Re: [dpdk-dev] [PATCH 07/10] linuxapp/eal_vfio: honor iova mode before mapping

[Jerin Jacob <jerin.jacob@caviumnetworks.com>]

-----Original Message-----
> Date: Wed, 5 Jul 2017 11:14:01 +0200
> From: Maxime Coquelin <maxime.coquelin@redhat.com>
> To: Santosh Shukla <santosh.shukla@caviumnetworks.com>,
>  thomas@monjalon.net, bruce.richardson@intel.com, dev@dpdk.org
> CC: jerin.jacob@caviumnetworks.com, hemant.agrawal@nxp.com,
>  shreyansh.jain@nxp.com, gaetan.rivet@6wind.com
> Subject: Re: [dpdk-dev] [PATCH 07/10] linuxapp/eal_vfio: honor iova mode
>  before mapping
> User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101
>  Thunderbird/52.1.0
> 
> 
> 
> On 06/08/2017 01:05 PM, Santosh Shukla wrote:
> > Check iova mode and accordingly map iova to pa or va.
> > 
> > Signed-off-by: Santosh Shukla<santosh.shukla@caviumnetworks.com>
> > Signed-off-by: Jerin Jacob<jerin.jacob@caviumnetworks.com>
> > ---
> >   lib/librte_eal/linuxapp/eal/eal_vfio.c | 10 ++++++++--
> >   1 file changed, 8 insertions(+), 2 deletions(-)
> > 
> > diff --git a/lib/librte_eal/linuxapp/eal/eal_vfio.c b/lib/librte_eal/linuxapp/eal/eal_vfio.c
> > index 04914406f..348b7a7f4 100644
> > --- a/lib/librte_eal/linuxapp/eal/eal_vfio.c
> > +++ b/lib/librte_eal/linuxapp/eal/eal_vfio.c
> > @@ -706,7 +706,10 @@ vfio_type1_dma_map(int vfio_container_fd)
> >   		dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
> >   		dma_map.vaddr = ms[i].addr_64;
> >   		dma_map.size = ms[i].len;
> > -		dma_map.iova = ms[i].phys_addr;
> > +		if (rte_eal_iova_mode() == RTE_IOVA_VA)
> > +			dma_map.iova = dma_map.vaddr;
> > +		else
> > +			dma_map.iova = ms[i].phys_addr;
> >   		dma_map.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
> 
> IIUC, it is changing default behavior for VFIO devices.
> 
> I see a possible problem, but I'm not sure the case is valid.
> 
> Imagine you have two devices in the iommu group, and the two devices are
> used in separate processes. Each process could try two different
> physical addresses at the same virtual address, and so the second map
> would fail.

IMO, Doesn't look like a problem. Here is the data flow

1) The vfio DMA map function(vfio_type1_dma_map()) will be called only
on primary process
http://dpdk.org/browse/dpdk/tree/lib/librte_eal/linuxapp/eal/eal_vfio.c#n359

2) On secondary process, DPDK rte_eal_huge_page_attach() will make sure
that, the Secondary process has the _same_ virtual address as primary or
exit from on attach.
http://dpdk.org/browse/dpdk/tree/lib/librte_eal/linuxapp/eal/eal_memory.c#n1452 

3) Since secondary process adds the mapped the virtual address in step (2).
in the page table in OS. On SMMU entry miss(When device
request from I/O transaction), OS will load the mapping and update the SMMU 
"context" with page tables from MMU.

Let me add the background for why this feature is required in DPDK to
enable NPU style co-processors.

The traditional NICs the Rx path code look like this:
1) On control path, Fill the mempool with buffers
2) on rx_burst(), alloc the mbuf from mempool
3) SW has the mbuf in hand(which is a virtual address) and program the
HW with mbuf->buf_physaddr)
4) Return the last pushed mbuf(will be updated by HW by now)


On NPU style co-processors, situation is different as the buffer recycling 
has been done in HW unlike SW model. Here is the data flow:
1) On control path, Fill the HW mempool with buffers(Obviously the IOVA
address, which is PA in existing model)
2) on rx_burst, HW gives you IOVA address(as address as step 1)
3) As application expects VA to operate on it, rx_burst() needs to
convert to VA from PAA. Which is very costly.
Instead with this IOVA as VA scheme, We can avoid the cost of converting
with help of IOMMU/SMMU.

This patch set auto detects the mode based available of type devices in
bus and provides an option to override mode based on eal argument, so we
don't foresee any issue with this approach and welcome any alternative
approaches.

Similar problem exists in another part of the code in DPDK,
http://dpdk.org/browse/dpdk/tree/drivers/bus/fslmc/fslmc_vfio.c#n231
Its a conditional compilation based approach with duplicating the vfio
code and we are trying to fix the problem in a generic way so that
everyone can get benefited out of it.

Comments are welcome.

/Jerin

> 
> By using physical addresses, you are safe against this problem.
> 
> Any thoughts?
> 
> Cheers,
> Maxime