On Mon, 12 Feb 2024 at 15:16, Manivannan Sadhasivam
<manivannan.sadhasivam@xxxxxxxxxx> wrote:
>
> On Fri, Feb 09, 2024 at 12:56:18PM +0200, Dmitry Baryshkov wrote:
> > On Fri, 9 Feb 2024 at 09:57, Manivannan Sadhasivam
> > <manivannan.sadhasivam@xxxxxxxxxx> wrote:
> > >
> > > On Fri, Feb 09, 2024 at 12:58:15PM +0530, Krishna Chaitanya Chundru wrote:
> > > >
> > > >
> > > > On 2/8/2024 8:49 PM, Dmitry Baryshkov wrote:
> > > > > On Thu, 8 Feb 2024 at 16:58, Krishna Chaitanya Chundru
> > > > > <quic_krichai@xxxxxxxxxxx> wrote:
> > > > > > On 2/8/2024 12:21 PM, Dmitry Baryshkov wrote:
> > > > > > > On Thu, 8 Feb 2024 at 08:14, Krishna Chaitanya Chundru
> > > > > > > <quic_krichai@xxxxxxxxxxx> wrote:
> > > > > > > >
> > > > > > > >
> > > > > > > >
> > > > > > > > On 2/7/2024 5:17 PM, Dmitry Baryshkov wrote:
> > > > > > > > > On Wed, 7 Feb 2024 at 12:42, Krishna chaitanya chundru
> > > > > > > > > <quic_krichai@xxxxxxxxxxx> wrote:
> > > > > > > > > >
> > > > > > > > > > Enable PCIe1 controller and its corresponding PHY nodes on
> > > > > > > > > > qcs6490-rb3g2 platform.
> > > > > > > > > >
> > > > > > > > > > PCIe switch is connected to PCIe1, PCIe switch has multiple endpoints
> > > > > > > > > > connected. For each endpoint a unique BDF will be assigned and should
> > > > > > > > > > assign unique smmu id. So for each BDF add smmu id.
> > > > > > > > > >
> > > > > > > > > > Signed-off-by: Krishna chaitanya chundru <quic_krichai@xxxxxxxxxxx>
> > > > > > > > > > ---
> > > > > > > > > > arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts | 42 ++++++++++++++++++++++++++++
> > > > > > > > > > 1 file changed, 42 insertions(+)
> > > > > > > > > >
> > > > > > > > > > diff --git a/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts b/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts
> > > > > > > > > > index 8bb7d13d85f6..0082a3399453 100644
> > > > > > > > > > --- a/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts
> > > > > > > > > > +++ b/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts
> > > > > > > > > > @@ -413,6 +413,32 @@ vreg_bob_3p296: bob {
> > > > > > > > > > };
> > > > > > > > > > };
> > > > > > > > > >
> > > > > > > > > > +&pcie1 {
> > > > > > > > > > + perst-gpios = <&tlmm 2 GPIO_ACTIVE_LOW>;
> > > > > > > > > > +
> > > > > > > > > > + pinctrl-0 = <&pcie1_reset_n>, <&pcie1_wake_n>;
> > > > > > > > > > + pinctrl-names = "default";
> > > > > > > > > > +
> > > > > > > > > > + iommu-map = <0x0 &apps_smmu 0x1c80 0x1>,
> > > > > > > > > > + <0x100 &apps_smmu 0x1c81 0x1>,
> > > > > > > > > > + <0x208 &apps_smmu 0x1c84 0x1>,
> > > > > > > > > > + <0x210 &apps_smmu 0x1c85 0x1>,
> > > > > > > > > > + <0x218 &apps_smmu 0x1c86 0x1>,
> > > > > > > > > > + <0x300 &apps_smmu 0x1c87 0x1>,
> > > > > > > > > > + <0x400 &apps_smmu 0x1c88 0x1>,
> > > > > > > > > > + <0x500 &apps_smmu 0x1c89 0x1>,
> > > > > > > > > > + <0x501 &apps_smmu 0x1c90 0x1>;
> > > > > > > > >
> > > > > > > > > Is the iommu-map really board specific?
