MPC885 hang in a lwarx/stwcx insn crossing a page boundary

[Joakim Tjernlund]

On Tue, 2007-04-24 at 07:55 +0200, Heiko Schocher wrote:
> Hello,
> 
> I have a MPC885 and a running 2.6.16 kernel on it. I use the CPU15
> Errata Patch from
> (http://ozlabs.org/pipermail/linuxppc-dev/2007-April/033789.html)
> 
> Now if a lwarx/stwcx insn crossing a page boundary, the CPU hangs in
> an endless loop, because the reservation allocated from the lwarx gets
> cleared in the case of a tlb miss. And the CPU15 workaround invalidates
> the tlbs around the page we're faulting in, so we cause a nasty infinite
> loop. I started a service request at Freescale, and they are searching
> why a tlbie clears the reservation or under which circumstances this
> appears.
> 
> Without the CPU15 Errata patch it runs fine, but i can reproduce the
> CPU15 Bug, so I need the or some CPU15 Errata patch ...
> 
> I also have a running 2.4.24-pre2 Kernel on this board, with the same
> CPU15 Errata patch, and there the hang doesnt occur!! Some ideas where
> the differences between 2.4.24-pre2 and 2.6.16 are?
> 
> I also didnt get this hang in userspace with the 2.6.16er Kernel.
> 
> So I see at the moment the following possibilities:
> 
> a) patching the gcc, so that the CPU15 Errata is no more necessary
>    My favorit.
> 
> b) upgrade the CPU15 Bug like in the MPC885 Errata Sheet from Freescale
>    (I think this is to much code for in an Exceptionhandler ...)
> 
> c) I made before every lwarcx a ".align" so that the insn is on one
>    page. This is actually running without seeing problems on the board.
>    But it is a "errata" for the "errata" ...
> 
> d) ...
> 
> Some suggestions, ideas? Especially, why the hang doesnt occur with a
> 2.4.24-pre2 kernel?

A guess:
The TLB Miss handler changed for the case when "Level 1 entry" is null.
2.6 will force a TLB Error to happen, but 2.4 will(at least used to)
bail out via DataAccess.

 Jocke

> 
> thanks
> Heiko