On 28.07.23 18:18, Linus Torvalds wrote:
On Thu, 27 Jul 2023 at 14:28, David Hildenbrand <david@xxxxxxxxxx> wrote:
This is my proposal on how to handle the fallout of 474098edac26
("mm/gup: replace FOLL_NUMA by gup_can_follow_protnone()") where I
accidentially missed that follow_page() and smaps implicitly kept the
FOLL_NUMA flag clear by *not* setting it if FOLL_FORCE is absent, to
not trigger faults on PROT_NONE-mapped PTEs.
Ugh.
I was hoping for that reaction, with the assumption that we would get
something cleaner :)
I hate how it uses FOLL_FORCE that is inherently scary.
I hate FOLL_FORCE, but I hate FOLL_NUMA even more, because to me it
is FOLL_FORCE in disguise (currently and before 474098edac26, if
FOLL_FORCE is set, FOLL_NUMA won't be set and the other way around).
Why do we have that "gup_can_follow_protnone()" logic AT ALL?
That's what I was hoping for.
Couldn't we just get rid of that disgusting thing, and just say that
GUP (and follow_page()) always just ignores NUMA hinting, and always
just follows protnone?
We literally used to have this:
if (!(gup_flags & FOLL_FORCE))
gup_flags |= FOLL_NUMA;
ie we *always* set FOLL_NUMA for any sane situation. FOLL_FORCE should
be the rare crazy case.
Yes, but my point would be that we now spell that "rare crazy case"
out for follow_page().
If you're talking about patch #1, I agree, therefore patch #3 to
avoid all that nasty FOLL_FORCE handling in GUP callers.
But yeah, if we can avoid all that, great.
The original reason for not setting FOLL_NUMA all the time is
documented in commit 0b9d705297b2 ("mm: numa: Support NUMA hinting
page faults from gup/gup_fast") from way back in 2012:
* If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
* would be called on PROT_NONE ranges. We must never invoke
* handle_mm_fault on PROT_NONE ranges or the NUMA hinting
* page faults would unprotect the PROT_NONE ranges if
* _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
* bitflag. So to avoid that, don't set FOLL_NUMA if
* FOLL_FORCE is set.
In handle_mm_fault(), we never call do_numa_page() if
!vma_is_accessible(). Same for do_huge_pmd_numa_page().
So, if we would ever end up triggering a page fault on
mprotect(PROT_NONE) ranges (i.e., via FOLL_FORCE), we
would simply do nothing.
At least that's the hope, I'll take a closer look just to make
sure we're good on all call paths.
but I don't think the original reason for this is *true* any more.
Because then two years later in 2014, in commit c46a7c817e66 ("x86:
define _PAGE_NUMA by reusing software bits on the PMD and PTE levels")
Mel made the code able to distinguish between PROT_NONE and NUMA
pages, and he changed the comment above too.
CCing Mel.
I remember that pte_protnone() can only distinguished between
NUMA vs. actual mprotect(PROT_NONE) by looking at the VMA -- vma_is_accessible().
Indeed, include/linux/pgtable.h:
/*
* Technically a PTE can be PROTNONE even when not doing NUMA balancing but
* the only case the kernel cares is for NUMA balancing and is only ever set
* when the VMA is accessible. For PROT_NONE VMAs, the PTEs are not marked
* _PAGE_PROTNONE so by default, implement the helper as "always no". It
* is the responsibility of the caller to distinguish between PROT_NONE
* protections and NUMA hinting fault protections.
*/
But I get the very very strong feeling that instead of changing the
comment, he should have actually removed the comment and the code.
So I get the strong feeling that all these FOLL_NUMA games should just
go away. You removed the FOLL_NUMA bit, but you replaced it with using
FOLL_FORCE.
So rather than make this all even more opaque and make it even harder
to figure out why we have that code in the first place, I think it
should all just be removed.
At least to me, spelling FOLL_FORCE in follow_page() now out is much
less opaque then getting that implied by lack of FOLL_NUMA.
The original reason for FOLL_NUMA simply does not exist any more. We
know exactly when a page is marked for NUMA faulting, and we should
simply *ignore* it for GUP and follow_page().
I think we should treat a NUMA-faulting page as just being present
(and not NUMA-fault it).
Am I missing something?
There was the case for "FOLL_PIN represents application behavior and
should trigger NUMA faults", but I guess that can be ignored.
