Re: [RFT] [SPARC] Emulate cmpxchg like parisc

[David Miller <davem@davemloft.net>]

From: Kyle McMartin <kyle@parisc-linux.org>
Date: Sat, 26 May 2007 10:45:09 -0400

> On Fri, May 25, 2007 at 10:00:36PM -0700, David Miller wrote:
> > > After some minor fixes this builds, and the DRM drivers also
> > > build again. I cannot test this since I do not have a machine with
> > > PCI or these cards.
> > > Removed your name in the comment, as that went out of fashion after
> > > we started using proper versioning systems.
> > > 
> > > I've got a ppc->sparc32 compiler if that would help...
> > 
> > So how in the world is something like this going to "work" with DRM?
> > Userland cannot disable interrupts and it can't take the magic lock
> > the kernel uses to provide mutual exclusion for the emulated
> > cmpxchg().
> > 
> 
> Hi David,
> 
> My knowledge of the DRM is weak, but as I understand it, the only time
> it is used is by the ioctl handlers, and not by userspace. I've added
> Dave Airlie to the CC list, hopefully he can enlighten us as to where
> else cmpxchg is used.
> 
> > Unless something %100 inside of the kernel will be the only consumers
> > of this, I'd recommend not adding it.  You can't let anything part of
> > a userland API try to make use of it, and DRM definitely falls into
> > that category.
> > 
> 
> istr there was a contentious thread on linux-arch about using cmpxchg in
> generic code a while ago. So it's possible architectures with crappy
> atomic support will need some kind of hack like this to even be able to
> use generic code in the future.

For purely kernel stuff it's fine, because the lock and the interrupt
disabling will be used consistently by all accesses.

But the words that operate on cmpxchg() in DRM are exported to
userspace, and therefore cannot make use of the locking scheme
that atomically challenged platforms need to make use of in
the kernel.