On Mon, 31 Jul 2023 at 09:20, David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> I modified it slightly: FOLL_HONOR_NUMA_FAULT is now set in
> is_valid_gup_args(), such that it will always be set for any GUP users,
> including GUP-fast.
But do we actually want that? It is actively crazy to honor NUMA
faulting at least for get_user_pages_remote().
So right now, GUP-fast requires us to honor NUMA faults, because
GUP-fast doesn't have a vma (which in turn is because GUP-fast doesn't
take any locks).
So GUP-fast can only look at the page table data, and as such *has* to
fail if the page table is inaccessible.
But GUP in general? Why would it want to honor numa faulting?
Particularly by default, and _particularly_ for things like
FOLL_REMOTE.
In fact, I feel like this is what the real rule should be: we simply
define that get_user_pages_fast() is about looking up the page in the
page tables.
So if you want something that acts like a page table lookup, you use
that "fast" thing. It's literally how it is designed. The whole - and
pretty much only - point of it is that it can be used with no locking
at all, because it basically acts like the hardware lookup does.
So then if KVM wants to look up a page in the page table, that is what
kvm should use, and it automatically gets the "honor numa faults"
behavior, not because it sets a magic flag, but simply because that is
how GUP-fast *works*.
But if you use the "normal" get/pin_user_pages() function, which looks
up the vma, at that point you are following things at a "software
level", and it wouldn't do NUMA faulting, it would just get the page.
(Ok, we have the whole "FAST_ONLY vs fall back" case, so "fast" can
look up the vma too, but read the above argument as "fast *can* be
done without vma, so fast must honor page table bits as per
hardware").
Linus
