Re: [PATCH v1] mm/vmalloc: fix exact allocations with an alignment > 1

[David Hildenbrand <david@redhat.com>]

>> No, that's leaking implementation details to the caller. And no, increasing
>> the range and eventually allocating something bigger (e.g., placing a huge
>> page where it might not have been possible) is not acceptable for KASAN.
>>
>> If you're terribly unhappy with this patch,
> Sorry to say but it simple does not make sense.
> 

Let's agree to disagree.

find_vmap_lowest_match() is imprecise now and that's an issue for exact 
allocations. We can either make it fully precise again (eventually 
degrading allocation performance) or just special-case exact allocations 
to fix the regression.

I decided to go the easy path and do the latter; I do agree that making 
find_vmap_lowest_match() fully precise again might be preferred -- we 
could have other allocations failing right now although there are still 
suitable holes.

I briefly thought about performing the search in 
find_vmap_lowest_match() twice. First, start the search without an 
extended range, and fallback to the extended range if that search fails. 
Unfortunately, I think that still won't make the function completely 
precise due to the way we might miss searching some suitable subtrees.

>>
>> please suggest something reasonable to fix exact allocations:
>> a) Fixes the KASAN use case.
>> b) Allows for automatic placement of huge pages for exact allocations.
>> c) Doesn't leak implementation details into the caller.
>>
> I am looking at it.

Thanks!

-- 
Thanks,

David / dhildenb