Re: [PATCH v5 1/2] PCI: iproc: Retry request when CRS returned from EP

From: Bjorn Helgaas
Date: Fri Aug 04 2017 - 10:36:59 EST


On Fri, Aug 04, 2017 at 07:48:53PM +0530, Oza Oza wrote:
> On Fri, Aug 4, 2017 at 6:57 PM, Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > On Fri, Aug 04, 2017 at 11:40:46AM +0530, Oza Oza wrote:
> >> On Fri, Aug 4, 2017 at 11:29 AM, Oza Oza <oza.oza@xxxxxxxxxxxx> wrote:
> >> > On Fri, Aug 4, 2017 at 12:27 AM, Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> >> >> On Thu, Aug 03, 2017 at 01:50:29PM +0530, Oza Oza wrote:
> >> >>> On Thu, Aug 3, 2017 at 2:34 AM, Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> >> >>> > On Thu, Jul 06, 2017 at 08:39:41AM +0530, Oza Pawandeep wrote:
> > ...
> >
> >> >>> > What about CRS status for a config *write*? There's nothing here to
> >> >>> > reissue those.
> >> >>>
> >> >>> No, we do not need there, because read will always be issued first
> >> >>> before any write.
> >> >>> so we do not need to implement write.
> >> >>
> >> >> How so? As far as I know, there's nothing in the spec that requires
> >> >> the first config access to a device to be a read, and there are
> >> >> reasons why we might want to do a write first:
> >> >> http://lkml.kernel.org/r/5952D144.8060609@xxxxxxxxxx
> >> >>
> >> >
> >> > I understand your point here. my thinking was during enumeration
> >> > process first read will always be issued
> >> > such as vendor/device id.
> >> > I will extend this implementation for write.
> >>
> >> I am sorry, but I just released that, it is not possible to implement
> >> retry for write.
> >> the reason is:
> >>
> >> we have indirect way of accessing configuration space access.
> >> for e.g.
> >> for config write:
> >>
> >> A) write to to addr register.
> >> B) write to data register
> >>
> >> now above those 2 registers are implemented by host bridge (not in
> >> PCIe core IP).
> >> there is no way of knowing for software, if write has to be retried.
> >>
> >> e.g. I can not read data register (step B) to check if write was successful.
> >> I have double checked this with internal ASIC team here.
> >
> > The bottom line is that you're saying this hardware cannot correctly
> > support CRS. Maybe the workaround you're proposing will work in many
> > cases, but we need to acknowledge in the code and changelog that there
> > are issues we might trip over.
>
> yes this is precisely right.
>
> 1) I will have to add notes in the code as you are suggesting.
> 2) I will add documentation notes in the Change-log.
>
> But even going forward, we will still have one more separate register
> in host bridge,
> which will be dedicated to CRS. but again to a very limited extent.
> because CRS software visibility bit will not have any effect, (e.g. HW
> is not going to consider it).

I assume your Root Capabilities CRS Software Visibility bit is zero,
indicating that the Root Port does not support returning CRS status to
software, and that the Root Control CRS Software Visibility Enable bit
is hardwired to zero.

This is worth mentioning somewhere because it makes it clear that your
config accessor will never return CRS status.

If these bits are not hardwired to zero, you should adjust your config
accessors to make them appear zero to higher-level software, so we
don't try to enable CRS.