[Linux-stm32] [PATCH RFC 2/6] dmaengine: Make of_dma_request_slave_channel pass a cookie to of_xlate

[Laurent Pinchart]

On Sat, Aug 02, 2025 at 02:37:54PM +0200, Konrad Dybcio wrote:
> On 8/1/25 2:00 PM, Laurent Pinchart wrote:
> > On Wed, Jul 30, 2025 at 02:37:54PM -0400, Frank Li wrote:
> >> On Wed, Jul 30, 2025 at 09:04:17PM +0300, Laurent Pinchart wrote:
> >>> On Wed, Jul 30, 2025 at 12:39:43PM -0400, Frank Li wrote:
> >>>> On Wed, Jul 30, 2025 at 11:33:29AM +0200, Konrad Dybcio wrote:
> >>>>> From: Konrad Dybcio <konrad.dybcio at oss.qualcomm.com>
> >>>>>
> >>>>> The DMA subsystem attempts to make it theoretically possible to pair
> >>>>> any DMA block with any user. While that's convenient from a
> >>>>> codebase sanity perspective, some blocks are more intertwined.
> >>>>>
> >>>>> One such case is the Qualcomm GENI, where each wrapper contains a
> >>>>> number of Serial Engine instances, each one of which can be programmed
> >>>>> to support a different protocol (such as I2C, I3C, SPI, UART, etc.).
> >>>>>
> >>>>> The GPI DMA it's designed together with, needs to receive the ID of the
> >>>>> protocol that's in use, to adjust its behavior accordingly. Currently,
> >>>>> that's done through passing that ID through device tree, with each
> >>>>> Serial Engine expressed NUM_PROTOCOL times, resulting in terrible
> >>>>> dt-bindings that are full of useless copypasta.
> >>>>>
> >>>>> In a step to cut down on that, let the DMA user give the engine driver
> >>>>> a hint at request time.
> >>>>>
> >>>>> Signed-off-by: Konrad Dybcio <konrad.dybcio at oss.qualcomm.com>
> >>>>> ---
> 
> [...]
> 
> >>>>> diff --git a/include/linux/of_dma.h b/include/linux/of_dma.h
> >>>>> index fd706cdf255c61c82ce30ef9a2c44930bef34bc8..9f9bc4207b85d48d73c25aad4b362e7c84c01756 100644
> >>>>> --- a/include/linux/of_dma.h
> >>>>> +++ b/include/linux/of_dma.h
> >>>>> @@ -19,7 +19,7 @@ struct of_dma {
> >>>>>  	struct list_head	of_dma_controllers;
> >>>>>  	struct device_node	*of_node;
> >>>>>  	struct dma_chan		*(*of_dma_xlate)
> >>>>> -				(struct of_phandle_args *, struct of_dma *);
> >>>>> +				(struct of_phandle_args *, struct of_dma *, void *);
> >>>>
> >>>> I suggest pass down more informaiton, like client's dev point. So we can
> >>>> auto create device link between client's dev and dma chan's device.
> >>>
> >>> Is .of_dma_xlate() really the right place to do that ? If you want to
> >>> create a device link for PM reasons, isn't it better created when the
> >>> channel is requested ? It should also be removed when the channel is
> >>> freed.
> >>
> >> I remember just need record client device pointer here.
> >>
> >>>>
> >>>> DMA Engineer device
> >>>>    DMA chan device
> >>>>        consumer clients' device.
> >>>>
> >>>> If consumer device runtime pm suspend can auto trigger DMA chan's device's
> >>>> runtime pm function.
> >>>>
> >>>> It will simplifly DMA engine's run time pm manage. Currently many DMA run
> >>>> time pm implement as, runtime_pm_get() when alloc and runtime_pm_put() at
> >>>> free channel.  But many devices request dma channel at probe, which make
> >>>> dma engine work at always 'on' state.
> >>>>
> >>>> But ideally, dma chan should be resume only when it is used to transfer.
> >>>
> >>> This is exactly what I was going to mention after reading the last
> >>> paragraph. Is there anything that prevents a DMA engine driver to
> >>> perform a rutime PM get() when a transfer is submitted
> >>
> >> DMA description is a queue, It is hard to track each descriptor submit and
> >> finished. espcially cycle buffer case.
> >>
> >> And according to dma engine API defination, submit a descriptor not
> >> neccessary to turn on clock, maybe just pure software operation, such as
> >> enqueue it to a software list.
> >>
> >> Many driver call dmaengine_submit() in irq context,  submit new descriptor
> >> when previous descriptor finished. runtime_pm_get() can NOT be called in
> >> atomic context.
> >>
> >> And some driver submit many descripor advance. Only issue_transfer() is
> >> actually trigger hardware to start transfer.
> >>
> >> Some client use cycle descripor, such audio devices.  Some audio devices
> >> have not free descriptor at their run time suspend function, just disable
> >> audio devices's clocks.  Audio devices run time suspend, which means no
> >> one use this dma channel, dma channel can auto suspend if built device link
> >> between audio device and dma chan devices.
> >>
> >> Some DMA client have not devices, such as memory to memory. for this kind
> >> case, it need keep chan always on.
> >>
> >> issue_transfer() can be call in atomic context. but trigger hardware transfer
> >> need clock and runtime_pm_get() can't be called in atomic context.
> >>
> >> Most case issue_transfer() is call in irq handle, which means device should
> >> already be in runtime resume statue.  DMA engine can safely access their
> >> register if using device link.
> > 
> > You have good points there, in particular the fact the issue_transfer()
> > can be called in interrupt context.
> > 
> > For me this calls for new DMA engine operations to "start/stop" the DMA
> > engine (better names are likely needed) from a client perspective.
> > 
> >>> and a put() when
> >>> it completes ? (Logically speaking, the actual implementation would
> >>> likely be a bit different in drivers, but the result would be similar.)
> 
> So.. do you folks want me to alter the patch in any way?

I think the runtime PM issue is orthogonal to the problem this series
addresses. It can be addressed separately.

That being said, I'm not a big fan of passing a void pointer to
.of_xlate() to carry device-specific information, in a device-specific
format. This seems prone to mismatch between clients and DMA engines.
.of_xlate() also seems the wrong place to do this. It would be cleaner
if we could use another operation, such as dmaengine_slave_config() for
instance.

-- 
Regards,

Laurent Pinchart