lkml.org 
[lkml]   [2022]   [Sep]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 05/15] cxl/acpi: Add probe function to detect restricted CXL hosts in RCD mode
On Thu, 1 Sep 2022 08:01:05 +0200
Robert Richter <rrichter@amd.com> wrote:

> On 31.08.22 11:08:04, Jonathan Cameron wrote:
> > On Wed, 31 Aug 2022 10:15:53 +0200
>
> > Robert Richter <rrichter@amd.com> wrote:
> >
> > > Restricted CXL device (RCD) mode (formerly CXL 1.1) uses a different
> > > enumeration scheme other than CXL VH (formerly CXL 2.0). In RCD mode a
> > > host/device (RCH-RCD) pair shows up as a legal PCIe hierarchy with an
> > > ACPI host bridge ("PNP0A08" or "ACPI0016" HID) and RCiEP connected to
> > > it with a description of the CXL device.
> > >
> > > Add function cxl_restricted_host_probe() to probe RCD enumerated
> > > devices. The function implements a loop that detects all CXL capable
> > > ACPI PCI root bridges in the system (RCD mode only). The iterator
> > > function cxl_find_next_rch() is introduced to walk through all of the
> > > CXL hosts. The loop will then enable all CXL devices connected to the
> > > host. For now, only implement an empty loop with an iterator that
> > > returns all pci host bridges in the system.
> > >
> > > The probe function is triggered by adding an own root device for RCHs.
> > > This is different to CXL VH where an ACPI "ACPI0017" root device
> > > exists. Its detection starts the CXL host detection. In RCD mode such
> > > a device does not necessarily exists, so solve this by creating a
> > > plain platform device that is not an ACPI device and is root only for
> > > RCHs.
> >
> > If I read this correctly that platform device is created whether or not
> > there are any cxl devices in the system?
> >
> > Can we create it only if we find some devices that will be placed
> > under it later?
>
> This would move the host detection from probe to init which I wanted
> to avoid to better control driver init order dependencies.

It's a bit nasty either way. I can see your reasoning, but
definitely not keen on it if there is a plausible way to avoid.
>
> I could add a put_device() at the end of a probe so that it will be
> released in case no other references use it. This implies the refcount
> is maintained for parent devices. Or this needs to be added to. So if
> there are no children (hosts) attached to the root device after probe,
> it will disappear.

Unless there is precedence for that, it'll be weird enough to be
hard to maintain. I guess I can live with the ugliness if we can't
add something new to ACPI to base this off.

>
> > > @@ -531,7 +566,41 @@ static struct platform_driver cxl_acpi_driver = {
> > > .id_table = cxl_test_ids,
> > > };
> > >
> > > -module_platform_driver(cxl_acpi_driver);
> > > +static void cxl_acpi_device_release(struct device *dev) { }
> >
> > Why the empty release? Perhaps introduce this only when it
> > does something.
>
> The core device driver requires this in device_release() to be setup.
>
> There is nothing to do as the device is kept in a static struct.
> That's why it's empty.
Ah got it. I'd failed to register the static structure.

>
> -Robert

\
 
 \ /
  Last update: 2022-09-01 12:11    [W:0.075 / U:3.024 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site