[RFT PATCH v3 12/27] of/address: Add infrastructure to declare MMIO as non-posted

[Arnd Bergmann]

On Wed, Mar 10, 2021 at 6:01 PM Rob Herring <robh at kernel.org> wrote:
>
> On Wed, Mar 10, 2021 at 1:27 AM Hector Martin <marcan at marcan.st> wrote:
> >
> > On 10/03/2021 07.06, Rob Herring wrote:
> > >> My main concern here is that this creates an inconsistency in the device
> > >> tree representation that only works because PCI drivers happen not to
> > >> use these code paths. Logically, having "nonposted-mmio" above the PCI
> > >> controller would imply that it applies to that bus too. Sure, it doesn't
> > >> matter for Linux since it is ignored, but this creates an implicit
> > >> exception that PCI buses always use posted modes.
> > >
> > > We could be stricter that "nonposted-mmio" must be in the immediate
> > > parent. That's kind of in line with how addressing already works.
> > > Every level has to have 'ranges' to be an MMIO address, and the
> > > address cell size is set by the immediate parent.
> > >
> > >> Then if a device comes along that due to some twisted fabric logic needs
> > >> nonposted nGnRnE mappings for PCIe (even though the actual PCIe ops will
> > >> end up posted at the bus anyway)... how do we represent that? Declare
> > >> that another "nonposted-mmio" on the PCIe bus means "no, really, use
> > >> nonposted mmio for this"?
> > >
> > > If we're strict, yes. The PCI host bridge would have to have "nonposted-mmio".
> >
> > Works for me; then let's just make it non-recursive.
> >
> > Do you think we can get rid of the Apple-only optimization if we do
> > this? It would mean only looking at the parent during address
> > resolution, not recursing all the way to the top, so presumably the
> > performance impact would be quite minimal.

Works for me.

> Yeah, that should be fine. I'd keep an IS_ENABLED() config check
> though. Then I'll also know if anyone else needs this.

Ok, makes sense.

Conceptually, I'd like to then see a check that verifies that the
property is only set for nodes whose parent also has it set, since
that is how AXI defines it: A bus can wait for the ack from its
child node, or it can acknowledge the write to its parent early.
However, this breaks down as soon as a bus does the early ack:
all its children by definition use posted writes (as seen by the
CPU), even if they wait for stores that come from other masters.

Does this make sense to you?

       Arnd