Re: [PATCH v6] hugetlb: don't delete vma_lock in hugetlb MADV_DONTNEED processing

[Nadav Amit]

On Oct 31, 2022, at 3:34 PM, Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:

> madvise(MADV_DONTNEED) ends up calling zap_page_range() to clear the page
> tables associated with the address range.  For hugetlb vmas,
> zap_page_range will call __unmap_hugepage_range_final.  However,
> __unmap_hugepage_range_final assumes the passed vma is about to be removed
> and deletes the vma_lock to prevent pmd sharing as the vma is on the way
> out.  In the case of madvise(MADV_DONTNEED) the vma remains, but the
> missing vma_lock prevents pmd sharing and could potentially lead to issues
> with truncation/fault races.
> 

[snip]

> index 978c17df053e..517c8cc8ccb9 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -3464,4 +3464,7 @@ madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
>  */
> #define  ZAP_FLAG_DROP_MARKER        ((__force zap_flags_t) BIT(0))
> 
> +/* Set in unmap_vmas() to indicate an unmap call.  Only used by hugetlb */
> +#define  ZAP_FLAG_UNMAP              ((__force zap_flags_t) BIT(1))

PeterZ wants to add ZAP_FLAG_FORCE_FLUSH that would be set on
zap_pte_range(). Not sure you would want to combine them both together, but
at least be aware of potential conflict.

https://lore.kernel.org/all/Y1f7YvKuwOl1XEwU@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/

[snip]

> +#ifdef CONFIG_ADVISE_SYSCALLS
> +/*
> + * Similar setup as in zap_page_range().  madvise(MADV_DONTNEED) can not call
> + * zap_page_range for hugetlb vmas as __unmap_hugepage_range_final will delete
> + * the associated vma_lock.
> + */
> +void clear_hugetlb_page_range(struct vm_area_struct *vma, unsigned long start,
> +				unsigned long end)
> +{
> +	struct mmu_notifier_range range;
> +	struct mmu_gather tlb;
> +
> +	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 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);
> +	update_hiwater_rss(vma->vm_mm);
> +	mmu_notifier_invalidate_range_start(&range);
> +
> +	__unmap_hugepage_range_locking(&tlb, vma, start, end, NULL, 0);
> +
> +	mmu_notifier_invalidate_range_end(&range);
> +	tlb_finish_mmu(&tlb);
> }
> +#endif

I hate ifdef’s. And the second definition of clear_hugetlb_page_range() is
confusing since it does not have an ifdef at all. . How about moving the
ifdef’s into the function like being done in io_madvise_prep()? I think it
is less confusing.

[ snip ]

> 
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1671,7 +1671,7 @@ void unmap_vmas(struct mmu_gather *tlb, struct maple_tree *mt,
> {
> 	struct mmu_notifier_range range;
> 	struct zap_details details = {
> -		.zap_flags = ZAP_FLAG_DROP_MARKER,
> +		.zap_flags = ZAP_FLAG_DROP_MARKER | ZAP_FLAG_UNMAP,
> 		/* Careful - we need to zap private pages too! */
> 		.even_cows = true,
> 	};
> @@ -1704,15 +1704,21 @@ void zap_page_range(struct vm_area_struct *vma, unsigned long start,
> 	MA_STATE(mas, mt, vma->vm_end, vma->vm_end);
> 
> 	lru_add_drain();
> -	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
> -				start, start + size);
> 	tlb_gather_mmu(&tlb, vma->vm_mm);
> 	update_hiwater_rss(vma->vm_mm);
> -	mmu_notifier_invalidate_range_start(&range);
> 	do {
> -		unmap_single_vma(&tlb, vma, start, range.end, NULL);
> +		mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma,
> +				vma->vm_mm,
> +				max(start, vma->vm_start),
> +				min(start + size, vma->vm_end));
> +		if (is_vm_hugetlb_page(vma))
> +			adjust_range_if_pmd_sharing_possible(vma,
> +				&range.start,
> +				&range.end);
> +		mmu_notifier_invalidate_range_start(&range);
> +		unmap_single_vma(&tlb, vma, start, start + size, NULL);

Is there a reason that you wouldn’t use range.start and range.end here?
At least for consistency.