Re: 4GB+DEBUG_PAGEALLOC oopses with 2.6.0-test3-mm1

[Ingo Molnar <mingo@elte.hu>]

On Mon, 11 Aug 2003, Manfred Spraul wrote:

> >(in theory it's possible that kernel-internal mounts pass in a pointer
> >where pointer + PAGE_SIZE is not a valid kernel address - if this happens
> >then we'd get a hard crash.)

> Exactly that happens.
> I'm running with CONFIG_PAGE_DEBUG, i.e. unallocated pages are marked as 
> non-present in the linear mapping.

this is not a bug technically, unless the mount options are in the last
linearly mapped page. It is a bug to copy those unallocated bytes, but
they do not get to relied upon. Note that the non-4G code copies them just
as much.

> Regarding the i386 trap handler: show_registers tries to hexdump the 
> current instructions. It did a __get_user, to avoid causing a fault when 
> %eip is invalid.
> Now it contains:
> 
> >      if ((user_mode(regs) && get_user(c, eip)) ||
> >           (!user_mode(regs) && __direct_get_user(c, eip))) {
> >              printk(" Bad EIP value.");
> >              break;
> >      }
> 
> I.e. it's already fixed.

yeah, this was one of the latest patches that went into Andrew's tree.  
(This fix enabled us to get rid of the do_page_fault() hacks.)

	Ingo