On Wed, 1 Oct 2014, Linus Torvalds wrote:
> On Wed, Oct 1, 2014 at 1:19 AM, Hugh Dickins <hu...@google.com> wrote:
>
> Can we please just get rid of _PAGE_NUMA. There is no excuse for it.
I'm no lover of _PAGE_NUMA, and hope that it can be simplified away
as you outline. What we have in 3.16+3.17 is already an attempt to
improve on what you hated before, but not obviously an improvement.
Mel is the one who knows these issues inside out: maybe he's been
blinkered, but I wouldn't dare to pull it apart without his input.
Myself, I'm not looking beyond fixing whatever is the bug for 3.17.
> > However, that would still not explain Dave's endless refaulting;
>
> Why not? You start out with a PROTNONE, trigger shrink_page_list() on
> a hugepage,.which calls add_to_swap(), which does
> split_huge_page_to_list(), which in turn calls __split_huge_page(),
> and that turns (_PAGE_PROTNONE) into (_PAGE_PROTNONE|_PAGE_NUMA),
> which you will then fault on forever, because the kernel thinks the
> page is present, but not a NUMA page.
I hesitate to admit, I still don't see it: please illuminate further.
We're talking about the loop in __split_huge_page_map(), where it does
entry = mk_pte(page + i, vma->vm_page_prot);
entry = maybe_mkwrite(pte_mkdirty(entry), vma);
if (!pmd_write(*pmd))
entry = pte_wrprotect(entry);
if (!pmd_young(*pmd))
entry = pte_mkold(entry);
if (pmd_numa(*pmd))
entry = pte_mknuma(entry);
, right? I only see that adding _PAGE_NUMA to _PAGE_PROTNONE if
pmd_numa(*pmd): but that would mean we had already gone wrong, setting
pmd_numa in a PROT_NONE vma, which task_numa_work takes care not to do;
or have mprotected an area to PROT_NONE without doing the pmd_mknonnuma.
Ah, we won't have mmap_sem in the add_to_swap case; so we could be
racing with an mprotect which already updated vm_flags and vm_page_prot,
but has not yet reached this pmd: is that a part of how you see it?
Or are you noticing a deficiency in the pmd locking? I have not
worked my way through that, so cannot guarantee it, but please
point me to the weakness where you see it.
But when you convince me on that, then I still don't see how we get to
doing the repeated write fault, instead of hitting access_error(), as
you pointed out originally. That still seems to require a PROTNONE
pte to be left behind in a VM_WRITE vma, which I do not see happening
here. pte_modify leaves _PAGE_NUMA alone, but updates _PAGE_PROTNONE.
The pte_numa<->pte_special confusion is messy, but I don't yet get
how it would manifest in the manner observed.
But certainly, a bug in the THP splitting feels just right to
match the frequency of the sightings: I hope you've got it.
Hugh
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
