Re: [PATCH v1 1/3] xen/riscv: implement software page table walking

[Jan Beulich <jbeulich@xxxxxxxx>]

On 29.01.2025 14:12, Oleksii Kurochko wrote:
> 
> On 1/28/25 9:14 AM, Jan Beulich wrote:
>> On 27.01.2025 18:22, Oleksii Kurochko wrote:
>>> On 1/27/25 1:57 PM, Jan Beulich wrote:
>>>> On 27.01.2025 13:29, Oleksii Kurochko wrote:
>>>>> On 1/27/25 11:06 AM, Jan Beulich wrote:
>>>>>> On 20.01.2025 17:54, Oleksii Kurochko wrote:
>>>>>>> RISC-V doesn't have hardware feature to ask MMU to translate
>>>>>>> virtual address to physical address ( like Arm has, for example ),
>>>>>>> so software page table walking in implemented.
>>>>>>>
>>>>>>> Signed-off-by: Oleksii Kurochko<oleksii.kurochko@xxxxxxxxx>
>>>>>>> ---
>>>>>>>     xen/arch/riscv/include/asm/mm.h |  2 ++
>>>>>>>     xen/arch/riscv/pt.c             | 56 
>>>>>>> +++++++++++++++++++++++++++++++++
>>>>>>>     2 files changed, 58 insertions(+)
>>>>>>>
>>>>>>> diff --git a/xen/arch/riscv/include/asm/mm.h 
>>>>>>> b/xen/arch/riscv/include/asm/mm.h
>>>>>>> index 292aa48fc1..d46018c132 100644
>>>>>>> --- a/xen/arch/riscv/include/asm/mm.h
>>>>>>> +++ b/xen/arch/riscv/include/asm/mm.h
>>>>>>> @@ -15,6 +15,8 @@
>>>>>>>     
>>>>>>>     extern vaddr_t directmap_virt_start;
>>>>>>>     
>>>>>>> +paddr_t pt_walk(vaddr_t va);
>>>>>> In the longer run, is returning just the PA really going to be 
>>>>>> sufficient?
>>>>>> If not, perhaps say a word on the limitation in the description.
>>>>> In the long run, this function's prototype looks like|paddr_t 
>>>>> pt_walk(vaddr_t root, vaddr_t va, bool is_xen)| [1]. However, I'm not 
>>>>> sure if it will stay that way,
>>>>> as I think|is_xen| could be skipped, since using|map_table()| should be 
>>>>> sufficient (as it now considers|system_state|) and I'm not really sure if 
>>>>> I need root argument
>>>>> as initial goal was to use this function for debug only purposes and I've 
>>>>> never used it for guest page table (stage-1) walking.
>>>>> Anyway, yes, it is still returning a physical address, and that seems 
>>>>> enough to me.
>>>>>
>>>>> Could you share your thoughts on what I should take into account for 
>>>>> returning value, probably, I am missing something really useful?
>>>> Often you care about the permissions as well. Sometimes it may even be 
>>>> relevant
>>>> to know the (super-)page size of the mapping.
>>> Perhaps it would be better to change the prototype to:
>>>     bool pt_walk(vaddr_t va, mfn_t *ret_pa);
>>> or even
>>>     void pt_walk(vaddr_t va, mfn_t *ret_pa);
>>>     In this case,|ret_pa = INVALID_MFN| could serve as a signal 
>>> that|pt_walk()| failed.
>>> If there's a need to return permissions or (super-)page size in the future, 
>>> another argument could be added.
>>>
>>> What do you think? Would this approach be better?
>>>
>>> I am also considering returning a structure containing the|mfn| 
>>> (or|paddr_t|) and adding other properties (such as permissions or
>>> page size) as needed in the future. Both solutions seem more or less 
>>> equivalent.
>> Imo the most natural thing for a page walking function would be to return the
>> leaf PTE (or the leaf-most not-present [or otherwise "no-access"] one). That
>> would provide (almost) all possible information to the caller. "Almost"
>> because depending on how page walk works, permissions may combine across page
>> table levels. Yet then (see also the "no-access" above) this would also
>> require further input, to specify the context for which the translation is
>> being seeked. For example, the intention to write may want to yield no valid
>> PTE when there are present ones down to the leaf, but effective permissions
>> say "read-only".
> 
> Perhaps returning the leaf PTE could be a really good option.
> 
> I'm not entirely sure I understand what you mean by "leaf-most not-present". 
> Could you please try to explain this moment one more time?
> My expectation was that the function should return an existing leaf PTE (from 
> which "access" rights could be determined)
> or|NULL| to indicate that no leaf PTE was found.

"no leaf PTE" may be for a variety of reasons. Hence why I think returning
the PTE at which the walk stopped (leaf or leaf-most not-present) is likely
best. Such a not-present PTE may, after all, still contain valuable
information; it's not like it has to be all zero.

> Another thing I'm curious about is whether this would be sufficient for 
> determining the level.
> It seems clear that, given a PTE and a virtual address, we could compute:
> |mask = VA | paddr_from_pte(pte)|

What would this value represent? No, from holding a PTE in your hands you
can't determine the level it came from. So yes, ...

> Then, iterating through each level, we could apply and understand on which 
> one level it was mapped:
> |mask & (BIT(XEN_PT_LEVEL_ORDER(i), UL) - 1)|.
> 
> If I haven't overlooked any other way to calculate the page table level, 
> would it be better to simply add another argument
> to|pt_walk()| to return the level.

... for callers who care doing this might then be necessary (this would be
a pointer parameter, and since I expect many callers wouldn't care about
the level, it likely wants to be permissible to pass in NULL).

Question then is whether it's better to hand back the level or the page
order of the mapping. On x86 we return the latter from P2M lookups, for
example.

Jan