* Yu Zhao <yuzhao@xxxxxxxxxx> [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