> > > > > > > > >
> > > > > > > > The iommu-map for PCIe varies if PCIe switch is connected.
> > > > > > > > For this platform a PCIe switch is connected and for that reason
> > > > > > > > we need to define additional smmu ID's for each BDF.
> > > > > > > >
> > > > > > > > For that reason we defined here as these ID's are applicable only
> > > > > > > > for this board.
> > > > > > >
> > > > > > > So, these IDs are the same for all boards, just being unused on
> > > > > > > devices which have no bridges / switches connected to this PCIe host.
> > > > > > > If this is correct, please move them to sc7280.dtsi.
> > > > > > >
> > > > > > Yes ID's will be same for all boards. we can move them sc7280.dtsi
> > > > > > but the BDF to smmu mapping will be specific to this board only.
> > > > > > if there is some other PCIe switch with different configuration is
> > > > > > connected to different board of same variant in future again these
> > > > > > mapping needs to updated.
> > > > >
> > > > > Could you possibly clarify this? Are they assigned one at a time
> > > > > manually? Or is it somehow handled by the board's TZ code, which
> > > > > assigns them sequentially to the known endpoints? And is it done via
> > > > > probing the link or via some static configuration?
> > > >
> > > > There is no assignment of SID's in TZ for PCIe.
> > > > PCIe controller has BDF to SID mapping table which we need to
> > > > program with the iommu map table.
> > > >
> > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/pci/controller/dwc/pcie-qcom.c?h=v6.8-rc3#n997
> > > >
> > > > Based upon switch the BDF to SID table will change for example I had two
> > > > switches with one switch has 2 PCIe ports and other has 3 ports one
> > > > embedded port which supports multiple functions.
> > > >
> > > > For the first switch the BDF's are
> > > > - 0x000(root complex),
> > > > - 0x100(USP),
> > > > - 0x208(DSP 0),
> > > > - 0x210(DSP 1),
> > > > - 0x300(endpoint connected to DSP 0),
> > > > - 0x400( endpoint connected to DSP 1).
> > > >
> > > > For 2nd switch the BDF's are
> > > > - 0x000(root complex),
> > > > - 0x100(USP),
> > > > - 0x208(embeeded DSP 0),
> > > > - 0x210(DSP 1),
> > > > - 0x218 (DSP 2),
> > > > - 0x300(embedded endpoint function 0),
> > > > - 0x301 (embedded endpoint function 1)
> > > > - 0x400( endpoint connected to DSP 1)
> > > > - 0x500(endpoint connected to DSP2).
> > > >
> > > > For these two switches we need different BDF to SID table so for that
> > > > reason we are keeping iommu map here as this is specific to this board.
> > > >
> > >
> > > I don't understand why the SID table has to change between PCIe devices. The SID
> > > mapping should be part of the SoC dtsi, where a single SID would be defined for
> > > the devices under a bus. And all the devices under the bus have to use the same
> > > SID.
> >
> > This sounds like a sane default, indeed. Nevertheless, I see a point
> > in having per-device-SID assignment. This increases isolation and can
> > potentially prevent security issues. However in such case SID
> > assignment should be handled in some automagic way. In other words,
> > there must be no need to duplicate the topology of the PCIe bus in the
> > iommu-maps property.
> >
>
> Yes, address space isolation is the primary motive behind this patch. But as
> you said, we should not do it by hardcoding the SIDs in the board DTS. It won't
> scale and is not a proper solution.
>
> Instead, the issue should be addressed in the IOMMU layer by working with the
> IOMMU folks.
>
> It should be noted that we _cannot_ use any arbitrary SID for PCIe bus. HYP/TZ
> will fault the transactions coming with different SIDs than the ones assigned
> to them. So we still need to pass that info from DT to IOMMU layer.
Yes, passing a range or a masked value sounds logical. Passing 1:1
mapping for a dynamic bus doesn't.
--
With best wishes
Dmitry
