Re: [BUG] kernel BUG at arch/i386/mm/highmem.c:15! on 2.6.23-rc8/rc9

[Hugh Dickins <hugh@veritas.com>]

On Thu, 4 Oct 2007, Linus Torvalds wrote:
> 
> Ouch.

Very much so.

> 
> The patch looks obviously correct, to the point that I don't understand 
> how this bug happened in the first place.  It seems to have been 
> introduced by Nick in d0217ac04ca6591841e5665f518e38064f4e65bd ("mm: fault 
> feedback #1") if I read it right.

Yes, I think you are reading that right.

> 
> So "do_nonlinear_fault()" would effectively end up unmapping the PTE 
> twice, right? And if I understand the problem right, this wasn't noticed 
> immediately, because it probably only matters on:
> 
>  - CONFIG_HIGHPTE
> *and*
>  - a filesystem that allows VM_NONLINEAR, which became very rare with 
>    dirty accounting.
> 
> Correct?

Correct.  And though I eventually realized the nonlinear dirty accounting
issue (which I did highlight in the patch comment), I'd forgotten all
about needing CONFIG_HIGHPTE in addition to CONFIG_HIGHMEM - yes, that
increased the probability of overlooking it too.

No excuses that we did miss it in all these months those patches have
been around; but it's an object lesson in the trouble rare paths give.

Hugh