Re: [RFC PATCH v2 01/20] mm: Provide pagesize to pmd_populate()

[Christophe Leroy]

Le 21/05/2024 à 13:57, Oscar Salvador a écrit :
> On Mon, May 20, 2024 at 04:24:51PM +0000, Christophe Leroy wrote:
>> I had a quick look at that document and it seems to provide a good
>> summary of MMU features and principles. However there are some
>> theoritical information which is not fully right in practice. For
>> instance when they say "Segment attributes. These fields define
>> attributes common to all pages in this segment.". This is right in
>> theory if you consider it from Linux page table topology point of view,
>> hence what they call a segment is a PMD entry for Linux. However, in
>> practice each page has its own L1 and L2 attributes and there is not
>> requirement at HW level to have all L1 attributes of all pages of a
>> segment the same.
> 
> Thanks for taking the time Christophe, highly appreciated.
> 
>   
>> rlwimi = Rotate Left Word Immediate then Mask Insert. Here it rotates
>> r10 by 23 bits to the left (or 9 to the right) then masks with
>> _PMD_PAGE_512K and inserts it into r11.
>>
>> It means _PAGE_HUGE bit is copied into lower bit of PS attribute.
>>
>> PS takes the following values:
>>
>> PS = 00 ==> Small page (4k or 16k)
>> PS = 01 ==> 512k page
>> PS = 10 ==> Undefined
>> PS = 11 ==> 8M page
> 
> I see, thanks for the explanation.
> 
>> That's a RFC, all ideas are welcome, I needed something to replace
>> hugepd_populate()
> 
> The only user interested in pmd_populate() having a sz parameter
> is 8xx because it will toggle _PMD_PAGE_8M in case of a 8MB mapping.
> 
> Would it be possible for 8xx to encode the 'sz' in the *pmd pointer
> prior to calling down the chain? (something like as we do for PTR_ERR()).
> Then pmd_populate_{kernel_}size() from 8xx, would extract it like:
> 
>   unsigned long sz = PTR_SIZE(pmd)
> 
> Then we would not need all these 'sz' parameters scattered.
> 
> Can that work?

Indeed _PMD_PAGE_8M can be set in set_huge_pte_at(), no need to do it 
atomically as part of pmd_populate, so I'll drop patches 1 and 2.

> 
> 
> PD: Do you know a way to emulate a 8xx VM? qemu seems to not have
> support support.
> 

I don't know any way. You are right that 8xx is not supported by QEMU 
unfortunately. I don't know how difficult it would be to add it to QEMU.

Christophe