[Pkg-uml-pkgs] Bug#495611: [uml-devel] Bug#495611: user-mode-linux: page_mapcount(page) went negative!

[Jeff Dike]

On Wed, Sep 03, 2008 at 12:01:19AM +0100, Graham Cobb wrote:
> The bisection is complete.  It took longer than I expected as many of the 
> versions git-bisect wanted to test would not build UML without tracking down 
> some additional patches. 
> 
> The bug seems to have been introduced with git commit 
> 3963333fe6767f15141ab2dc3b933721c636c212 (uml: cover stubs with a VMA). 

Excellent, the backtrace now makes a bit of sense.  It looks like the
VMAs at the top of the address space are being overmapped by a call to
mmap.  The question is why does it look like there's a big enough hole
there, when it ends up unmapping the stubs in order to make room for
the mmap.

> With hindsight I suppose I could have guessed that would be the commit: the 
> crash seemed to be when unmapping a VMA.  If my reading of the page flags is 
> correct the problem is that the page has the PG_reserved flag set, so what is 
> it doing in the VMA which is being unmapped?

A page of kernel code is mapped into the process.  This page (and all
other pages containing kernel text) is marked reserved during boot.

> Let me know if you want me to run some more tests, for example if you want to 
> add some printk's in the code to understand more about what is going on.

I'm going to need some more information.  I'll get back to you with a
patch...

				Jeff

-- 
Work email - jdike at linux dot intel dot com