Re: [PATCH v8 1/2] hugetlb: don't delete vma_lock in hugetlb MADV_DONTNEED processing

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

On 11/10/22 14:22, Nadav Amit wrote:
> On Nov 10, 2022, at 1:48 PM, Mike Kravetz <mike.kravetz@oracle.com> wrote:
> 
> >>> void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
> >>> 			  unsigned long end, struct page *ref_page,
> >>> 			  zap_flags_t zap_flags)
> >>> {
> >>> +	struct mmu_notifier_range range;
> >>> 	struct mmu_gather tlb;
> >>> 
> >>> +	mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
> >>> +				start, end);
> >>> +	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
> >>> 	tlb_gather_mmu(&tlb, vma->vm_mm);
> >>> +
> >>> 	__unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
> >> 
> >> Is there a reason for not using range.start and range.end?
> > 
> > After calling adjust_range_if_pmd_sharing_possible, range.start - range.end
> > could be much greater than the range we actually want to unmap.  The range
> > gets adjusted to account for pmd sharing if that is POSSIBLE.  It does not
> > know for sure if we will actually 'unshare a pmd'.
> > 
> > I suppose adjust_range_if_pmd_sharing_possible could be modified to actually
> > check if unmapping will result in unsharing, but it does not do that today.
> 
> Thanks for the explanation. It’s probably me, but I am still not sure that I
> understand the the different between __unmap_hugepage_range() using (start,
> end) and __zap_page_range_single() using (address, range.end). Perhaps it
> worth a comment in the code?

__zap_page_range_single is wrong.  It should have been updated to use
the range address - (address + size).

At Peter's suggestion the mmu notifier updates will be broken out in a
separate patch.  I will also add comments, to make this easier to follow.

> But anyhow… shouldn’t unmap_hugepage_range() call
> mmu_notifier_invalidate_range_start()?

Yes, thanks!

-- 
Mike Kravetz