Re: [dpdk-dev] [PATCH 0/4] librte_pmd_virtio :Fix: virtio_pci.h non-existent virtio feature bit-flag tested

[Stephen Hemminger <stephen@networkplumber.org>]

On Thu,  5 Jun 2014 15:39:15 +0100
Alan Carew <alan.carew@intel.com> wrote:

> This series addresses an issue with librte_pmd_virtio where the offset to the
> virtio device specific header may be incorrect depending on whether MSI-X has
> been enabled or not.
> 
> If MSI-X is configured the device specific header is placed at byte offset 24
> relative to the IO base address. 
> If MSI-X is not configured the device specific header is placed at
> byte offset 20. 
> 
> The following macro defined in virtio_pci.h is used to test the
> presence of the MSI-X header and determine the correct offset: 
> #define VIRTIO_PCI_CONFIG(hw) (((hw)->guest_features & VIRTIO_PCI_FLAG_MSIX) ? 24 : 20) 
> 
> However, VIRTIO_PCI_FLAG_MSIX is not a guest_feature nor is it part of the
> Virtio Specification and resolves to the VIRTIO_NET_F_MAC feature as both
> are #defined as 0x20. 
> 
> VIRTIO_PCI_FLAG_MSIX or similar flag should instead be set by the kernel
> driver allocating resources and passed to user space for testing. 
> i.e. 
> #define VIRTIO_PCI_CONFIG(hw) (((hw)->intr_mode & IGBUIO_MSIX_INTR_MODE) ? 24 : 20)
> 
> To enable this testing of interrupt mode, this series allows for the kernel
> driver(igb_uio) to place the configured interrupt mode into a sysfs entry.
> sysfs is then parsed by eal_pci to determine the configured mode, which
> allows all user space devices to correctly determine the interrupt mode,
> including virtio_ethdev.
> 
> This series should be applied prior to Anatoly Burakov's
> [VFIO] Add VFIO support to DPDK series
> 
> Alan Carew (4):
>   igb_uio: Add interrupt_mode sysfs entry for igb_uio devices
>   eal_pci: Add interrupt mode to rte_pci_device and parsing to eal_pci
>   FreeBSD: Adds the equivalent interrupt mode setting and parsing
>   virtio: Fixes the VIRTIO_PCI_CONFIG macro to use the correct offset
>     to the Virtio header
> 
>  lib/librte_eal/bsdapp/eal/eal_pci.c                |   44 ++++++++++
>  lib/librte_eal/bsdapp/nic_uio/nic_uio.c            |   14 +++
>  lib/librte_eal/common/Makefile                     |    1 +
>  lib/librte_eal/common/include/rte_pci.h            |    2 +
>  .../common/include/rte_pci_dev_feature_defs.h      |   85 ++++++++++++++++++++
>  .../common/include/rte_pci_dev_features.h          |   70 ++++++++++++++++
>  lib/librte_eal/linuxapp/eal/eal_pci.c              |   78 ++++++++++++++++++
>  lib/librte_eal/linuxapp/igb_uio/igb_uio.c          |   48 +++++++++---
>  lib/librte_pmd_virtio/virtio_ethdev.c              |    1 +
>  lib/librte_pmd_virtio/virtio_pci.h                 |    4 +-
>  10 files changed, 336 insertions(+), 11 deletions(-)
>  create mode 100755 lib/librte_eal/common/include/rte_pci_dev_feature_defs.h
>  create mode 100755 lib/librte_eal/common/include/rte_pci_dev_features.h
> 

I took the simpler approach. Since DPDK devices doesn't really need the multiple
vectors of MSI-X at all, I just converted igb_uio to use MSI instead.