only in-tree system-mode use of HOST_PAGE_ALIGN
>>>>>> (all usermode code is a non-issue as not looking to support user-mode
>>>>>> multi-arch yet).
>>>>>>
>>>>>> The problem I face, is that this macro depends on translate-all.c's
>>>>>> qemu_host_page_size which in turn, depends on TARGET_PAGE_SIZE. I'm
>>>>>> hoping that one day, TARGET_PAGE_SIZE will be a variable and the users
>>>>>> of it will know to get the correct value depending on their CPU
>>>>>> specific code location. vfio is the only one I can't handle. My
>>>>>> knowledge on vfio is near-0, but my thinking is, since this is not
>>>>>> arch specific code can we instead use the raw host page alignment
>>>>>> rather that the CPU arch specific one?
>>>>
>>>>
>>>>
>>>> What is "raw host page" here? I thought qemu_host_page_size is the one,
>>>> where does it depend on TARGET_PAGE_SIZE?
>>>>
>>>>
>>>
>>> In translate-all.c:
>>>
>>> void page_size_init(void)
>>> {
>>>       /* NOTE: we can always suppose that qemu_host_page_size >=
>>>          TARGET_PAGE_SIZE */
>>>       qemu_real_host_page_size = getpagesize();
>>>       if (qemu_host_page_size == 0) {
>>>           qemu_host_page_size = qemu_real_host_page_size;
>>>       }
>>>       if (qemu_host_page_size < TARGET_PAGE_SIZE) {
>>>           qemu_host_page_size = TARGET_PAGE_SIZE;
>>>       }
>>>       qemu_host_page_mask = ~(qemu_host_page_size - 1);
>>> }
>>>
>>> It is clamped to be at least as big as a the TARGET_PAGE_SIZE.
>>
>>
>>
>> Ah, just this one. TARGET_PAGE_SIZE is 4K for PPC64 and the actual page size
>> can be 4K and 64K so that branch never works for PPC64.
>>
>>
>>>>> I think we could replace our use of HOST_PAGE_ALIGN with something based
>>>>> only on the host's getpagesize().  I don't see that we really care about
>>>>> the target page size for this usage.  Alexey, I think you're the only
>>>>> arch where host and target page sizes can actually be different, do you
>>>>> agree?  Thanks,
>>>>
>>>>
>>>>
>>>> Strongly agree. Where it really matters (MSIX), it is already
>>>> qemu_host_page_size and HOST_PAGE_ALIGN and I am a bit scared by that
>>>> "raw
>>>> host page alignment" :)
>>>>
>>>
>>> So the change would be an alternate macro based on
>>> qemu_real_host_page_size which is what I am calling "raw host page
>>> alignment".
>>
>>
>> Good, I like it. Cannot it be just a qemu_host_page_size? It is a bit
>> confusing to have both qemu_host_page_size and qemu_real_host_page_size when
>> even the first name suggests it is something non-static and sort of real :)
>>
>
> Yes based on this, HOST_PAGE_ALIGN is badly named. It is really
> aligning to both host and target. Some rename options:
>
> QEMU_PAGE_ALIGN
> HOST_TARGET_PAGE_ALIGN
> QEMU_HOST_PAGE_ALIGN
> or just unqualified PAGE_ALIGN?


Sorry, I have bad taste so I am not the one to ask :) I'd stick to 
HOST_PAGE_ALIGN.




-- 
Alexey