Re: [PATCH v2 1/8] dmaengine: Add API to combine configuration and preparation (sg and single)

[Vinod Koul <vkoul@kernel.org>]

On 09-03-26, 10:42, Frank Li wrote:
> On Mon, Mar 09, 2026 at 12:04:59PM +0100, Vinod Koul wrote:
> > On 18-12-25, 10:56, Frank Li wrote:
> > > Previously, configuration and preparation required two separate calls. This
> > > works well when configuration is done only once during initialization.
> > >
> > > However, in cases where the burst length or source/destination address must
> > > be adjusted for each transfer, calling two functions is verbose and
> > > requires additional locking to ensure both steps complete atomically.
> > >
> > > Add a new API dmaengine_prep_config_single() and dmaengine_prep_config_sg()
> > > and callback device_prep_config_sg() that combines configuration and
> > > preparation into a single operation. If the configuration argument is
> > > passed as NULL, fall back to the existing implementation.
> > >
> > > Add a new API dmaengine_prep_config_single_safe() and
> > > dmaengine_prep_config_sg_safe() for re-entrancy, which require driver
> > > implement callback device_prep_config_sg().
> >
> > Okay to add API
> >
> > > +	struct dma_async_tx_descriptor *(*device_prep_config_sg)(
> > > +		struct dma_chan *chan, struct scatterlist *sgl,
> > > +		unsigned int sg_len, enum dma_transfer_direction direction,
> > > +		unsigned long flags, struct dma_slave_config *config,
> > > +		void *context);
> >
> > Do we want to have drivers implement one more callback. It does not make
> > sense to me. Why not handle this in framework and have it call the
> > respective lower level APIs.
> 
> To avoid use addtional lock! suppose each API is re-entriable.
> 
> thread 1:  call dmaengine_prep_config_sg_safe()
> thread 2:  call dmaengine_prep_config_sg_safe()
> 
> If DMA engine driver implement device_prep_config_sg, thread 1 and thread 2
> can run parallel.
> 
> If driver have not implement this callback, it have to use mutex make sure
> config and prep atomic.
> 
> https://lore.kernel.org/dmaengine/20260109-edma_dymatic-v1-0-9a98c9c98536@nxp.com/
> show finial opitimziation result, which depend on this. If can't call
> prep() function parallel, which will kill performace.

Which seems to be 10% in your case.

> > Drivers should implement simple apis and collectively the functionality
> > can come from the framework.
> >
> > Would you consider revising as such. Bonus all existing drivers can
> > start using this API, no change required for drivers in that case'
> 
> Not that simple, some devices just call config at probe, especial fix
> FIFO address and burst length.
> 
> Call config and prep only need for the case, which need adjust src/dst
> address, burst length or other parameter for each transfer.

Correct. In the cases where config is done once they can invoke with
NULL config. I would like the middle layer to handle complexities and
drivers should be simpler

> 
> Frank
> 
> >
> >
> > --
> > ~Vinod

-- 
~Vinod