On Tue, Sep 02, 2025 at 12:48:28PM -0500, Bjorn Helgaas wrote: > On Fri, Jun 20, 2025 at 11:55:06PM +0800, Hans Zhang wrote: > > Current PCIe initialization logic may leave root ports operating with > > non-optimal Maximum Payload Size (MPS) settings. While downstream device > > configuration is handled during bus enumeration, root port MPS values > > inherited from firmware or hardware defaults ... > > Apparently Root Port MPS configuration is different from that for > downstream devices? pci_host_probe() will call pci_scan_root_bus_bridge(), which will call pci_scan_single_device(), which will call pci_device_add(), which will call pci_configure_device(), which will call pci_configure_mps(). This will be done for both bridges and endpoints. The bridge will be scanned/added first, before devices behind the bridge. While pci_configure_device()/pci_configure_mps() will be called for both bridges and endpoints, pci_configure_mps() will do an early return for devices where pci_upstream_bridge() returns NULL, i.e. for devices where that does not have an upstream bridge, i.e. for the root bridge itself: https://github.com/torvalds/linux/blob/v6.17-rc4/drivers/pci/probe.c#L2181-L2182 So MPS will not be touched for root bridges. This patch ensures that MPS for root bridges gets initialized to MPSS (Max supported MPS). Later, when pci_configure_device()/pci_configure_mps() is called for a device behind the bridge, if the MPSS of the device behind the bridge is smaller than the MPS of the bridge, the code reduces the MPS of the bridge: https://github.com/torvalds/linux/blob/v6.17-rc4/drivers/pci/probe.c#L2205 My only question with this patch is if there is a bridge behind a bridge, will the bridge behind the bridge still have pci_pcie_type() == PCI_EXP_TYPE_ROOT_PORT ? If so, perhaps we should modify this patch from: + if (pci_pcie_type(dev) == PCI_EXP_TYPE_ROOT_PORT && + pcie_bus_config != PCIE_BUS_TUNE_OFF) { + pcie_write_mps(dev, 128 << dev->pcie_mpss); + } + if (!bridge || !pci_is_pcie(bridge)) return; to: + if (!bridge && pci_pcie_type(dev) == PCI_EXP_TYPE_ROOT_PORT && + pcie_bus_config != PCIE_BUS_TUNE_OFF) { + pcie_write_mps(dev, 128 << dev->pcie_mpss); + } + if (!bridge || !pci_is_pcie(bridge)) return; > > During host controller probing phase, when PCIe bus tuning is enabled, > > the implementation now configures root port MPS settings to their > > hardware-supported maximum values. Specifically, when configuring the MPS > > for a PCIe device, if the device is a root port and the bus tuning is not > > disabled (PCIE_BUS_TUNE_OFF), the MPS is set to 128 << dev->pcie_mpss to > > match the Root Port's maximum supported payload size. The Max Read Request > > Size (MRRS) is subsequently adjusted through existing companion logic to > > maintain compatibility with PCIe specifications. > > > > Note that this initial setting of the root port MPS to the maximum might > > be reduced later during the enumeration of downstream devices if any of > > those devices do not support the maximum MPS of the root port. > > > > Explicit initialization at host probing stage ensures consistent PCIe > > topology configuration before downstream devices perform their own MPS > > negotiations. This proactive approach addresses platform-specific > > requirements where controller drivers depend on properly initialized > > root port settings, while maintaining backward compatibility through > > PCIE_BUS_TUNE_OFF conditional checks. Hardware capabilities are fully > > utilized without altering existing device negotiation behaviors. > > This last paragraph seems kind of like marketing without any real > content. Is there something important in there? > > Nits: > s/root port/Root Port/ > > Reword "implementation now configures" to be clear about whether "now" > refers to before this patch or after. > > Update the MRRS "to maintain compatibility" part. I'm dubious about > there being a spec compatibility issue with respect to MRRS. Cite the > relevant section if there is an issue. I'm not sure why the commit message mentions MRRS at all. Sure, pcie_write_mrrs() might set MRRS to MPS, but that is existing logic and not really related to the change in this patch IMO. Kind regards, Niklas
