RE: [PATCH v2 00/26] IB: Optimize DMA mapping

From: Estrin, Alex
Date: Fri Jan 13 2017 - 21:05:27 EST




> -----Original Message-----
> From: Bart Van Assche [mailto:Bart.VanAssche@xxxxxxxxxxx]
> Sent: Friday, January 13, 2017 5:00 PM
> To: Estrin, Alex <alex.estrin@xxxxxxxxx>; dledford@xxxxxxxxxx
> Cc: linux-kernel@xxxxxxxxxxxxxxx; linux-rdma@xxxxxxxxxxxxxxx;
> gregkh@xxxxxxxxxxxxxxxxxxx
> Subject: Re: [PATCH v2 00/26] IB: Optimize DMA mapping
>
> On Fri, 2017-01-13 at 21:08 +0000, Estrin, Alex wrote:
> > It didn't fix the failure... Apparently there is an issue with generic
> > mapping itself.
>
> Hello Alex,
>
> The generic mapping code in lib/dma-virt.c works fine with at least the rxe
> driver. Additionally, as far as I can see the mapping code in lib/dma-virt.c
> is identical to the mapping in drivers/infiniband/sw/rdmavt/dma.c. What is
> not clear to me however is how my patches can have caused the SDMA engine
> error messages? As far as I can see the code in
> drivers/infiniband/hw/hfi1/sdma.c uses PCI DMA mapping operations instead of
> the generic dma_virt_ops. Is that correct?

Hi Bart,
If you refer to code like this
from sdma.c:
/* dma map the coalesce buffer */
addr = dma_map_single(&dd->pcidev->dev,
.....
please see hfi1/verbs.c @ hfi1_register_ib_device()

'dma_device' field is initialized here
-> ibdev->dma_device = &dd->pcidev->dev;
followed by
ret = rvt_register_device(&dd->verbs_dev.rdi);
-> set_dma_ops(rdi->ibdev.dma_device, &dma_virt_ops);

Thanks,
Alex.