Re: [dpdk-dev] [RFC] eal/memory: introducing an option to set iova as va

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

On Mon, Jun 05, 2017 at 10:24:11AM +0530, santosh wrote:
> Hi Bruce,
> 
> 
> On Friday 02 June 2017 02:57 PM, Bruce Richardson wrote:
> > On Fri, Jun 02, 2017 at 09:54:46AM +0530, santosh wrote:
> >> Ping?
> >>
> >> On Wednesday 24 May 2017 09:41 PM, Santosh Shukla wrote:
> >>
> >>> Some NPU hardware like OCTEONTX follows push model to get
> >>> the packet from the pktio device. Where packet allocation
> >>> and freeing done by the HW. Since HW can operate only on
> >>> IOVA with help of SMMU/IOMMU, When packet receives from the
> >>> Ethernet device, It is the IOVA address(which is PA in existing scheme).
> >>>
> >>> Mapping IOVA as PA is expensive on those HW, where every
> >>> packet needs to be converted to VA from PA/IOVA.
> >>>
> >>> This patch proposes the scheme where the user can set IOVA
> >>> as VA by using an eal command line argument. That helps to
> >>> avoid costly lookup for VA in SW by leveraging the SMMU
> >>> translation feature.
> >>>
> >>> Signed-off-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
> >>> ---
> > Hi,
> >
> > I agree this is a problem that needs to be solved, but this doesn't look
> > like a particularly future-proofed solution. Given that we should
> > use the IOMMU on as many platforms as possible for protection, we
> > probably need to find an automatic way for DPDK to use IO addresses
> > correctly. Is this therefore better done as part of the VFIO and
> > UIO-specific code in EAL - as that is the part that knows how the memory
> > mapping is done, and in the VFIO case, what address ranges were
> > programmed in. The mempool driver was something else I considered but it
> > is probably too high a level to implement this.
> 
> The other approach which we evaluated, Its detail:
> 0) Introduce a new bus api whose job is to detect iommu capable devices on that
> bus {/ are those devices bind to iommu capable driver or not?}. Let's call that
> api rte_bus_chk_iommu_dev();
> 
> 1) The scheme is like If _all_ the devices bind to iommu kdrv then return iova=va
> 2) Otherwise switch to default mode i.e.. iova=pa.
> 3) Based on rte_bus_chk_iommu_dev() return value, 
> accordingly program iova=va Or iova=pa in vfio_type1/spapr_map(). 
> 
> 4) User from the command line can always override iova=va, 
> in case if he wants to default scheme( iova=pa mode). For that purpose - Introduce eal
> option something like --iova-pa Or --override-iova Or --iova-default 
> or some better name.
> 
> Proposed API snap:
> 
> enum iova_mode {
>     iova_va;
>     iova_pa;
>     iova_unknown;
> };
> 
> /**
>  * Look for iommu devices on that Bus.
>  * And find out that those devices bind to iommu
>  * capable driver example vfio.
>  *
>  *
>  * @return
>  *      On success return valid iova mode (iova_va or iova_pa)
>  *      On failure return iova_unkown.
>  */
> typedef int (*rte_bus_chk_iommu_dev_t)(void);
> 
> 
> By this approach, 
> - We can automatically detect iova is va or pa
> and then program accordingly. 
> - Also, the user can always switch to default iova mode.
> - Drivers like dpaa2 can use this API to detect iova mode then 
> program dma_map accordingly. Currently they are doing in ifdef-way.
> 
> Comments? thoughts? Or if anyone has better proposal then, please
> suggest.
> 

That sounds a more complete solution. However, it's probably a lot of
work to implement. :-)

I also wonder if we want to simplify things a little and disallow
mixed-mode operation i.e. all devices have to use UIO or all use VFIO?
Would that help to allow simplification or other options. Having a whole
new bus type seems strange for this. Can each bus just report whether
it's members require physical addresses. Then the EAL can manage a
single flag to report whether we are using VA or PA?

/Bruce