Re: [PATCH V4 3/4] mm/sparse-vmemmap: Generalise vmemmap_populate_hugepages()

[Huacai Chen <chenhuacai@kernel.org>]

Hi, Will,

On Wed, Jul 20, 2022 at 5:34 PM David Hildenbrand <david@redhat.com> wrote:
>
> On 14.07.22 14:34, Huacai Chen wrote:
> > Oh, Sudarshan Rajagopalan's Email has changed, Let's update.
> >
> > Huacai
> >
> > On Fri, Jul 8, 2022 at 5:47 PM Huacai Chen <chenhuacai@kernel.org> wrote:
> >>
> >> +Dan Williams
> >> +Sudarshan Rajagopalan
> >>
> >> On Thu, Jul 7, 2022 at 12:17 AM Will Deacon <will@kernel.org> wrote:
> >>>
> >>> On Tue, Jul 05, 2022 at 09:07:59PM +0800, Huacai Chen wrote:
> >>>> On Tue, Jul 5, 2022 at 5:29 PM Will Deacon <will@kernel.org> wrote:
> >>>>> On Mon, Jul 04, 2022 at 07:25:25PM +0800, Huacai Chen wrote:
> >>>>>> diff --git a/mm/sparse-vmemmap.c b/mm/sparse-vmemmap.c
> >>>>>> index 33e2a1ceee72..6f2e40bb695d 100644
> >>>>>> --- a/mm/sparse-vmemmap.c
> >>>>>> +++ b/mm/sparse-vmemmap.c
> >>>>>> @@ -686,6 +686,60 @@ int __meminit vmemmap_populate_basepages(unsigned long start, unsigned long end,
> >>>>>>       return vmemmap_populate_range(start, end, node, altmap, NULL);
> >>>>>>  }
> >>>>>>
> >>>>>> +void __weak __meminit vmemmap_set_pmd(pmd_t *pmd, void *p, int node,
> >>>>>> +                                   unsigned long addr, unsigned long next)
> >>>>>> +{
> >>>>>> +}
> >>>>>> +
> >>>>>> +int __weak __meminit vmemmap_check_pmd(pmd_t *pmd, int node, unsigned long addr,
> >>>>>> +                                    unsigned long next)
> >>>>>> +{
> >>>>>> +     return 0;
> >>>>>> +}
> >>>>>> +
> >>>>>> +int __meminit vmemmap_populate_hugepages(unsigned long start, unsigned long end,
> >>>>>> +                                      int node, struct vmem_altmap *altmap)
> >>>>>> +{
> >>>>>> +     unsigned long addr;
> >>>>>> +     unsigned long next;
> >>>>>> +     pgd_t *pgd;
> >>>>>> +     p4d_t *p4d;
> >>>>>> +     pud_t *pud;
> >>>>>> +     pmd_t *pmd;
> >>>>>> +
> >>>>>> +     for (addr = start; addr < end; addr = next) {
> >>>>>> +             next = pmd_addr_end(addr, end);
> >>>>>> +
> >>>>>> +             pgd = vmemmap_pgd_populate(addr, node);
> >>>>>> +             if (!pgd)
> >>>>>> +                     return -ENOMEM;
> >>>>>> +
> >>>>>> +             p4d = vmemmap_p4d_populate(pgd, addr, node);
> >>>>>> +             if (!p4d)
> >>>>>> +                     return -ENOMEM;
> >>>>>> +
> >>>>>> +             pud = vmemmap_pud_populate(p4d, addr, node);
> >>>>>> +             if (!pud)
> >>>>>> +                     return -ENOMEM;
> >>>>>> +
> >>>>>> +             pmd = pmd_offset(pud, addr);
> >>>>>> +             if (pmd_none(READ_ONCE(*pmd))) {
> >>>>>> +                     void *p;
> >>>>>> +
> >>>>>> +                     p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
> >>>>>> +                     if (p) {
> >>>>>> +                             vmemmap_set_pmd(pmd, p, node, addr, next);
> >>>>>> +                             continue;
> >>>>>> +                     } else if (altmap)
> >>>>>> +                             return -ENOMEM; /* no fallback */
> >>>>>
> >>>>> Why do you return -ENOMEM if 'altmap' here? That seems to be different to
> >>>>> what we currently have on arm64 and it's not clear to me why we're happy
> >>>>> with an altmap for the pmd case, but not for the pte case.
> >>>> The generic version is the same as X86. It seems that ARM64 always
> >>>> fallback whether there is an altmap, but X86 only fallback in the no
> >>>> altmap case. I don't know the reason of X86, can Dan Williams give
> >>>> some explaination?
> >>>
> >>> Right, I think we need to understand the new behaviour here before we adopt
> >>> it on arm64.
> >> Hi, Dan,
> >> Could you please tell us the reason? Thanks.
> >>
> >> And Sudarshan,
> >> You are the author of adding a fallback mechanism to ARM64,  do you
> >> know why ARM64 is different from X86 (only fallback in no altmap
> >> case)?
>
> I think that's a purely theoretical issue: I assume that in any case we
> care about, the altmap should be reasonably sized and aligned such that
> this will always succeed.
>
> To me it even sounds like the best idea to *consistently* fail if there
> is no more space in the altmap, even if we'd have to fallback to PTE
> (again, highly unlikely that this is relevant in practice). Could
> indicate an altmap-size configuration issue.

Does David's explanation make things clear? Moreover, I think Dan's
dedicated comments "no fallback" implies that his design is carefully
considered. So I think the generic version using the X86 logic is just
OK.

Huacai
>
> --
> Thanks,
>
> David / dhildenb
>
>