On Thu, Jan 14, 2016 at 01:13:20AM +0200, Laurent Pinchart wrote:
> Hi Vinod,
>
> (CC'ing Linus as he's mentioned)
>
> On Wednesday 13 January 2016 14:55:50 Niklas Söderlund wrote:
> > * Vinod Koul <vinod.koul@xxxxxxxxx> [2016-01-13 19:06:01 +0530]:
> > > On Mon, Jan 11, 2016 at 03:17:46AM +0100, Niklas Söderlund wrote:
> > >> Enable slave transfers to devices behind IPMMU:s by mapping the slave
> > >> addresses using the dma-mapping API.
> > >>
> > >> Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@xxxxxxxxxxxx>
> > >> ---
> > >>
> > >> drivers/dma/sh/rcar-dmac.c | 64 +++++++++++++++++++++++++++++++++++++---
> > >> 1 file changed, 60 insertions(+), 4 deletions(-)
> > >>
> > >> diff --git a/drivers/dma/sh/rcar-dmac.c b/drivers/dma/sh/rcar-dmac.c
> > >> index 7820d07..da94809 100644
> > >> --- a/drivers/dma/sh/rcar-dmac.c
> > >> +++ b/drivers/dma/sh/rcar-dmac.c
> > >> @@ -13,6 +13,7 @@
> > >> #include <linux/dma-mapping.h>
> > >> #include <linux/dmaengine.h>
> > >> #include <linux/interrupt.h>
> > >> +#include <linux/iommu.h>
> > >> #include <linux/list.h>
> > >> #include <linux/module.h>
> > >> #include <linux/mutex.h>
> > >> @@ -1101,6 +1102,24 @@ rcar_dmac_prep_dma_cyclic(struct dma_chan *chan,
> > >> dma_addr_t buf_addr,
> > >> return desc;
> > >> }
> > >>
> > >> +static dma_addr_t __rcar_dmac_dma_map(struct dma_chan *chan,
> > >> phys_addr_t addr,
> > >> + size_t size, enum dma_data_direction dir)
> > >> +{
> > >> + struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
> > >> + struct page *page = phys_to_page(addr);
> > >> + size_t offset = addr - page_to_phys(page);
> > >> + dma_addr_t map = dma_map_page(chan->device->dev, page, offset, size,
> > >> + dir);
> > >
> > > Hmmmm, dmaengine APIs for slave cases expect that client has already
> > > ammped and provided an address which the dmaengine understands. So doing
> > > this in driver here does not sound good to me
> >
> > It was my understanding that clients do not do this mapping and in fact
> > are expected not to. Is this not what Linus Walleij is trying to address
> > in '[PATCH] dmaengine: use phys_addr_t for slave configuration'?
> There's a problem somewhere and we need to fix it. Clients currently pass
> physical addresses and the DMA engine API expects a DMA address. There's only
> two ways to fix that, either modify the API to expect a phys_addr_t, or modify
> the clients to provide a dma_addr_t.
Okay I am in two minds for this, doing phys_addr_t seems okay but somehow I
feel we should rather pass dma_addr_t and dmaengien driver get a right dma
address to use and thus fix the clients, that maybe the right thing to do
here, thoughts...?
The assumption from API was always that the client should perform the
mapping...
> The struct device used to map buffer through the DMA mapping API needs to be
> the DMA engine struct device, not the client struct device. As the client is
> not expected to have access to the DMA engine device I would argue that DMA
> engines should perform the mapping and the API should take a phys_addr_t.
That is not a right assumption. Once the client gets a channel, they have
access to dmaengine device and should use that to map. Yes the key is to map
using dmaengine device and not client device. You can use chan->device->dev.
>
> Vinod, unless you have reasons to do it otherwise, can we get your ack on this
> approach and start hammering at the code ? The problem has remained known and
> unfixed for too long, we need to move on.
--
~Vinod
--
To unsubscribe from this list: send the line "unsubscribe dmaengine" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
