Hugetlb HPTE hashing weirdness

[Adam Litke]

On Fri, 2005-09-23 at 15:15 +1000, Benjamin Herrenschmidt wrote:
> > Attached is my somewhat hacky hugetlb test suite.  Build it with 'make
> > all'.  I've been reproducing with:
> > 	for i in `seq 1 10`; do ./mmap-gettest 1 10; done
> > After about the third or fourth iteration, you should start to see test
> > failures.
> 
> What machine do you use ?

I've been doing my testing on a p650 in bare metal mode.  Do you need
any other specific info?

> I can reproduce the problem on a g5 here but strangely not on any LPAR
> machine I've tested so far (all POWER5)... 
> 
> I did find a bug in the hugepage code though I doubt it's the culprit,
> I'll test a fix on the g5.

That patch (as included in 2.6.14-rc2-git5) makes the problem disappear
for the prefault case but it is still broken for demand faulting.  Seems
some HPTEs are still not flushed.  I'll try on some LPAR systems and see
if I can reproduce there.

-- 
Adam Litke - (agl at us.ibm.com)
IBM Linux Technology Center