On Tue, Jul 30, 2019 at 01:54:06PM -0700, Dave Hansen wrote:
> On 7/30/19 1:46 PM, Uladzislau Rezki wrote:
> >> + /*
> >> + * If required width exeeds current VA block, move
> >> + * base downwards and then recheck.
> >> + */
> >> + if (base + end > va->va_end) {
> >> + base = pvm_determine_end_from_reverse(&va, align) - end;
> >> + term_area = area;
> >> + continue;
> >> + }
> >> +
> >> /*
> >> * If this VA does not fit, move base downwards and recheck.
> >> */
> >> - if (base + start < va->va_start || base + end > va->va_end) {
> >> + if (base + start < va->va_start) {
> >> va = node_to_va(rb_prev(&va->rb_node));
> >> base = pvm_determine_end_from_reverse(&va, align) - end;
> >> term_area = area;
> >> --
> >> 2.21.0
> >>
> > I guess it is NUMA related issue, i mean when we have several
> > areas/sizes/offsets. Is that correct?
>
> I don't think NUMA has anything to do with it. The vmalloc() area
> itself doesn't have any NUMA properties I can think of. We don't, for
> instance, partition it into per-node areas that I know of.
>
> I did encounter this issue on a system with ~100 logical CPUs, which is
> a moderate amount these days.
>
Percpu memory does have this restriction when we embed the first chunk
as we need to preserve the offsets. So that is when we'd require
multiple areas in the vma.
I didn't see the original patches come through, but this seems like it
restores the original functionality. FWIW, this way of finding space
isn't really smart, so it's possible we want to revisit this.
Acked-by: Dennis Zhou <dennis@xxxxxxxxxx>
Thanks,
Dennis
