Re: [dpdk-dev] [dpdk-dev,3/3] nfp: set device dma mask

[Alejandro Lucero <alejandro.lucero@netronome.com>]

On Thu, May 12, 2016 at 4:03 PM, Jan Viktorin <viktorin@rehivetech.com>
wrote:

> On Thu, 12 May 2016 15:34:00 +0100
> "Alejandro.Lucero" <alejandro.lucero@netronome.com> wrote:
>
> > - Just hugepages within the supported range will be available.
>
> Again the hyphen is redundant here.
>
> By the way, this text does not describe the change well. If I understood
> the whole patch set (I am not quite sure now), the initialization would
> fail if there are hugepages out of the given DMA mask. Am I wrong?
>
>
You are right.


> I'd expect something like "NFP supports DMA address in range ...".
>
>
That is a good idea. I was thinking on adding a memseg dump info as well
which would help to understand this issue and other related to memory
allocation.


> >
> > Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com>
> >
> > ---
> > drivers/net/nfp/nfp_net.c | 11 +++++++++++
> >  1 file changed, 11 insertions(+)
> >
> > diff --git a/drivers/net/nfp/nfp_net.c b/drivers/net/nfp/nfp_net.c
> > index ea5a2a3..e0e444a 100644
> > --- a/drivers/net/nfp/nfp_net.c
> > +++ b/drivers/net/nfp/nfp_net.c
> > @@ -115,6 +115,14 @@ enum nfp_qcp_ptr {
> >       NFP_QCP_WRITE_PTR
> >  };
> >
> > +#ifndef DMA_64BIT_MASK
> > +#define DMA_64BIT_MASK  0xffffffffffffffffULL
> > +#endif
> > +
> > +#ifndef DMA_BIT_MASK
> > +#define DMA_BIT_MASK(n) (((n) == 64) ? DMA_64BIT_MASK : ((1ULL<<(n))-1))
> > +#endif
>
> This is quite a generic code, I'd put it into the EAL. Probably, it should
> be renamed to something like RTE_DMA_BIT_MASK.


OK.


>
>
> +
> >  /*
> >   * nfp_qcp_ptr_add - Add the value to the selected pointer of a queue
> >   * @q: Base address for queue structure
> > @@ -2441,6 +2449,9 @@ nfp_net_init(struct rte_eth_dev *eth_dev)
> >       /* Recording current stats counters values */
> >       nfp_net_stats_reset(eth_dev);
> >
> > +     /* Setting dma_mask */
> > +     eth_dev->data->dma_mask = DMA_BIT_MASK(40);
>
> Can we read this from /sys/bus/pci/devices/*/dma_mask_bits? I am not sure
> whether is this generic enough but I can see dma_mask_bits for the PCI
> devices on my PC.
>
>
The kernel adds a default dma mask when device scanning (at least for PCI
devices). It is a device driver who knows about specific DMA addressing
limitations. For example, this is done with UIO (igb_uio) and the using
sysfs would be fine (but then you should add support for specifying a dma
mask in igb_uio as a module param) but this is not true for VFIO.



> Regards
> Jan
>
> > +
> >       return 0;
> >  }
> >
>
>
>
>
> --
>    Jan Viktorin                  E-mail: Viktorin@RehiveTech.com
>    System Architect              Web:    www.RehiveTech.com
>    RehiveTech
>    Brno, Czech Republic
>