Re: [virtio] [PATCH] pci: missing documentation for dealing with 64 bit config fields

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

On Mon, Apr 06, 2015 at 06:17:56PM +0000, James Bottomley wrote:
> On Thu, 2015-04-02 at 12:23 +0200, Cornelia Huck wrote:
> > On Thu, 2 Apr 2015 12:06:41 +0200
> > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > 
> > > On Thu, Apr 02, 2015 at 11:01:45AM +0200, Cornelia Huck wrote:
> > > > On Wed, 1 Apr 2015 20:42:13 +0200
> > > > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > 
> > > > > +For device configuration access, the driver MUST use 8-bit wide
> > > > > +accesses for 8-bit wide fields, 16-bit wide and aligned accesses
> > > > > +for 16-bit wide fields and 32-bit wide and aligned accesses for
> > > > > +32-bit and 64-bit wide fields. For 64-bit fields, the driver MAY
> > > > > +access each of the high and low 32-bit parts of the field
> > > > > +independently.
> > > > 
> > > > Doesn't the last sentence follow from "use 32-bit wide accesses"
> > > > already?
> > > 
> > > Someone might assume that driver must always access low part,
> > > then high part. Last sentence says it does not have to.
> > > 
> > 
> > OK, that clears it up then.
> 
> Not entirely.  There's a hidden gotcha here in terms of fields which
> perform actions.  For instance the status register is used to transmit
> actions and information from the guest to the host.  Problems happen if
> any action register is wider than 32 bits, because a 32 bit OS cannot
> now set it atomically.
> 
> There a various ways around this: we could say no action register may
> ever be wider than 32 bits (easy) or we could prescribe a mechanism for
> using 64 bit action registers on 32 bits (locking + defined initiation,
> say writing to the lower 32 bits triggers the action, so on 32 bit
> architectures, you lock, write upper, then write lower, then unlock).
> 
> I'd actually say it's useful clarification to code the former into the
> Spec until such time as someone finds a need for a 64 bit action
> register.
> 
> James
> 

This seems to belong to appendix B: creating new device types.
We already have this text there:
	Remember that configuration fields over 32 bits wide might not be
	atomically writable by the driver.

Is this sufficient? If not, we can add
	Therefore, it is generally not a good idea to have
	such writeable fields, and especially to have
	driver writes into such fields trigger device actions".

Is this what you had in mind?

-- 
MST