Re: [PATCH v2] PCI: quirks: Add warm reset quirk for IBM Spyre accelerator

[Bjorn Helgaas <helgaas@kernel.org>]

On Fri, Aug 21, 2026 at 06:04:01PM +0530, Avinash Roy wrote:
> From: Shivaprasad G Bhat <sbhat@linux.ibm.com>
> 
> The IBM Spyre accelerator (PCI ID 1014:06a7) requires a dedicated warm
> reset method for reliable EEH recovery on pSeries/ppc64 platforms.

It sounds like if you plugged a Spyre into an x86/arm64/etc platform,
it would still need a dedicated warm reset quirk because the issue is
with Spyre, not with pSeries.

> FLR alone is insufficient because it does not re-initialize firmware
> state machine for the current version of the card. The existing API
> pci_set_pcie_reset_state() is used to drive the platform-level
> PERST# using the EEH interfaces by adding a device specific quirk.
> 
> The reset function is placed in arch/powerpc/platforms/pseries/pci.c
> where the EEH infrastructure is already present, declared in
> pseries.h, and registered in the generic pci_dev_reset_methods[] table
> in drivers/pci/quirks.c under a CONFIG_PPC_PSERIES guard.
> 
> This reset method is only registered when CONFIG_PPC_PSERIES is enabled,
> so it is not available on other platforms.
> 
> Reset sequence:
>   - Assert PERST# via pcie_warm_reset, hold 250 ms
>   - Deassert PERST# via pcie_deassert_reset
>   - Wait 250 ms for PCIe link retrain 	
> 
> Signed-off-by: Avinash Roy <avinash.roy1@linux.ibm.com>
> Signed-off-by: Shivaprasad G Bhat <sbhat@linux.ibm.com>
> 
> ---
> Changelog
> v1: https://lore.kernel.org/linux-pci/20260805125100.1-1-avinash.roy1@linux.ibm.com/
> 
>    Reworded the commit message to highlight the warm reset requirement being
>    a need of the current card and not a platform deficiency.
> 
> diff --git a/arch/powerpc/platforms/pseries/pci.c b/arch/powerpc/platforms/pseries/pci.c
> index 84e4ffe957a8..7aa58d57208f 100644
> --- a/arch/powerpc/platforms/pseries/pci.c
> +++ b/arch/powerpc/platforms/pseries/pci.c
> @@ -11,6 +11,7 @@
>  #include <linux/kernel.h>
>  #include <linux/pci.h>
>  #include <linux/string.h>
> +#include <linux/delay.h>
>  
>  #include <asm/eeh.h>
>  #include <asm/pci-bridge.h>
> @@ -236,6 +237,41 @@ static void fixup_winbond_82c105(struct pci_dev* dev)
>  DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_WINBOND, PCI_DEVICE_ID_WINBOND_82C105,
>  			 fixup_winbond_82c105);
>  
> +/*
> + * Reset Spyre adapter using pci_set_pcie_reset_state()
> + * This is specifically for PPC platforms where EEH (Enhanced Error Handling)
> + * requires this reset method for proper device recovery.

This comment still suggests that the issue is related to the platform,
not the adapter itself.  I know reset_spyre() will only *work* on
pSeries because pcibios_set_pcie_reset_state() is only implemented
there.

> + */
> +int reset_spyre(struct pci_dev *dev, bool probe)
> +{
> +	int ret;
> +
> +	if (probe)
> +		return 0;
> +
> +	/* Assert warm reset */
> +	ret = pci_set_pcie_reset_state(dev, pcie_warm_reset);
> +	if (ret) {
> +		pci_err(dev, "Failed to assert reset: %d\n", ret);
> +		return ret;
> +	}
> +
> +	/* Wait for reset to take effect */
> +	msleep(250);
> +
> +	/* Deassert reset */
> +	ret = pci_set_pcie_reset_state(dev, pcie_deassert_reset);
> +	if (ret) {
> +		pci_err(dev, "Failed to deassert reset: %d\n", ret);
> +		return ret;
> +	}
> +
> +	/* Wait for device to recover */
> +	msleep(250);
> +
> +	return 0;
> +}
> +
>  static enum pci_bus_speed prop_to_pci_speed(u32 prop)
>  {
>  	switch (prop) {
> diff --git a/arch/powerpc/platforms/pseries/pseries.h b/arch/powerpc/platforms/pseries/pseries.h
> index 3968a6970fa8..79c194aeb3c6 100644
> --- a/arch/powerpc/platforms/pseries/pseries.h
> +++ b/arch/powerpc/platforms/pseries/pseries.h
> @@ -92,6 +92,10 @@ extern struct pci_controller_ops pseries_pci_controller_ops;
>  int pseries_msi_allocate_domains(struct pci_controller *phb);
>  void pseries_msi_free_domains(struct pci_controller *phb);
>  
> +/* PCI device reset method for Spyre adapter */
> +struct pci_dev;
> +int reset_spyre(struct pci_dev *dev, bool probe);
> +
>  extern int CMO_PrPSP;
>  extern int CMO_SecPSP;
>  extern unsigned long CMO_PageSize;
> diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
> index b09f27f7846f..1b8ab5d9764b 100644
> --- a/drivers/pci/quirks.c
> +++ b/drivers/pci/quirks.c
> @@ -4237,6 +4237,11 @@ static int reset_hinic_vf_dev(struct pci_dev *pdev, bool probe)
>  	return 0;
>  }
>  
> +#ifdef CONFIG_PPC_PSERIES
> +/* Defined in arch/powerpc/platforms/pseries/pci.c */
> +int reset_spyre(struct pci_dev *dev, bool probe);
> +#endif

Given that we need these ifdefs anyway, I think it would be simpler to
just include the reset_spyre() implementation here.  It doesn't look
like it needs anything pSeries-specific, since
pci_set_pcie_reset_state() is already available.

>  static const struct pci_dev_reset_methods pci_dev_reset_methods[] = {
>  	{ PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_82599_SFP_VF,
>  		 reset_intel_82599_sfp_virtfn },
> @@ -4252,6 +4257,10 @@ static const struct pci_dev_reset_methods pci_dev_reset_methods[] = {
>  		reset_chelsio_generic_dev },
>  	{ PCI_VENDOR_ID_HUAWEI, PCI_DEVICE_ID_HINIC_VF,
>  		reset_hinic_vf_dev },
> +#ifdef CONFIG_PPC_PSERIES
> +	{ PCI_VENDOR_ID_IBM, 0x06a7,
> +		reset_spyre},
> +#endif
>  	{ 0 }
>  };