[PATCH v13 08/14] mm: multi-gen LRU: support page table walks

[Liam Howlett]

* Yu Zhao <yuzhao at google.com> [220706 18:26]:
> On Wed, Jul 06, 2022 at 04:00:17PM -0600, Yu Zhao wrote:
> 
> ...
> 
> > +/*
> > + * Some userspace memory allocators map many single-page VMAs. Instead of
> > + * returning back to the PGD table for each of such VMAs, finish an entire PMD
> > + * table to reduce zigzags and improve cache performance.
> > + */
> > +static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
> > +			 unsigned long *vm_start, unsigned long *vm_end)
> > +{
> > +	unsigned long start = round_up(*vm_end, size);
> > +	unsigned long end = (start | ~mask) + 1;
> > +
> > +	VM_WARN_ON_ONCE(mask & size);
> > +	VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
> > +
> > +	while (args->vma) {
> > +		if (start >= args->vma->vm_end) {
> > +			args->vma = args->vma->vm_next;
> > +			continue;
> > +		}
> > +
> > +		if (end && end <= args->vma->vm_start)
> > +			return false;
> > +
> > +		if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args)) {
> > +			args->vma = args->vma->vm_next;
> > +			continue;
> > +		}
> > +
> > +		*vm_start = max(start, args->vma->vm_start);
> > +		*vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
> > +
> > +		return true;
> > +	}
> > +
> > +	return false;
> > +}
> 
> To make the above work on top of the Maple Tree:
> 
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 7096ff7836db..c0c1195da803 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -3689,23 +3689,14 @@ static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk
>  {
>  	unsigned long start = round_up(*vm_end, size);
>  	unsigned long end = (start | ~mask) + 1;
> +	VMA_ITERATOR(vmi, args->mm, start);
>  
>  	VM_WARN_ON_ONCE(mask & size);
>  	VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
>  
> -	while (args->vma) {
> -		if (start >= args->vma->vm_end) {
> -			args->vma = args->vma->vm_next;
> +	for_each_vma_range(vmi, args->vma, end) {
> +		if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
>  			continue;
> -		}
> -
> -		if (end && end <= args->vma->vm_start)
> -			return false;
> -
> -		if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args)) {
> -			args->vma = args->vma->vm_next;
> -			continue;
> -		}
>  
>  		*vm_start = max(start, args->vma->vm_start);
>  		*vm_end = min(end - 1, args->vma->vm_end - 1) + 1;

This looks correct to me.  There are a few subtle things that happen
with this change but it all works out rather nicely.


Thanks,
Liam