On Tue, Oct 25, 2022 at 12:07:47AM -0500, Tyler Hicks wrote:
> On 2022-10-20 15:24:37, Bjorn Helgaas wrote:
> > On Wed, Oct 19, 2022 at 01:25:59PM -0500, Tyler Hicks wrote:
> > > On 2022-06-10 23:01:31, Zhiqiang Hou wrote:
> > > > From: Hou Zhiqiang <Zhiqiang.Hou@xxxxxxx>
> > > >
> > > > The commit 27d868b5e6cf ("PCI: Set MPS to match upstream bridge")
> > > > made the device's MPS matches upstream bridge for PCIE_BUS_DEFAULT
> > > > mode, so that it's more likely that a hot-added device will work in
> > > > a system with an optimized MPS configuration.
> > > >
> > > > Obviously, the Linux itself optimizes the MPS settings in the
> > > > PCIE_BUS_SAFE and PCIE_BUS_PERFORMANCE mode, so let's do this also
> > > > for these modes.
> > > >
> > > > Signed-off-by: Hou Zhiqiang <Zhiqiang.Hou@xxxxxxx>
> > >
> > > I wanted to give a little more information about the issue we're seeing.
> > > We're having trouble retaining the optimized Max Payload Size (MPS)
> > > value of a PCIe endpoint after hotplug/rescan. In this case,
> > > `pcie_ports=native` and `pci=pcie_bus_safe` are set on the cmdline and
> > > we expect the Linux kernel to retain the MPS value. Commit 27d868b5e6cf
> > > preserved the MPS value when using the default PCIe bus mode
> > > (PCIE_BUS_DEFAULT) and we're hopeful that this can be extended to other
> > > modes, as well.
> >
> > Thanks, Tyler. I need help understanding what's going on here. An
> > example of the topology and what happens and what *should* happen
> > might help.
>
> Hey Bjorn and Keith! Thanks for both of your responses and your
> patience. They spurred good investigations on my side and I'm learning a
> lot as I go.
>
> > Some MPS configuration is done per-device in pci_configure_mps(), and
> > some considers a hierarchy in pcie_bus_configure_settings(). In the
> > current tree, in the PCIE_BUS_SAFE case:
> >
> > - pci_configure_mps() does nothing (except for RCiEPs).
> >
> > - pcie_bus_configure_settings(bus) looks at the hierarchy rooted at
> > the bridge leading to "bus". If the hierarchy contains a hotplug
> > Switch Downstream Port, it sets MPS and MRRS to 128 for
> > everything.
> >
> > If it does not contain such a bridge, it finds the smallest
> > MPS_Supported ("smpss") of any device in the hierarchy and sets
> > MPS and MRRS to that for everything.
> >
> > If you boot with a hotplug Root Port leading to an empty slot, I think
> > the RP MPS will end up being whatever BIOS put there.
>
> I've been talking to the firmware folks on why SAFE mode was selected,
> based on Keith's question from Wednesday. From what I've been told,
> U-Boot doesn't seed the RP MPS with a value so the kernel sees a value
> of 128 for p_mps in pci_configure_mps() when using the DEFAULT mode.
> Apparently UEFI does seed the RP MPS but we don't get that with U-Boot.
> Therefore, SAFE mode was selected to get an initial, tuned RP MPS value
> set to 256.
Are there any devices below the RP at the time we set MPS=256?
> > A subsequent hot-add will do nothing in pci_configure_mps(), and
> > pcie_bus_configure_settings() looks like it would set the RP and EP
> > MPS to the minimum of the RP and EP MPS_Supported.
> >
> > Is that what you see? What would you like to see instead?
>
> No, not quite. After hot-add, we see the EP MPS set to 128 with SAFE
> mode even though the RP MPS is 256.
Can you add the relevant topology here so we can work through the
concrete details? Is the endpoint hot-added directly below a Root
Port? Or is there a switch involved? What are the MPS_Supported
values for all the devices? If there's a switch in the picture, it
looks like we currently restrict to 128, I think because it's possible
an endpoint that can only do 128 may be added.
Bjorn
