Re: [PATCH] PCI: Skip the CXL state update for a VF

From: Dave Jiang

Date: Thu Nov 19 2026 - 11:50:58 EST




On 11/19/26 9:42 AM, Joshua Machauer wrote:
> set_pcie_cxl() runs for every device that is added. It first updates
> the CXL state of the upstream bridge, which walks that bridge's
> extended capabilities and recurses up to the root port. It then walks
> the device's own extended capabilities for the Flex Bus Port DVSEC.
>
> For a VF neither walk can change anything. A CXL device exposes its
> Flex Bus Port DVSEC through Function 0 (CXL 4.0, section 8.1.8), so
> the search in a VF finds nothing and the VF's is_cxl stays false.
> pci_upstream_bridge() resolves a VF through its PF, so the bridge is
> the one that was already updated when the PF was added.
>
> Enabling SR-IOV adds the VFs one after another, and each add repeats
> both walks against the same bridges.
>
> On an NVIDIA Vera system, enabling 256 VFs on a PF whose upstream
> bridge has 18 extended capabilities takes 296 ms instead of 402 ms.
> Enabling 32 VFs on a Rubin GPU takes 218 ms instead of 235 ms. The
> times are for the write to sriov_numvfs, with the PF bound to
> pci-pf-stub and VF driver autoprobe disabled, so no VF driver probes.
>
> Return early for a VF. Every device keeps the is_cxl value it gets
> today.
>
> Signed-off-by: Joshua Machauer <machaujo@xxxxxxxxx>

commit log a bit wordy, otherwise LGTM

Reviewed-by: Dave Jiang <dave.jiang@xxxxxxxxx>

> ---
> drivers/pci/probe.c | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index 9a178e0e5f7f..34acebd1c706 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -1717,6 +1717,14 @@ static void set_pcie_cxl(struct pci_dev *dev)
> if (!pci_is_pcie(dev))
> return;
>
> + /*
> + * A CXL device exposes its Flex Bus Port DVSEC through Function 0,
> + * so a VF has none. Its upstream bridge is the PF's, which was
> + * updated when the PF was added.
> + */
> + if (dev->is_virtfn)
> + return;
> +
> /*
> * Update parent's CXL state because alternate protocol training
> * may have changed