Re: [PATCH] dma: Add Keystone Packet DMA Engine driver

From: Vinod Koul <vinod.koul@intel.com>
To: Arnd Bergmann <arnd@arndb.de>
Cc: linux-arm-kernel@lists.infradead.org,
	Santosh Shilimkar <santosh.shilimkar@ti.com>,
	Mark Rutland <mark.rutland@arm.com>,
	devicetree@vger.kernel.org, Russell King <linux@arm.linux.org.uk>,
	linux-kernel@vger.kernel.org, Sandeep Nair <sandeep_n@ti.com>,
	Grant Likely <grant.likely@linaro.org>,
	Rob Herring <robh+dt@kernel.org>,
	dmaengine@vger.kernel.org
Subject: Re: [PATCH] dma: Add Keystone Packet DMA Engine driver
Date: Tue, 18 Mar 2014 21:21:35 +0530	[thread overview]
Message-ID: <20140318155135.GF1976@intel.com> (raw)
In-Reply-To: <4424649.TIey3uEAed@wuerfel>

On Tue, Mar 18, 2014 at 04:38:25PM +0100, Arnd Bergmann wrote:
> On Tuesday 18 March 2014 20:54:44 Vinod Koul wrote:
> > On Mon, Mar 17, 2014 at 03:37:47PM -0400, Santosh Shilimkar wrote:
> > > >> To simplify this bit more, you can think of this as DMA channels, flows
> > > >> are allocated and DMA channels are enabled by DMA engine and they remains
> > > >> enabled always as long as the channel in use. Enablling dma channel
> > > >> actually don't start the DMA transfer but just sets up the connection/pipe
> > > >> with peripheral and memory and vice a versa.
> > > >>
> > > >> All the descriptor management, triggering, sending completion interrupt or
> > > >> hardware signal to DMAEngine all managed by centralised QMSS.
> > > >>
> > > >> Actual copy of data is still done by DMA hardware but its completely
> > > >> transparent to software. DMAEngine hardware takes care of that in the
> > > >> backyard.
> > > > So you will use the dmaengine just for setting up the controller. Not for actual
> > > > transfers. Those would be governed by the QMSS, right?
> > > >
> > > Correct.
> > >  
> > > > This means that someone expecting to use dmaengine API will get confused about
> > > > this and doing part (alloc) thru dmaengine and rest (transfers) using some other
> > > > API. This brings to me the design approach, does it really make sense creating
> > > > dmaengine driver for this when we are not fully complying to the API
> > > > 
> > > Thats fair. The rationale behind usage of DMEngine was that its the closest
> > > available subsystem which can be leveraged for this hardware. We can
> > > pretty much use all the standard DMAEngine device tree parsing as well as
> > > the config API to setup DMAs. 
> > > 
> > > I think you made your stand clear, just to confirm, you don't prefer this
> > > driver to be a DMAEngine driver considering it doesn't fully complying to
> > > the APIs. We could document the deviation of 'transfer' handling to avoid
> > > any confusion.
> > Yup, a user will just get confused as the driver doenst conform the dmaengine
> > API. Unless someone comes up witha  strong argument on why it should be
> > dmaengine driver and what befits we see form such a model, i would like a
> > damengine driver to comply to standard API and usage.
> 
> I think it would be possible to turn the QMSS driver into a library and have
> the packet DMA code use the proper dmaengine API by calling into that code.
> 
> The main user of packet DMA (the ethernet driver) would however still have
> to call into QMSS directly, so I'm not sure if it's worth the effort.

Wouldn't that make clients use a standard API and also help in transaction
management by using existing infrastructure

-- 
~Vinod