Terry Bowman wrote:
> CXL and AER drivers need the ability to identify CXL devices and CXL port
> devices.
>
> First, add set_pcie_cxl() with logic checking for CXL Flexbus DVSEC
> presence. The CXL Flexbus DVSEC presence is used because it is required
> for all the CXL PCIe devices.[1]
>
> Add boolean 'struct pci_dev::is_cxl' with the purpose to cache the CXL
> Flexbus presence.
>
> Add pcie_is_cxl() as a macro to return 'struct pci_dev::is_cxl'.
>
> Add pcie_is_cxl_port() to check if a device is a CXL Root Port, CXL
> Upstream Switch Port, or CXL Downstream Switch Port. Also, verify the
> CXL Extensions DVSEC for Ports is present.[1]
>
> [1] CXL 3.1 Spec, 8.1.1 PCIe Designated Vendor-Specific Extended
> Capability (DVSEC) ID Assignment, Table 8-2
>
> Signed-off-by: Terry Bowman <terry.bowman@xxxxxxx>
> Reviewed-by: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx>
> Reviewed-by: Dave Jiang <dave.jiang@xxxxxxxxx>
> Reviewed-by: Fan Ni <fan.ni@xxxxxxxxxxx>
> ---
> drivers/pci/pci.c | 13 +++++++++++++
> drivers/pci/probe.c | 10 ++++++++++
> include/linux/pci.h | 5 +++++
> include/uapi/linux/pci_regs.h | 3 ++-
> 4 files changed, 30 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index 869d204a70a3..a2d8b41dd043 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -5032,6 +5032,19 @@ static u16 cxl_port_dvsec(struct pci_dev *dev)
> PCI_DVSEC_CXL_PORT);
> }
>
> +inline bool pcie_is_cxl(struct pci_dev *pci_dev)
> +{
> + return pci_dev->is_cxl;
> +}
> +
> +bool pcie_is_cxl_port(struct pci_dev *dev)
> +{
> + if (!pcie_is_cxl(dev))
> + return false;
> +
> + return (cxl_port_dvsec(dev) > 0);
At first I was concerned that this adds a capability list walk during
error handling, but patch 17 takes pcie_is_cxl_port() out of the
handles_cxl_errors() path.
It is still used in the aer_probe() path which means enumeration can
potentially race a CXL link up event.
I think this is fine for now because the CXL core has the same top-down
vs bottom-up race, and the CXL SBR code also shares the same race
problem.
A follow-on change needs to arrange for cxl_port_probe() to
enable/disable internal errors, because that path knows that a link has
been negotiated with an endpoint and that the CXL link details should be
stable.
> +}
> +
> static bool cxl_sbr_masked(struct pci_dev *dev)
> {
> u16 dvsec, reg;
> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index b6536ed599c3..7737b9ce7a83 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -1676,6 +1676,14 @@ static void set_pcie_thunderbolt(struct pci_dev *dev)
> dev->is_thunderbolt = 1;
> }
>
> +static void set_pcie_cxl(struct pci_dev *dev)
> +{
> + u16 dvsec = pci_find_dvsec_capability(dev, PCI_VENDOR_ID_CXL,
> + PCI_DVSEC_CXL_FLEXBUS);
> + if (dvsec)
> + dev->is_cxl = 1;
> +}
Similar race problem here as it is premature to check for this DVSEC on
disconnected ports.
For now, lets add a comment to include/uapi/linux/pci_regs.h along the
lines of:
diff --git a/include/uapi/linux/pci_regs.h b/include/uapi/linux/pci_regs.h
index 3445c4970e4d..32df7abdd23c 100644
--- a/include/uapi/linux/pci_regs.h
+++ b/include/uapi/linux/pci_regs.h
@@ -1208,7 +1208,13 @@
#define PCI_DOE_DATA_OBJECT_DISC_RSP_3_PROTOCOL 0x00ff0000
#define PCI_DOE_DATA_OBJECT_DISC_RSP_3_NEXT_INDEX 0xff000000
-/* Compute Express Link (CXL r3.1, sec 8.1.5) */
+/*
+ * Compute Express Link (CXL r3.1, sec 8.1)
+ *
+ * Note that CXL DVSEC id 3 and 7 to be ignored when the CXL link state
+ * is "disconnected" (CXL r3.1, sec 9.12.3). Re-enumerate these
+ * registers on downstream link-up events.
+ */
#define PCI_DVSEC_CXL_PORT 3
#define PCI_DVSEC_CXL_PORT_CTL 0x0c
#define PCI_DVSEC_CXL_PORT_CTL_UNMASK_SBR 0x00000001
...to at least remind our future selves that there is work to do here to
make the implementation robust against hot-plug scenarios.
With that you can add:
Reviewed-by: Dan Williams <dan.j.williams@xxxxxxxxx>
