Re: [PATCH v4 2/3] PCI: Add soft reset method as last resort

[Alex Williamson <alex@shazbot.org>]

On Mon, 18 May 2026 14:48:34 +0200
Jose Ignacio Tornos Martinez <jtornosm@redhat.com> wrote:

> Add a software-initiated "soft" reset method that attempts D3hot->D0
> transition as an absolute last resort when all other reset methods
> have failed.
> 
> Some devices incorrectly advertise NoSoftRst+ (blocking PM reset) but
> the D3hot transition does provide sufficient reset for certain use cases,
> particularly VFIO passthrough scenarios. This method provides a "better
> than nothing" option when the device would otherwise have no reset
> capability.
> 
> The method only becomes available when:
> - pci_pm_reset() is unavailable (typically blocked by NoSoftRst+)
> - pci_d3cold_reset() is unavailable (no platform _PR3 support)
> - Device has PM capability (required for D3hot transition)
> 
> Extract the D3hot transition logic into a shared helper function
> (pci_do_d3hot_transition) used by both pci_pm_reset and pci_soft_reset.
> 
> Reset hierarchy with this change:
> 1. device_specific
> 2. acpi
> 3. flr
> 4. af_flr
> 5. pm (proper method, checks NoSoftRst)
> 6. bus
> 7. cxl_bus
> 8. d3cold (requires _PR3)
> 9. soft (NEW - D3hot without NoSoftRst check, absolute last resort)
> 
> Signed-off-by: Jose Ignacio Tornos Martinez <jtornosm@redhat.com>
> ---
> v4: Implements D3hot transition as last resort when pm/d3cold unavailable
> v3: https://lore.kernel.org/all/20260513122349.268753-1-jtornosm@redhat.com/
> 
>  drivers/pci/pci.c   | 98 ++++++++++++++++++++++++++++++++++-----------
>  include/linux/pci.h |  2 +-
>  2 files changed, 76 insertions(+), 24 deletions(-)

NAK.  This cannot happen as a general case, it will cause vfio-pci to
report reset capabilities for essentially all devices, whether
validated or not.

The suggestion was that for devices where this has proven "better than
nothing", we could think about a device specific version of this,
matching devices IDs, not a fall-through for any device.

Given the "partial reset" nature of this, even on the target device, I
still wonder though whether userspace cannot already handle this by
forcing the power state through sysfs prior to assigning.  Thanks,

