Re: [PATCH] x86_64: fix page table size

["Yinghai Lu" <yhlu.kernel@gmail.com>]

On Feb 7, 2008 12:09 AM, Ingo Molnar <mingo@elte.hu> wrote:
>
> * Yinghai Lu <Yinghai.Lu@Sun.COM> wrote:
>
> > [PATCH] x86_64: fix page table size
>
> > Entering add_active_range(0, 1048576, 67239936) 2 entries of 3200 used
> > end_pfn_map = 67239936
> > Kernel panic - not syncing: Overlapping early reservations 8000-109fff
> > PGTABLE to 9bc00-9dbff EBDA
> >
> > change back the logic. we DO need extra space for pmds when
> > direct_gbpages is not there.
>
> > @@ -370,7 +370,7 @@ static void __init find_early_table_space(unsigned long end)
> >
> >       puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
> >       tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
> > -     if (direct_gbpages) {
> > +     if (!direct_gbpages) {
>
> thanks Yinghai, applied!
>
> I'm wondering why this bug didnt trigger more widely. It seems to me it
> needs some serious amount of RAM to trigger this bug - correct?

yes.  need 256g.
128g without patch it still work. because round_up(.., PAGE_SIZE) get
some extra.

>
> btw., it would be nice to have some "lots of RAM simulation" debugging
> code which would just _fake_ a really large e820 map and would in the
> end throw away the 'fake' pages later during bootup. Perhaps tell the
> early allocator to never allocate into these fake areas [via an struct
> e820 entry flag], but all our sizing code and the boot bitmaps, etc.
> would be sized accordingly, as if we had this much RAM - and we'd
> trigger these nuances. We could put this into a new "fakemem=128GB" boot
> option and hence we could boot with fakemem=128GB on a 2GB box and could
> at least hope to be able to boot [with some serious amount of RAM wasted
> on over-sized pagetables, allocator bitmaps and mem_map[]]. Hm?

sound interesting. but need some variable to prevent using non exist page.

YH