On Fri, Jun 15, 2018 at 12:14:21PM -0400, Rik van Riel wrote:
> On Fri, 2018-06-15 at 18:09 +0300, Michael S. Tsirkin wrote:
> > On Mon, Jun 11, 2018 at 11:18:57AM -0400, nilal@xxxxxxxxxx wrote:
> > > From: Nitesh Narayan Lal <nilal@xxxxxxxxxx>
>
> > > + write_seqlock(&guest_page_lock);
> > > + while (idx < MAX_FGPT_ENTRIES) {
> > > + unsigned long pfn = free_page_obj[idx].pfn;
> > > + unsigned long pfn_end = free_page_obj[idx].pfn +
> > > + free_page_obj[idx].pages -
> > > 1;
> > > + bool prev_free = false;
> > > +
> > > + while (pfn <= pfn_end) {
> >
> > It seems like this can be MAX_FGPT_ENTRIES^2 so 1000000 cycles in the
> > inner-most loop in the worst case, all with interrupts disabled and
> > all
> > kind of MM locks held. This kind of latency likely won't be
> > acceptable
> > to the MM crowd even if it only happens once in a while.
>
> That makes me wonder if instead of spending time making
> this code perfect, it would make time to also implement
> a prototype of your bitmap page hinting idea, and compare
> the two.
>
> This project is probably at the "50% effort" stage, and
> the bitmap page hinting approach seems like an interesting
> possible different tradeoff...
Since you mention a bitmap, this is suspeciously similar to
frontswap. Would it help to support tmem on kvm?
> --
> All Rights Reversed.
