On Tue, Mar 30, 2010 at 10:45:48PM +0200, Juergen Lock wrote:
> On Tue, Mar 30, 2010 at 01:33:15PM -0700, Richard Henderson wrote:
> > On 03/30/2010 01:09 PM, Juergen Lock wrote:
> > > Oh sorry if that was not clear, things go into swap if I _replace_ the
> > > endaddr ~0ul (which caused the assert) with the max value the assert
> > > still tolerates i.e.
> > > ((abi_ulong)1 << L1_MAP_ADDR_SPACE_BITS) - 1
> > > which in this case seems to be 0x7fffffffffff:
> >
> > Yes, I got that. And I see from ...
> >
> > > #3 0x0000000060012731 in page_set_flags (start=140737488224256,
> > > end=18446744073709551615, flags=32)
> >
> > ... here that the range we're reserving is
> >
> > 0x7ffffffe0000 ... 0x7fffffffffff
> >
> > which is a mere 128k range. Which ought to allocate no more than
> > a single leaf page table (and thus N-1 pages for the N-level table).
> >
> > Which doesn't answer the question of why you'd wind up running out
> > of memory.
>
> Ah yeah our mails crossed each other :) Yeah I don't know...
Ok found the problem: Reads from /compat/linux/proc/self/maps are
broken i.e. return partial lines if (at least I think that's the reason)
the mappings change between read() calls. (It got a line like this
after the first line that originally caused the assert:
fe0000-800000000000 rwxp 00020000 00:00 0
)
Well good that we now use kinfo_getvmmap(), I hope noone needs
bsd-user on FreeBSD 6.x...
Oh well...
Juergen
