Using GPMI-NAND driver on iMX28 using 3.4-rc1?

[Huang Shijie]

于 2012年04月06日 10:23, Shawn Guo 写道:
> On Thu, Apr 05, 2012 at 05:57:41PM +0530, Vinod Koul wrote:
>> On Thu, 2012-04-05 at 05:03 -0700, Sam Gandhi wrote:
>>> 2012/4/4 Huang Shijie<b32955 at freescale.com>:
> ...
>>>>>> flash_erase /dev/mtd1 0 0
>>>>>> Erasing 1------------[ cut here ]------------
>>>>>> kernel BUG at /home/sam/linux/drivers/dma/dmaengine.h:53!
>>>> the mxs-dma has added some patches about the cookie.
>>>> The bug is in the dmaengine.h.
>>>>
>>>> So let more people know this bug.
>>>>
> I'm still trying to understand why I did run into the BUG_ON when I
> test it with mxs mmc and sound drivers.
>
> ...
>
>> This means you are trying to mark a cookie complete when it is already
>> marked so!, hence dmaengine screams.
>> The bug is is mxs-dma.
>> Let me know if below fixes it (assuming you are not using cyclic API,
>> that would need fixup as well)
>>
>> diff --git a/drivers/dma/mxs-dma.c b/drivers/dma/mxs-dma.c
>> index c81ef7e..5ddd84e 100644
>> --- a/drivers/dma/mxs-dma.c
>> +++ b/drivers/dma/mxs-dma.c
>> @@ -399,6 +399,10 @@ static struct dma_async_tx_descriptor
>> *mxs_dma_prep_slave_sg(
>>   		ccw->bits&= ~CCW_DEC_SEM;
>>   	} else {
>>   		idx = 0;
>> +		/* assign cookie here,
>> +		 * hopefully for above case we dont need it
>> +		 */
>> +		dma_cookie_assign(&mxs_chan->desc);
> Isn't it done in mxs_dma_tx_submit() already? The gpmi driver somehow

It's maybe too late to assign the DMA cookie after mxs_dma_enable_chan() 
in mxs_dma_tx_submit().
The interrupt may arise before the dma_cookie_assign() finishes.

Why mmc/audio do not have this bug? their interrupt arise too slow.

I tested the following code :
==============================================================

diff --git a/drivers/dma/mxs-dma.c b/drivers/dma/mxs-dma.c
index 5978113..0f5b09a 100644
--- a/drivers/dma/mxs-dma.c
+++ b/drivers/dma/mxs-dma.c
@@ -202,10 +202,12 @@ static struct mxs_dma_chan *to_mxs_dma_chan(struct 
dma_cha
static dma_cookie_t mxs_dma_tx_submit(struct dma_async_tx_descriptor *tx)
{
struct mxs_dma_chan *mxs_chan = to_mxs_dma_chan(tx->chan);
+ dma_cookie_t c;

+ c = dma_cookie_assign(tx);
mxs_dma_enable_chan(mxs_chan);

- return dma_cookie_assign(tx);
+ return c;
}

static void mxs_dma_tasklet(unsigned long data)


> misses the call to dmaengine_submit() for some case?
>
> Regards,
> Shawn
>
>>   	}
>>
>>   	if (direction == DMA_TRANS_NONE) {
>>
>> --