[+to Alex, +cc kvm]
On Fri, Nov 15, 2024 at 03:22:02PM +0800, Jian-Hong Pan wrote:
> PCI devices' parameters on the VMD bus have been programmed properly
> originally. But, cleared after pci_reset_bus() and have not been restored
> correctly. This leads the link's L1.2 between PCIe Root Port and child
> device gets wrong configs.
17423360a27a ("PCI/ASPM: Save L1 PM Substates Capability for
suspend/resume") appeared in v6.9 and added a bug in the save/restore
path: we save the device L1SS state, but restore both the device L1SS
state and the parent's L1SS state (when the parent state may be junk)
afterwards.
Jian-Hong sees this on VMD, which makes sense because vmd uses
pci_reset_bus() when enabling the VMD domain.
But this should be a problem for any device reset, so I added you,
Alex, in case you've seen this with the resets VFIO does?
We also save/restore for suspend, but I suppose we don't notice the
problem there because in that case we save state for *all* devices,
so the parent state should be valid when we restore.
> Here is a failed example on ASUS B1400CEAE with enabled VMD. Both PCIe
> bridge and NVMe device should have the same LTR1.2_Threshold value.
> However, they are configured as different values in this case:
>
> 10000:e0:06.0 PCI bridge [0604]: Intel Corporation 11th Gen Core Processor PCIe Controller [8086:9a09] (rev 01) (prog-if 00 [Normal decode])
> ...
> Capabilities: [200 v1] L1 PM Substates
> L1SubCtl1: PCI-PM_L1.2- PCI-PM_L1.1- ASPM_L1.2+ ASPM_L1.1-
> T_CommonMode=0us LTR1.2_Threshold=0ns
> L1SubCtl2: T_PwrOn=0us
>
> 10000:e1:00.0 Non-Volatile memory controller [0108]: Sandisk Corp WD Blue SN550 NVMe SSD [15b7:5009] (rev 01) (prog-if 02 [NVM Express])
> ...
> Capabilities: [900 v1] L1 PM Substates
> L1SubCtl1: PCI-PM_L1.2- PCI-PM_L1.1- ASPM_L1.2+ ASPM_L1.1-
> T_CommonMode=0us LTR1.2_Threshold=101376ns
> L1SubCtl2: T_PwrOn=50us
I think T_PwrOn should also be the same for both devices, FWIW.
In fact, I think L1SS should be configured identically for both ends
of the link, with the exceptions of Link Activation and
Common_Mode_Restore_Time, which are RsvdP for the Upstream Port.
> Here is VMD mapped PCI device tree:
>
> -+-[0000:00]-+-00.0 Intel Corporation Device 9a04
> | ...
> \-[10000:e0]-+-06.0-[e1]----00.0 Sandisk Corp WD Blue SN550 NVMe SSD
> \-17.0 Intel Corporation Tiger Lake-LP SATA Controller
>
> When pci_reset_bus() resets the bus [e1] of the NVMe, it only saves and
> restores NVMe's state before and after reset. Then, when it restores the
> NVMe's state, ASPM code restores L1SS for both the parent bridge and the
> NVMe in pci_restore_aspm_l1ss_state(). The NVMe's L1SS is restored
> correctly. But, the parent bridge's L1SS is restored with a wrong value 0x0
> because the parent bridge's L1SS wasn't saved by pci_save_aspm_l1ss_state()
> before reset.
>
> To avoid pci_restore_aspm_l1ss_state() restore wrong value to the parent's
> L1SS config like this example, make pci_save_aspm_l1ss_state() save the
> parent's L1SS config, if the PCI device has a parent.
>
> Link: https://lore.kernel.org/linux-pci/CAPpJ_eexU0gCHMbXw_z924WxXw0+B6SdS4eG9oGpEX1wmnMLkQ@xxxxxxxxxxxxxx/
> Closes: https://bugzilla.kernel.org/show_bug.cgi?id=218394
> Fixes: 17423360a27a ("PCI/ASPM: Save L1 PM Substates Capability for suspend/resume")
> Suggested-by: Ilpo Järvinen <ilpo.jarvinen@xxxxxxxxxxxxxxx>
> Signed-off-by: Jian-Hong Pan <jhp@xxxxxxxxxxxxx>
> Reviewed-by: Ilpo Järvinen <ilpo.jarvinen@xxxxxxxxxxxxxxx>
> Reviewed-by: David E. Box <david.e.box@xxxxxxxxxxxxxxx>
> ---
> v9:
> - Drop the v8 fix about drivers/pci/pcie/aspm.c. Use this in VMD instead.
>
> v10:
> - Drop the v9 fix about drivers/pci/controller/vmd.c
> - Fix in PCIe ASPM to make it symmetric between pci_save_aspm_l1ss_state()
> and pci_restore_aspm_l1ss_state()
>
> v11:
> - Introduce __pci_save_aspm_l1ss_state as a resusable helper function
> which is same as the original pci_configure_aspm_l1ss
> - Make pci_save_aspm_l1ss_state invoke __pci_save_aspm_l1ss_state for
> both child and parent devices
> - Smooth the commit message
>
> v12:
> - Update the commit message
>
> v13:
> - Tweak the commit message to make it more like a general fix
> - When pci_alloc_dev() prepares the pci_dev, it sets the pci_dev's bus.
> So, let pci_save_aspm_l1ss_state() access pdev's bus directly.
> - Add comment in pci_save_aspm_l1ss_state() to describe why it does not
> save both the PCIe device and the parent's L1SS config like
> pci_restore_aspm_l1ss_state() directly.
>
> drivers/pci/pcie/aspm.c | 18 +++++++++++++++++-
> 1 file changed, 17 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
> index 28567d457613..0bcd060aab32 100644
> --- a/drivers/pci/pcie/aspm.c
> +++ b/drivers/pci/pcie/aspm.c
> @@ -79,7 +79,7 @@ void pci_configure_aspm_l1ss(struct pci_dev *pdev)
> ERR_PTR(rc));
> }
>
> -void pci_save_aspm_l1ss_state(struct pci_dev *pdev)
> +static void __pci_save_aspm_l1ss_state(struct pci_dev *pdev)
> {
> struct pci_cap_saved_state *save_state;
> u16 l1ss = pdev->l1ss;
> @@ -101,6 +101,22 @@ void pci_save_aspm_l1ss_state(struct pci_dev *pdev)
> pci_read_config_dword(pdev, l1ss + PCI_L1SS_CTL1, cap++);
> }
>
> +void pci_save_aspm_l1ss_state(struct pci_dev *pdev)
> +{
> + struct pci_dev *parent = pdev->bus->self;
> +
> + __pci_save_aspm_l1ss_state(pdev);
> +
> + /*
> + * Save parent's L1 substate configuration, if the parent has not saved
> + * state. It avoids pci_restore_aspm_l1ss_state() restore wrong value to
> + * parent's L1 substate configuration. However, the parent might be
> + * nothing, if pdev is a PCI bridge.
> + */
> + if (parent && !parent->state_saved)
> + __pci_save_aspm_l1ss_state(parent);
> +}
> +
> void pci_restore_aspm_l1ss_state(struct pci_dev *pdev)
> {
> struct pci_cap_saved_state *pl_save_state, *cl_save_state;
> --
> 2.47.0
>
