[RFC PATCH v3 1/6] arch: introduce set_direct_map_valid_noflush()

[Patrick Roy]

On Mon, 2024-11-11 at 12:12 +0000, Vlastimil Babka wrote:
> On 10/31/24 10:57, David Hildenbrand wrote:
>> On 30.10.24 14:49, Patrick Roy wrote:
>>> From: "Mike Rapoport (Microsoft)" <rppt at kernel.org>
>>>
>>> From: Mike Rapoport (Microsoft) <rppt at kernel.org>
>>>
>>> Add an API that will allow updates of the direct/linear map for a set of
>>> physically contiguous pages.
>>>
>>> It will be used in the following patches.
>>>
>>> Signed-off-by: Mike Rapoport (Microsoft) <rppt at kernel.org>
>>> Signed-off-by: Patrick Roy <roypat at amazon.co.uk>
>>
>>
>> [...]
>>
>>>   #ifdef CONFIG_DEBUG_PAGEALLOC
>>>   void __kernel_map_pages(struct page *page, int numpages, int enable)
>>>   {
>>> diff --git a/include/linux/set_memory.h b/include/linux/set_memory.h
>>> index e7aec20fb44f1..3030d9245f5ac 100644
>>> --- a/include/linux/set_memory.h
>>> +++ b/include/linux/set_memory.h
>>> @@ -34,6 +34,12 @@ static inline int set_direct_map_default_noflush(struct page *page)
>>>      return 0;
>>>   }
>>>
>>> +static inline int set_direct_map_valid_noflush(struct page *page,
>>> +                                           unsigned nr, bool valid)
>>
>> I recall that "unsigned" is frowned upon; "unsigned int".
>>
>>> +{
>>> +    return 0;
>>> +}
>>
>> Can we add some kernel doc for this?
>>
>> In particular
>>
>> (a) What does it mean when we return 0? That it worked? Then, this
> 
> Seems so.
> 
>>      dummy function looks wrong. Or this it return the
> 
> That's !CONFIG_ARCH_HAS_SET_DIRECT_MAP and other functions around do it the
> same way. Looks like the current callers can only exist with the CONFIG_
> enabled in the first place.

Yeah, it looks a bit weird, but these functions seem to generally return
0 if the operation is not supported. ARM specifically has 

	if (!can_set_direct_map())
		return 0;

inside `set_direct_map_invalid_{noflush,default}`. Documenting this
definitely cannot hurt, I'll keep it on my todo list for the next
iteration :)

>>      number of processed entries? Then we'd have a possible "int" vs.
>>      "unsigned int" inconsistency.
>>
>> (b) What are the semantics when we fail halfway through the operation
>>      when processing nr > 1? Is it "all or nothing"?
> 
> Looking at x86 implementation it seems like it can just bail out in the
> middle, but then I'm not sure if it can really fail in the middle, hmm...

If I understood Mike correctly when talking about this at LPC, then it
can only fail if during break-up of huge mappings, it fails to allocate
page tables to hold the lower-granularity mappings (which happens before
any present bits are modified).

Best,
Patrick