Re: [PATCH v3 3/9] cxl/mem: Adjust cxl_mem_find_port() to find an RCH's port

[Dan Williams <dan.j.williams@intel.com>]

Robert Richter wrote:
> On 14.11.22 15:45:19, Dan Williams wrote:
> > Robert Richter wrote:
> > > The PCIe software view of an RCH and RCD is different to VH mode. An
> > > RCD is paired with an RCH and shows up as RCiEP with a parent already
> > > pointing to a PCI bridge (struct pci_host_bridge). In contrast, in VH
> > > mode an PCI Express Endpoint is a PCI type 0 device with a PCI type 1
> > > device as parent (struct pci_dev, most of the time a downstream switch
> > > port, but could also be a root port). The following hierarchy applies
> > > in VH mode:
> > > 
> > >  CXL memory device, cxl_memdev                               endpoint
> > >  └──PCIe Endpoint (type 0), pci_dev                           |
> > >     └──Downstream Port (type 1), pci_dev (Nth switch)        portN
> > >        └──Upstream Port (type 1), pci_dev (Nth switch)        |
> > >           :                                                   :
> > >           └──Downstream Port (type 1), pci_dev (1st switch)  port1
> > >              └──Upstream Port (type 1), pci_dev (1st switch)  |
> > >                 └──Root Port (type 1), pci_dev                |
> > >                    └──PCI host bridge, pci_host_bridge       port0
> > >                       :                                       |
> > >                       :..ACPI0017, acpi_dev                  root
> > > 
> > >  (There can be zero or any other number of switches in between.)
> > > 
> > > An iterator through the grandparents takes us to the root port which
> > > is registered as dport to the bridge. The next port an endpoint is
> > > connected to can be determined by using the grandparent of the memory
> > > device as a dport_dev in cxl_mem_find_port().
> > > 
> > > The same does not work in RCD mode where only an RCiEP is connected to
> > > the host bridge:
> > > 
> > >  CXL memory device, cxl_memdev                               endpoint
> > >  └──PCIe Endpoint (type 0), pci_dev                           |
> > >     └──PCI host bridge, pci_host_bridge                      port0
> > >        :                                                      |
> > >        :..ACPI0017, acpi_dev                                 root
> > > 
> > > Here, an endpoint is directly connected to the host bridge without a
> > > type 1 PCI device (root or downstream port) in between. To link the
> > > endpoint to the correct port, the endpoint's PCI device (parent of the
> > > memory device) must be taken as dport_dev arg in cxl_mem_find_port().
> > > 
> > > Change cxl_mem_find_port() to find an RCH's port.
> > > 
> > > Signed-off-by: Robert Richter <rrichter@amd.com>
> > > ---
> > >  drivers/cxl/core/port.c | 38 ++++++++++++++++++++++++++++++++++++++
> > >  1 file changed, 38 insertions(+)
> > > 
> > > diff --git a/drivers/cxl/core/port.c b/drivers/cxl/core/port.c
> > > index 0431ed860d8e..d10c3580719b 100644
> > > --- a/drivers/cxl/core/port.c
> > > +++ b/drivers/cxl/core/port.c
> > > @@ -1354,6 +1354,14 @@ static int add_port_attach_ep(struct cxl_memdev *cxlmd,
> > >  	return rc;
> > >  }
> > >  
> > > +static inline bool is_cxl_restricted(struct cxl_memdev *cxlmd)
> > > +{
> > > +	struct device *parent = cxlmd->dev.parent;
> > > +	if (!dev_is_pci(parent))
> > > +		return false;
> > > +	return pci_pcie_type(to_pci_dev(parent)) == PCI_EXP_TYPE_RC_END;
> > > +}
> > > +
> > >  int devm_cxl_enumerate_ports(struct cxl_memdev *cxlmd)
> > >  {
> > >  	struct device *dev = &cxlmd->dev;
> > > @@ -1433,9 +1441,39 @@ int devm_cxl_enumerate_ports(struct cxl_memdev *cxlmd)
> > >  }
> > >  EXPORT_SYMBOL_NS_GPL(devm_cxl_enumerate_ports, CXL);
> > >  
> > > +/*
> > > + * CXL memory device and port hierarchy:
> > > + *
> > > + * VH mode:
> > > + *
> > > + * CXL memory device, cxl_memdev                               endpoint
> > > + * └──PCIe Endpoint (type 0), pci_dev                           |
> > > + *    └──Downstream Port (type 1), pci_dev (Nth switch)        portN
> > > + *       └──Upstream Port (type 1), pci_dev (Nth switch)        |
> > > + *          :                                                   :
> > > + *          └──Downstream Port (type 1), pci_dev (1st switch)  port1
> > > + *             └──Upstream Port (type 1), pci_dev (1st switch)  |
> > > + *                └──Root Port (type 1), pci_dev                |
> > > + *                   └──PCI host bridge, pci_host_bridge       port0
> > > + *                      :                                       |
> > > + *                      :..ACPI0017, acpi_dev                  root
> > > + *
> > > + * (There can be zero or any other number of switches in between.)
> > > + *
> > > + * RCD mode:
> > > + *
> > > + * CXL memory device, cxl_memdev                               endpoint
> > > + * └──PCIe Endpoint (type 0), pci_dev                           |
> > > + *    └──PCI host bridge, pci_host_bridge                      port0
> > > + *       :                                                      |
> > > + *       :..ACPI0017, acpi_dev                                 root
> > > + */
> > >  struct cxl_port *cxl_mem_find_port(struct cxl_memdev *cxlmd,
> > >  				   struct cxl_dport **dport)
> > >  {
> > > +	if (is_cxl_restricted(cxlmd))
> > > +		return find_cxl_port(cxlmd->dev.parent, dport);
> > > +
> > >  	return find_cxl_port(grandparent(&cxlmd->dev), dport);
> > 
> > I do not see why this change is needed. For example:
> > 
> > # readlink -f /sys/bus/cxl/devices/mem0
> > /sys/devices/pci0000:38/0000:38:00.0/mem0
> > # cxl list -BT
> > [
> >   {
> >     "bus":"root0",
> >     "provider":"ACPI.CXL",
> >     "nr_dports":1,
> >     "dports":[
> >       {
> >         "dport":"pci0000:38",
> >         "id":49
> >       }
> >     ]
> >   }
> > ]
> > 
> > ...so, in this case, the grandparent of "mem0" is "pci0000:38", and
> > "pci0000:38" is a dport. Unmodified cxl_mem_find_port() will do the
> > right thing and find that this CXL RCIEP is directly connected to
> > "root0".
> 
> find_cxl_port() uses the dport_dev, not the uport_dev. A lookup of
> pci0000:38 gives the cxl root (ACPI.CXL).

...but that is what I would expect. I.e. that RCDs appear directly
connected to the cxl root with no intervening cxl_port instance, and RCH
host-bridges only serve the role of dport devices. This also matches the
diagram from 9.11.8 CXL Devices Attached to an RCH where a
downstream-port RCRB in the host bridge is directly connected to the
RCIEP endpoint.