Re: [PATCH 1/3] dt-bindings: dmaengine: Add one new cell to present hardware slave id

[Vinod Koul]

On 19-02-19, 17:49, Baolin Wang wrote:
> Hi Geert,
> 
> On Tue, 19 Feb 2019 at 17:30, Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> >
> > Hi Baolin,
> >
> > On Tue, Feb 19, 2019 at 4:15 AM Baolin Wang <baolin.wang@xxxxxxxxxx> wrote:
> > > On Mon, 18 Feb 2019 at 20:23, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > > > On Mon, Feb 18, 2019 at 11:52 AM Baolin Wang <baolin.wang@xxxxxxxxxx> wrote:
> > > > > On Mon, 18 Feb 2019 at 18:31, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > > > > >
> > > > > > On Tue, Feb 12, 2019 at 9:25 AM Baolin Wang <baolin.wang@xxxxxxxxxx> wrote:
> > > > > > > On Fri, 1 Feb 2019 at 19:53, Baolin Wang <baolin.wang@xxxxxxxxxx> wrote:
> > > > > > > > On Thu, 31 Jan 2019 at 00:52, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > > > > > > > > On Tue, Jan 22, 2019 at 2:21 PM Baolin Wang <baolin.wang@xxxxxxxxxx> wrote:
> > > > > > > > > >
> > > > > > > > > >  Client:
> > > > > > > > > >  DMA clients connected to the Spreadtrum DMA controller must use the format
> > > > > > > > > > -described in the dma.txt file, using a two-cell specifier for each channel.
> > > > > > > > > > -The two cells in order are:
> > > > > > > > > > +described in the dma.txt file, using a three-cell specifier for each channel.
> > > > > > > > > > +The three cells in order are:
> > > > > > > > > >  1. A phandle pointing to the DMA controller.
> > > > > > > > > >  2. The channel id.
> > > > > > > > > > +3. The hardware slave id which is used for clients to trigger DMA engine
> > > > > > > > > > +automatically.
> > > > > > > > >
> > > > > > > > > I notice that this is an incompatible binding change. Is that necessary?
> > > > > > > > > If the current code works, I'd suggest allowing both #dma-cells=<2>
> > > > > > > > > and <3>, and then implementing both in the driver.
> > > > > > > >
> > > > > > > > Yes, this is necessary.
> > > > > > > >
> > > > > > > > Yes, current code can work, but the problem is that the DMA clients
> > > > > > > > must add one property (something like "sprd,slave-id") to specify the
> > > > > > > > slave id. So considering this, we want to change the dma-cells to 2,
> > > > > > > > including dma channel and dma slave id, which can avoid introducing
> > > > > > > > some similar properties for DMA clients.
> > > > > > > >
> > > > > > > > Now there are no DMA clients in mainline for Spreadtrum platform, and
> > > > > > > > we want to upstream our first DMA clients: SPI controller. So no other
> > > > > > > > drivers need to change when we changing dma cells. Thanks.
> > > > > > >
> > > > > > > Do you have any other concerns about this patch set? If not, I think
> > > > > > > Vinod can apply this patch set. Thanks.
> > > > > >
> > > > > > Sorry for the late reply. Yes, this makes sense since there are no
> > > > > > existing users then. For the DT changes going through the dmaengine
> > > > > > tree
> > > > > >
> > > > > > Acked-by: Arnd Bergmann <arnd@xxxxxxxx>
> > > > >
> > > > > Thanks for your reviewing.
> > > > >
> > > > > >
> > > > > > One more question, to make sure we don't need to edit it again:
> > > > > > Why do you need both a 'channel id' and a 'slave id' here? Is this
> > > > > > a strict hardware requirement for your DMA engine? In many
> > > > > > other designs, only a DMA request line number needs to
> > > > > > be described, and I think this would correspond to what you
> > > > > > call the 'hardware slave id' in your documentation.
> > > > >
> > > > > I try to explain why we need the slave id.
> > > > >
> > > > > For our DMA engine driver, we have software request mode and hardware
> > > > > request mode. For software request mode, the DMA engine driver need
> > > > > trigger DMA to start transfer manually. But for hardware request mode,
> > > > > we just set one unique slave id corresponding to the slave hardware to
> > > > > the DMA engine, then the slave hardware can trigger DMA automatically.
> > > > > And the slave id is not always same with the channel id according to
> > > > > the SoC design, so we add one cell to specify the slave id.
> > > >
> > > > I did understand the need for a slave-id, I was instead wondering about
> > > > the channel-id number. On many SoCs, all channels are equal, and you
> > > > just have to pick one of those with the right capabilities for a particular
> > > > slave.
> > >
> > > Yes, all channels are equal. We just set a unique slave id for the
> > > channel for a particular slave. For example, the SPI slave can use
> > > channel 10 for tx transfer by setting slave id 11, or it also can use
> > > channel 9 for tx transfer by setting same slave id 11.
> >
> > So the channel selection is software policy, not hardware description, and
> > thus doesn't belong in DT?
> >
> > Can't the DMA engine driver allocate channels dynamically, removing the
> > need to specify this in DT?
> 
> In theory we can do as you suggested. But we still want to
> manage/assign the DMA channel resources manually for one SoC, we can
> make sure some important DMA slaves (such as audio)  can request a DMA
> channel at runtime firstly, another benefit is that it is easy to
> debug since we can easily know which channel is assigned for this
> slave.

Are  you suggesting that you have more users than channels available?

I dont think it is hard to debug if channels are dynamic in
nature (for example we can print channel number and you know which one
are you talking about, fwiw i have worked on a such a system where we
grabbed the free channel)

-- 
~Vinod