[dpdk-dev] [PATCH v5] vfio: Support for no-IOMMU mode

[Burakov, Anatoly]

> 2016-01-27 16:50, Anatoly Burakov:
> > --- a/lib/librte_eal/linuxapp/eal/eal_pci_vfio.c
> > +++ b/lib/librte_eal/linuxapp/eal/eal_pci_vfio.c
> > +int vfio_type1_dma_map(int);
> > +int vfio_noiommu_dma_map(int);
> 
> WARNING:AVOID_EXTERNS: externs should be avoided in .c files I agree with
> checkpatch, they should be static ;)
> 

Will fix that.

> > --- a/lib/librte_eal/linuxapp/eal/eal_vfio.h
> > +++ b/lib/librte_eal/linuxapp/eal/eal_vfio.h
> > +/* older kernels may not have no-IOMMU mode */ #ifndef
> > +VFIO_NOIOMMU_IOMMU #define VFIO_NOIOMMU_IOMMU 8 #endif
> 
> Shouldn't it be defined privately in .c file?

We already have other VFIO-related definitions in that file, specifically the PCI defines that aren't present in earlier kernels. This definition is similar in nature - it will be present in kernels starting from 4.5 (when NOIOMMU was introduced), but earlier kernels will need this defined. I didn't want to go similar route with redefining everything VFIO-related, but maybe it makes sense in this case for consistency's sake? E.g.

#define RTE_VFIO_TYPE1 VFIO_TYPE1_IOMMU [we're already in an ifdef linux >= 3.6, so define type1 unconditionally]
#if linux < 4.5
#define RTE_VFIO_NOIOMMU 8
#else
#define RTE_VFIO_NOIOMMU VFIO_NOIOMMU_IOMMU
#endif

Or something like that?

Thanks,
Anatoly