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

[Vinod Koul <vinod.koul@intel.com>]

On Wed, Mar 12, 2014 at 03:50:32AM +0800, Santosh Shilimkar wrote:
> On Tuesday 11 March 2014 06:23 PM, Vinod Koul wrote:
> > On Fri, Feb 28, 2014 at 05:56:40PM -0500, Santosh Shilimkar wrote:
> >> From: Sandeep Nair <sandeep_n@ti.com>
> >>
> >> The Packet DMA driver sets up the dma channels and flows for the
> >> QMSS(Queue Manager SubSystem) who triggers the actual data movements
> >> across clients using destination queues. Every client modules like
> >> NETCP(Network Coprocessor), SRIO(Serial Rapid IO) and CRYPTO
> >> Engines has its own instance of packet dma hardware. QMSS has also
> >> an internal packet DMA module which is used as an infrastructure
> >> DMA with zero copy.
> >>
> >> Patch adds DMAEngine driver for Keystone Packet DMA hardware.
> >> The specifics on the device tree bindings for Packet DMA can be
> >> found in:
> >> 	Documentation/devicetree/bindings/dma/keystone-pktdma.txt
> >>
> >> The driver implements the configuration functions using standard DMAEngine
> >> apis. The data movement is managed by QMSS device driver.
> > Pls use subsystem appropriate name so here would have been dmaengine: ...
> > 
> > So i am still missing stuff like prepare calls, irq, descriptor management to
> > call this a dmaengine driver.
> >
> > I guess you need to explain a bit more why the data movement is handled by some
> > other driver and not by this one
> >
> To expand above statement, Packet DMA hardware blocks on Keystone SOCs
> are DMAEngines. QMSS is centralised subsystem manages multiple functionalities
> including triggering dma transfers, descriptor management and completion
> irqs. There are separate instance of packet DMA hardware block per client
> device. We program the DMA hardware to allocate channels and flows. So
> the packet DMA resouces like dma channels and dma flows are configured
> and managed through the DMAEngine driver. Thats why we implement only
> device_alloc_chan_resources, device_free_chan_resources and device_control
> DMAEngine APIs.
Sorry am bit lost. If its a dmaengine then you still need to program the
transfers, how does that part work?

> >> +#define BITS(x)			(BIT(x) - 1)
> > this might get confusing, perhaps a better name could be given?
> > 
> Would "BIT_MASK" be ok with you ?
something which would imply its x -1, am not really good with name so no
suggestions :)

> >> +
> >> +#define BUILD_CHECK_REGS()						\
> >> +	do {								\
> >> +		BUILD_BUG_ON(sizeof(struct reg_global)   != 32);	\
> >> +		BUILD_BUG_ON(sizeof(struct reg_chan)     != 32);	\
> >> +		BUILD_BUG_ON(sizeof(struct reg_rx_flow)  != 32);	\
> >> +		BUILD_BUG_ON(sizeof(struct reg_tx_sched) !=  4);	\
> >> +	} while (0)
> > why is this required, do you want to use __packed__ to ensure right size?
> >
> This is just to ensure the register sanity. We should use __packed__ as
> well. We can take the BUILD_CHECK_REGS() out if you don't prefer it.
putting packed ensures so no need for this check.

-- 
~Vinod