Re: [PATCH for 4.20? v3 1/3] xen/riscv: implement software page table walking

[Jan Beulich <jbeulich@xxxxxxxx>]

On 12.02.2025 12:13, Oleksii Kurochko wrote:
> 
> On 2/10/25 5:32 PM, Jan Beulich wrote:
>> On 07.02.2025 14:13, Oleksii Kurochko wrote:
>>> --- a/xen/arch/riscv/pt.c
>>> +++ b/xen/arch/riscv/pt.c
>>> @@ -185,6 +185,57 @@ static int pt_next_level(bool alloc_tbl, pte_t 
>>> **table, unsigned int offset)
>>>       return XEN_TABLE_NORMAL;
>>>   }
>>>   
>>> +/*
>>> + * _pt_walk() performs software page table walking and returns the pte_t of
>>> + * a leaf node or the leaf-most not-present pte_t if no leaf node is found
>>> + * for further analysis.
>>> + * Additionally, _pt_walk() returns the level of the found pte.
>> That's optional, which I think wants expressing here.
>>
>>> + */
>>> +static pte_t *_pt_walk(vaddr_t va, unsigned int *pte_level)
>>> +{
>>> +    const mfn_t root = get_root_page();
>>> +    unsigned int level;
>>> +    pte_t *table;
>>> +
>>> +    DECLARE_OFFSETS(offsets, va);
>>> +
>>> +    table = map_table(root);
>> This mapping operation doesn't look to have a counterpart. Aiui ...
>>
>>> +    /*
>>> +     * Find `table` of an entry which corresponds to `va` by iterating for 
>>> each
>>> +     * page level and checking if the entry points to a next page table or
>>> +     * to a page.
>>> +     *
>>> +     * Two cases are possible:
>>> +     * - ret == XEN_TABLE_SUPER_PAGE means that the entry was found;
>>> +     *   (Despite the name) XEN_TABLE_SUPER_PAGE also covers 4K mappings. 
>>> If
>>> +     *   pt_next_level() is called for page table level 0, it results in 
>>> the
>>> +     *   entry being a pointer to a leaf node, thereby returning
>>> +     *   XEN_TABLE_SUPER_PAGE, despite of the fact this leaf covers 4k 
>>> mapping.
>>> +     * - ret == XEN_TABLE_MAP_NONE means that requested `va` wasn't 
>>> actually
>>> +     *   mapped.
>>> +     */
>>> +    for ( level = HYP_PT_ROOT_LEVEL; ; --level )
>>> +    {
>>> +        int ret = pt_next_level(false, &table, offsets[level]);
>> ... the mapping may be replaced here, but a new mapping will then still
>> be held by this function and ...
>>
>>> +        if ( ret == XEN_TABLE_MAP_NONE || ret == XEN_TABLE_SUPER_PAGE )
>>> +            break;
>>> +
>>> +        ASSERT(level);
>>> +    }
>>> +
>>> +    if ( pte_level )
>>> +        *pte_level = level;
>>> +
>>> +    return table + offsets[level];
>>> +}
>> ... the final one then be transferred to the caller.
>>
>>> +pte_t pt_walk(vaddr_t va, unsigned int *pte_level)
>>> +{
>>> +    return *_pt_walk(va, pte_level);
>>> +}
>> Hence aiui there needs to be an unmap operation here.
> 
> As _pt_walk() returns page table entry, it is needed to transform entry to 
> table.
> 
> Do we have any function in Xen for that?

I don't understand. Both unmap_domain_page() and pmap_unmap() ignore
the low bits of the passed in address.

Jan

> Or the best I can do is:
>    pte_t *table = *_pt_walk(va, pte_level) - 
> TABLE_OFFSET(pt_linear_offset(*pte_level, va)
> (of course, it is needed to check if pte_level isn't null and do some extra 
> actions if it is NULL)
> 
> As an option, as all page tables are PAGE_SIZE aligned, we could use 
> PAGE_OFFSET():
>   pte_t *entry = _pt_walk(va, pte_level);
>   pte_t *table = PAGE_OFFSET(entry);
> 
> ~ Oleksii
> 
>