Re: [PATCH v1 1/1] bus/pci: add RTE_PCI_DRV_NEED_IOVA_AS_PA flag

[Bruce Richardson <bruce.richardson@intel.com>]

On Sat, Oct 10, 2026 at 05:25:25PM +0800, liujie5@linkdatatechnology.com wrote:
> From: liufeng1 <liufeng1@sugon.com>
> 
> Add a new PCI driver flag, RTE_PCI_DRV_NEED_IOVA_AS_PA, that allows a
> driver to tell the PCI bus layer it cannot work with IOVA as VA and
> must use IOVA as PA.
> 
> The existing RTE_PCI_DRV_NEED_IOVA_AS_VA flag already provides the
> opposite requirement. Some PCI devices, however, are managed by a
> kernel driver that uses an identity-mapped IOMMU domain (e.g. when the
> 		kernel is booted with iommu=pt). The user-space PMD for such devices
> must use IOVA as PA, because the DMA addresses programmed by DPDK must
> match the identity mapping programmed by the kernel driver.
> 
> Without this flag, a PCI driver bound to a non-VFIO/UIO kernel driver
> has no way to force IOVA as PA. pci_device_iova_mode() returns
> RTE_IOVA_DC for those devices, and rte_eal_init() then sees an enabled
> IOMMU and selects RTE_IOVA_VA, which breaks identity-domain PMDs and
> forces users to pass --iova-mode=pa on the command line.
> 
> When RTE_PCI_DRV_NEED_IOVA_AS_PA is set, pci_device_iova_mode()
> returns RTE_IOVA_PA for the device. The requirement is propagated
> through rte_pci_get_iommu_class() to rte_bus_get_iommu_class(), so
> rte_eal_init() selects IOVA as PA automatically.
> 
> This is a generic mechanism: any PCI driver whose underlying kernel
> driver uses an identity IOMMU domain can set the flag.
> 
> Signed-off-by: Jie Liu <liujie5@linkdatatechnology.com>
> ---
>  drivers/bus/pci/bus_pci_driver.h | 2 ++
>  drivers/bus/pci/linux/pci.c      | 4 ++++
>  drivers/bus/pci/pci_common.c     | 3 +--
>  3 files changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/bus/pci/bus_pci_driver.h b/drivers/bus/pci/bus_pci_driver.h
> index c04ebddf59..5d5efbbb3e 100644
> --- a/drivers/bus/pci/bus_pci_driver.h
> +++ b/drivers/bus/pci/bus_pci_driver.h
> @@ -140,6 +140,8 @@ struct rte_pci_driver {
>  #define RTE_PCI_DRV_KEEP_MAPPED_RES 0x0020
>  /** Device driver needs IOVA as VA and cannot work with IOVA as PA */
>  #define RTE_PCI_DRV_NEED_IOVA_AS_VA 0x0040
> +/** Device driver needs IOVA as PA and cannot work with IOVA as VA */
> +#define RTE_PCI_DRV_NEED_IOVA_AS_PA 0x0080
>  
>  /**
>   * Register a PCI driver.
> diff --git a/drivers/bus/pci/linux/pci.c b/drivers/bus/pci/linux/pci.c
> index bcf0a409da..bf7125e214 100644
> --- a/drivers/bus/pci/linux/pci.c
> +++ b/drivers/bus/pci/linux/pci.c
> @@ -609,6 +609,8 @@ pci_device_iova_mode(const struct rte_pci_driver *pdrv,
>  			iova_mode = RTE_IOVA_PA;
>  		else if ((pdrv->drv_flags & RTE_PCI_DRV_NEED_IOVA_AS_VA) != 0)
>  			iova_mode = RTE_IOVA_VA;
> +		else if ((pdrv->drv_flags & RTE_PCI_DRV_NEED_IOVA_AS_PA) != 0)
> +			iova_mode = RTE_IOVA_PA;
>  		break;
>  	}
>  
> @@ -620,6 +622,8 @@ pci_device_iova_mode(const struct rte_pci_driver *pdrv,
>  	default:
>  		if ((pdrv->drv_flags & RTE_PCI_DRV_NEED_IOVA_AS_VA) != 0)
>  			iova_mode = RTE_IOVA_VA;
> +		else if ((pdrv->drv_flags & RTE_PCI_DRV_NEED_IOVA_AS_PA) != 0)
> +			iova_mode = RTE_IOVA_PA;
>  		break;
>  	}
>  	return iova_mode;

Do we need a sanity check somewhere that we don't have two drivers with
conflicting flags set i.e. one NEED_IOVA_AS_VA and one NEED_IOVA_AS_PA? It
should never happen, I imagine, but things that should never happen have
this annoying habit of sometimes happening! :-)

/Bruce