On Tue, Jun 15, 2004 at 08:56:11PM -0700, Andrew Morton wrote:
> > I found this problem on a MIPS machine. The problem is
> > likely to happen on other register-rich RISC arches too.
> > cmdcnt needs to be volatile since it is modified by
> > irq routine and read by normal process context.
> volatile is not the preferred way to fix this up. This points at either a
> locking error in the psmouse driver or a missing "memory" thingy in the
> mips port somewhere.
> Please describe the bug which led to this patch. Where was it getting stuck?
My current BK tree has this fixed using atomic bitfields, which do
compilation and memory barriers. I plan to sync it to Linus post 2.6.7.
SuSE Labs, SuSE CR