Re: [PATCHv3 RFC] uio: add generic driver for PCI 2.3 devices

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Jul 14, 2009 at 04:25:07PM -0700, Chris Wright wrote:
> * Michael S. Tsirkin (mst@redhat.com) wrote:
> > On Tue, Jul 14, 2009 at 11:03:38AM -0700, Chris Wright wrote:
> > > * Michael S. Tsirkin (mst@redhat.com) wrote:
> > > > - remove irqcontrol: user can enable interrupts by
> > > >   writing command register directly
> > > 
> > > Sorry if I gave you the impression that removing was needed.
> > > I actually think the irqcontrol was useful since it's atomic.
> > 
> > What I'm saying, it is not strictly needed (user can safely write 0 in
> > that register and worst case you just get an extra interrupt).  So let's
> > make the decision on what does irqcontrol do when we have a pressing
> > need for an extra kernel/user interface.
> 
> OK.  My concern is theoretical (as in the current design wouldn't
> trigger an issue):
> 
> cmd = pread()
>              ------+
>                     \
> cmd &= ~INTX_DISABLE +----+
>                     /     |
>             -------+      |
> pwrite(cmd)               |
>                           -
> During this window Command Reg can change[1] and cmd is stale
> 
> [1] due to irqhandler (doesn't matter in this case since it touches
>     same bit).  or due to some other thread updating Command Reg (also
>     doesn't matter since this does not happen right now).

Right. Note that the limitation that only a single thread should touch
config space through sysfs applies to any sub-byte field. If we wanted
to make such ops atomic, pci devices would need some kind of compare and
swap ioctl.

-- 
MST