Received-SPF: pass (google.com: domain of linux-kernel+bounces-156581-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Date: Wed, 24 Apr 2024 11:56:08 +0300
From: Mika Westerberg <mika.westerberg@linux.intel.com>
To: Mario Limonciello <mario.limonciello@amd.com>
Cc: Esther Shimanovich <eshimanovich@chromium.org>,
	Dmitry Torokhov <dmitry.torokhov@gmail.com>,
	Lukas Wunner <lukas@wunner.de>, Bjorn Helgaas <bhelgaas@google.com>,
	linux-pci@vger.kernel.org, linux-kernel@vger.kernel.org,
	Rajat Jain <rajatja@google.com>
Subject: Re: [PATCH v4] PCI: Relabel JHL6540 on Lenovo X1 Carbon 7,8
Message-ID: <20240424085608.GE112498@black.fi.intel.com>
References: <20240123061820.GL2543524@black.fi.intel.com>
 <CA+Y6NJFMDcB7NV49r2WxFzcfgarRiWsWO0rEPwz43PKDiXk61g@mail.gmail.com>
 <CA+Y6NJGz6f8hE4kRDUTGgCj+jddUyHeD_8ocFBkARr7w90jmBw@mail.gmail.com>
 <20240416050353.GI112498@black.fi.intel.com>
 <CA+Y6NJF6+s5zUZeaWtagpMt8Qu0a1oE+3re3c6EsppH+ZsuMRQ@mail.gmail.com>
 <20240419044945.GR112498@black.fi.intel.com>
 <CA+Y6NJEpWpfPqHO6=Z1XFCXZDUq1+g6EFryB+Urq1=h0PhT+fg@mail.gmail.com>
 <7d68a112-0f48-46bf-9f6d-d99b88828761@amd.com>
 <20240423053312.GY112498@black.fi.intel.com>
 <7197b2ce-f815-48a1-a78e-9e139de796b7@amd.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <7197b2ce-f815-48a1-a78e-9e139de796b7@amd.com>

On Tue, Apr 23, 2024 at 11:59:36AM -0500, Mario Limonciello wrote:
> On 4/23/2024 00:33, Mika Westerberg wrote:
> > On Mon, Apr 22, 2024 at 02:21:18PM -0500, Mario Limonciello wrote:
> > > On 4/22/2024 14:17, Esther Shimanovich wrote:
> > > > Thanks for the explanation! I still don't fully understand how that
> > > > would work for my use case.
> > > > 
> > > > Perhaps it would be better for me to describe the case I am trying to
> > > > protect against.
> > > > 
> > > > To rehash, this quirk was written for devices with discrete
> > > > Thunderbolt controllers.
> > > > 
> > > > For example,
> > > > CometLake_CPU -> AlpineRidge_Chip -> USB-C Port
> > > > This device has the ExternalFacingPort property in ACPI.
> > > > My quirk relabels the Alpine Ridge chip as "fixed" and
> > > > external-facing, so that devices attached to the USB-C port could be
> > > > labeled as "removable"
> > > > 
> > > > Let's say we have a TigerLake CPU, which has integrated
> > > > Thunderbolt/USB4 capabilities:
> > > > 
> > > > TigerLake_ThunderboltCPU -> USB-C Port
> > > > This device also has the ExternalFacingPort property in ACPI and lacks
> > > > the usb4-host-interface property in the ACPI.
> > > > 
> > > > My worry is that someone could take an Alpine Ridge Chip Thunderbolt
> > > > Dock and attach it to the TigerLake CPU
> > > > 
> > > > TigerLake_ThunderboltCPU -> USB-C Port -> AlpineRidge_Dock
> > > > 
> > > > If that were to happen, this quirk would incorrectly label the Alpine
> > > > Ridge Dock as "fixed" instead of "removable".
> > > > 
> > > > My thinking was that we could prevent this scenario from occurring if
> > > > we filtered this quirk not to apply on CPU's like Tiger Lake, with
> > > > integrated Thunderbolt/USB4 capabilities.
> > > > 
> > > > ExternalFacingPort is found both on the Comet Lake ACPI and also on
> > > > the Tiger Lake ACPI. So I can't use that to distinguish between CPUs
> > > > which don't have integrated Thunderbolt, like Comet Lake, and CPUs
> > > > with integrated Thunderbolt, like Tiger Lake.
> > > > 
> > > > I am looking for something that can tell me if the device's Root Port
> > > > has the Thunderbolt controller upstream to it or not.
> > > > Is there anything like that?
> > > > Or perhaps should I add a check which compares the name of the
> > > > device's CPU with a list of CPUs that this quirk can be applied to?
> > > > Or is there some way I can identify the Thunderbolt controller, then
> > > > determine if it's upstream or downstream from the root port?
> > > > Or are Alpine Ridge docks not something to worry about at all?
> > > 
> > > My thought was once you have a device as untrusted, everything else
> > > connected to it should "also" be untrusted.
> > 
> > I think what you are looking for is that anything behind a PCIe tunnel
> > should not have this applied. IIRC the AMD GPU or some code there were
> > going to add identification of "virtual" links to the bandwidth
> > calculation functionality.
> > 
> > @Mario, do you remember if this was done already and if that could maybe
> > be re-used here?
> 
> Yeah there was a series that I worked on a few spins a while back
> specifically in the context of eGPUs to identify virtual links and take them
> out of bandwidth calculations.
> 
> It didn't get merged, I recall it got stalled on various feedback and I
> didn't dust it off because the series also did prompt discussions about the
> reasoning that amdgpu was doing this in the first place.  It turned out to
> be a bad assumption in the code and I instead made a change to amdgpu to not
> look at the whole topology but just the link partner
> (466a7d115326ece682c2b60d1c77d1d0b9010b4f if anyone is curious).

Okay that makes sense. Thanks!