Re: [RFC PATCH 04/26] hugetlb: make huge_pte_lockptr take an explicit shift argument.

[Mike Kravetz <mike.kravetz@oracle.com>]

On 06/30/22 09:23, James Houghton wrote:
> On Thu, Jun 30, 2022 at 2:35 AM Muchun Song <songmuchun@bytedance.com> wrote:
> >
> > On Wed, Jun 29, 2022 at 03:24:45PM -0700, Mike Kravetz wrote:
> > > On 06/29/22 14:39, James Houghton wrote:
> > > > On Wed, Jun 29, 2022 at 2:04 PM Mike Kravetz <mike.kravetz@oracle.com> wrote:
> > > > >
> > > > > On 06/29/22 14:09, Muchun Song wrote:
> > > > > > On Mon, Jun 27, 2022 at 01:51:53PM -0700, Mike Kravetz wrote:
> > > > > > > On 06/24/22 17:36, James Houghton wrote:
> > > > > > > > This is needed to handle PTL locking with high-granularity mapping. We
> > > > > > > > won't always be using the PMD-level PTL even if we're using the 2M
> > > > > > > > hugepage hstate. It's possible that we're dealing with 4K PTEs, in which
> > > > > > > > case, we need to lock the PTL for the 4K PTE.
> > > > > > >
> > > > > > > I'm not really sure why this would be required.
> > > > > > > Why not use the PMD level lock for 4K PTEs?  Seems that would scale better
> > > > > > > with less contention than using the more coarse mm lock.
> > > > > > >
> > > > > >
> > > > > > Your words make me thing of another question unrelated to this patch.
> > > > > > We __know__ that arm64 supports continues PTE HugeTLB. huge_pte_lockptr()
> > > > > > did not consider this case, in this case, those HugeTLB pages are contended
> > > > > > with mm lock. Seems we should optimize this case. Something like:
> > > > > >
> > > > > > diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
> > > > > > index 0d790fa3f297..68a1e071bfc0 100644
> > > > > > --- a/include/linux/hugetlb.h
> > > > > > +++ b/include/linux/hugetlb.h
> > > > > > @@ -893,7 +893,7 @@ static inline gfp_t htlb_modify_alloc_mask(struct hstate *h, gfp_t gfp_mask)
> > > > > >  static inline spinlock_t *huge_pte_lockptr(struct hstate *h,
> > > > > >                                            struct mm_struct *mm, pte_t *pte)
> > > > > >  {
> > > > > > -       if (huge_page_size(h) == PMD_SIZE)
> > > > > > +       if (huge_page_size(h) <= PMD_SIZE)
> > > > > >                 return pmd_lockptr(mm, (pmd_t *) pte);
> > > > > >         VM_BUG_ON(huge_page_size(h) == PAGE_SIZE);
> > > > > >         return &mm->page_table_lock;
> > > > > >
> > > > > > I did not check if elsewhere needs to be changed as well. Just a primary
> > > > > > thought.
> > > >
> > > > I'm not sure if this works. If hugetlb_pte_size(hpte) is PAGE_SIZE,
> > > > then `hpte.ptep` will be a pte_t, not a pmd_t -- I assume that breaks
> > > > things. So I think, when doing a HugeTLB PT walk down to PAGE_SIZE, we
> > > > need to separately keep track of the location of the PMD so that we
> > > > can use it to get the PMD lock.
> > >
> > > I assume Muchun was talking about changing this in current code (before
> > > your changes) where huge_page_size(h) can not be PAGE_SIZE.
> > >
> >
> > Yes, that's what I meant.
> 
> Right -- but I think my point still stands. If `huge_page_size(h)` is
> CONT_PTE_SIZE, then the `pte_t *` passed to `huge_pte_lockptr` will
> *actually* point to a `pte_t` and not a `pmd_t` (I'm pretty sure the
> distinction is important). So it seems like we need to separately keep
> track of the real pmd_t that is being used in the CONT_PTE_SIZE case
> (and therefore, when considering HGM, the PAGE_SIZE case).

Ah yes, that is correct.  We would be passing in a pte not pmd in this
case.

> 
> However, we *can* make this optimization for CONT_PMD_SIZE (maybe this
> is what you originally meant, Muchun?), so instead of
> `huge_page_size(h) == PMD_SIZE`, we could do `huge_page_size(h) >=
> PMD_SIZE && huge_page_size(h) < PUD_SIZE`.
> 

Another 'optimization' may exist in hugetlb address range scanning code.
We currently have something like:

for addr=start, addr< end, addr += huge_page_size
	pte = huge_pte_offset(addr)
	ptl = huge_pte_lock(pte)
	...
	...
	spin_unlock(ptl)

Seems like ptl will be the same for all entries on the same pmd page.
We 'may' be able to go from 512 lock/unlock cycles to 1.
-- 
Mike Kravetz