Alex 
 
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index 839903b59698..8dad386bd65d 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -4437,6 +4437,43 @@ static int pci_af_flr(struct pci_dev *dev, bool probe)
>  	return ret;
>  }
>  
> +/**
> + * pci_do_d3hot_transition - Perform D3hot->D0 power state transition
> + * @dev: Device to transition
> + *
> + * Common helper to perform D3hot->D0 transition for PM-based reset methods.
> + * Handles IOMMU preparation, state transition, and waiting for device ready.
> + */
> +static int pci_do_d3hot_transition(struct pci_dev *dev)
> +{
> +	u16 csr;
> +	int ret;
> +
> +	if (dev->current_state != PCI_D0)
> +		return -EINVAL;
> +
> +	ret = pci_dev_reset_iommu_prepare(dev);
> +	if (ret) {
> +		pci_err(dev, "failed to stop IOMMU for a PCI reset: %d\n", ret);
> +		return ret;
> +	}
> +
> +	pci_read_config_word(dev, dev->pm_cap + PCI_PM_CTRL, &csr);
> +	csr &= ~PCI_PM_CTRL_STATE_MASK;
> +	csr |= PCI_D3hot;
> +	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
> +	pci_dev_d3_sleep(dev);
> +
> +	csr &= ~PCI_PM_CTRL_STATE_MASK;
> +	csr |= PCI_D0;
> +	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
> +	pci_dev_d3_sleep(dev);
> +
> +	ret = pci_dev_wait(dev, "PM D3hot->D0", PCIE_RESET_READY_POLL_MS);
> +	pci_dev_reset_iommu_done(dev);
> +	return ret;
> +}
> +
>  /**
>   * pci_pm_reset - Put device into PCI_D3 and back into PCI_D0.
>   * @dev: Device to reset.
> @@ -4455,7 +4492,6 @@ static int pci_af_flr(struct pci_dev *dev, bool probe)
>  static int pci_pm_reset(struct pci_dev *dev, bool probe)
>  {
>  	u16 csr;
> -	int ret;
>  
>  	if (!dev->pm_cap || dev->dev_flags & PCI_DEV_FLAGS_NO_PM_RESET)
>  		return -ENOTTY;
> @@ -4467,28 +4503,7 @@ static int pci_pm_reset(struct pci_dev *dev, bool probe)
>  	if (probe)
>  		return 0;
>  
> -	if (dev->current_state != PCI_D0)
> -		return -EINVAL;
> -
> -	ret = pci_dev_reset_iommu_prepare(dev);
> -	if (ret) {
> -		pci_err(dev, "failed to stop IOMMU for a PCI reset: %d\n", ret);
> -		return ret;
> -	}
> -
> -	csr &= ~PCI_PM_CTRL_STATE_MASK;
> -	csr |= PCI_D3hot;
> -	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
> -	pci_dev_d3_sleep(dev);
> -
> -	csr &= ~PCI_PM_CTRL_STATE_MASK;
> -	csr |= PCI_D0;
> -	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
> -	pci_dev_d3_sleep(dev);
> -
> -	ret = pci_dev_wait(dev, "PM D3hot->D0", PCIE_RESET_READY_POLL_MS);
> -	pci_dev_reset_iommu_done(dev);
> -	return ret;
> +	return pci_do_d3hot_transition(dev);
>  }
>  
>  /**
> @@ -4530,6 +4545,42 @@ static int pci_d3cold_reset(struct pci_dev *dev, bool probe)
>  	return pci_set_power_state(dev, PCI_D0);
>  }
>  
> +/**
> + * pci_soft_reset - Software-initiated reset via D3hot as last resort
> + * @dev: PCI device to reset
> + * @probe: if true, check if soft reset is supported; if false, perform reset
> + *
> + * Attempt a software-initiated reset via D3hot->D0 transition as an absolute
> + * last resort when all other reset methods have failed. This method only
> + * becomes available if the device has PM capability, pci_pm_reset() is blocked
> + * (typically by NoSoftRst+), and pci_d3cold_reset() is not available.
> + *
> + * Some devices incorrectly advertise NoSoftRst+ but D3hot transition does
> + * provide sufficient reset for certain use cases (e.g., VFIO passthrough).
> + * This method provides a "better than nothing" option when the device would
> + * otherwise have no reset capability.
> + *
> + * Returns 0 if device can be/was reset this way, -ENOTTY if a better reset
> + * method is available (pm or d3cold) or device lacks PM capability, or other
> + * negative error code on failure.
> + */
> +static int pci_soft_reset(struct pci_dev *dev, bool probe)
> +{
> +	if (pci_pm_reset(dev, true) == 0)
> +		return -ENOTTY;
> +
> +	if (pci_d3cold_reset(dev, true) == 0)
> +		return -ENOTTY;
> +
> +	if (!dev->pm_cap)
> +		return -ENOTTY;
> +
> +	if (probe)
> +		return 0;
> +
> +	return pci_do_d3hot_transition(dev);
> +}
> +
>  /**
>   * pcie_wait_for_link_status - Wait for link status change
>   * @pdev: Device whose link to wait for.
> @@ -5105,6 +5156,7 @@ const struct pci_reset_fn_method pci_reset_fn_methods[] = {
>  	{ pci_reset_bus_function, .name = "bus" },
>  	{ cxl_reset_bus_function, .name = "cxl_bus" },
>  	{ pci_d3cold_reset, .name = "d3cold" },
> +	{ pci_soft_reset, .name = "soft" },
>  };
>  
>  /**
> diff --git a/include/linux/pci.h b/include/linux/pci.h
> index 1ca7b880ead7..bcd2987b868b 100644
> --- a/include/linux/pci.h
> +++ b/include/linux/pci.h
> @@ -51,7 +51,7 @@
>  			       PCI_STATUS_PARITY)
>  
>  /* Number of reset methods used in pci_reset_fn_methods array in pci.c */
> -#define PCI_NUM_RESET_METHODS 9
> +#define PCI_NUM_RESET_METHODS 10
>  
>  #define PCI_RESET_PROBE		true
>  #define PCI_RESET_DO_RESET	false