[PATCH v5 3/3] dmaengine: xilinx_dma: Fix race condition in the driver for multiple descriptor scenario

Vinod Koul vinod.koul at intel.com
Thu Jan 12 21:36:27 PST 2017


On Thu, Jan 12, 2017 at 02:19:49PM +0000, Appana Durga Kedareswara Rao wrote:
> Hi Vinod,
> 
> 	Thanks for the review...    
> 
> > On Sat, Jan 07, 2017 at 12:15:30PM +0530, Kedareswara rao Appana wrote:
> > > When driver is handling AXI DMA SoftIP When user submits multiple
> > > descriptors back to back on the S2MM(recv) side with the current
> > > driver flow the last buffer descriptor next bd points to a invalid
> > > location resulting the invalid data or errors in the DMA engine.
> > 
> > Can you rephrase this, it a bit hard to understand.
> 
> When DMA is receiving packets h/w expects the descriptors
> Should be in the form of a ring (I mean h/w buffer descriptor
> Next descriptor field should always point to valid address
> So that when DMA engine go and fetch that next descriptor it always 
> Sees a valid address).
>
> 
> But with the current driver implementation when user queues
> Multiple descriptors the last descriptor next descriptor field
> Pointing to an invalid location causing data corruption or 
> Errors from the DMA h/w engine...
> 
> To avoid this issue creating a Buffer descriptor Chain during 
> Channel allocation and using those buffer descriptors for processing
> User requested data.

Is it not doable to to modify the next pointer to point to subsequent
transaction. IOW you are modifying tail descriptor to point to subsequent
descriptor.

Btw how and when does DMA stop, assuming it is circular it never would,
isn't there a valid/stop flag associated with a descriptor which tells DMA
engine what to do next


Btw there is something wrong with your MUA perhaps line are titlecased for
no reason. This is typically behavious of non linux tool which may not be
great tool for this work.

> 
> Please let me know if the above explanation is not clear will explain in detail....
> 
> > 
> > >
> > > This patch fixes this issue by creating a BD Chain during
> > 
> > whats a BD?
> 
> Buffer descriptor.

Thats nowhere mentioned..

> 
> > 
> > > channel allocation itself and use those BD's.
> > >
> > > Signed-off-by: Kedareswara rao Appana <appanad at xilinx.com>
> > > ---
> > >
> > >  drivers/dma/xilinx/xilinx_dma.c | 133
> > > +++++++++++++++++++++++++---------------
> > >  1 file changed, 83 insertions(+), 50 deletions(-)
> > >
> > > diff --git a/drivers/dma/xilinx/xilinx_dma.c
> > > b/drivers/dma/xilinx/xilinx_dma.c index 0e9c02e..af2159d 100644
> > > --- a/drivers/dma/xilinx/xilinx_dma.c
> > > +++ b/drivers/dma/xilinx/xilinx_dma.c
> > > @@ -163,6 +163,7 @@
> > >  #define XILINX_DMA_BD_SOP		BIT(27)
> > >  #define XILINX_DMA_BD_EOP		BIT(26)
> > >  #define XILINX_DMA_COALESCE_MAX		255
> > > +#define XILINX_DMA_NUM_DESCS		255
> > 
> > why 255?
> 
> It is not an h/w limitation 
> Allocating 255 descriptors (Each descriptor is capable of sending 7MB data)
> So roughly using allocated descriptors DMA engine can transfer 1GB data
> And in the driver we are reusing the allocated descriptors when they are free.
> 
> Regards,
> Kedar.

-- 
~Vinod



More information about the linux-arm-kernel mailing list