Russell King - ARM Linux wrote:

On Tue, Jun 28, 2011 at 09:57:40AM +0200, Linus Walleij wrote:

quoted

From: Ulf Hansson <redacted>

The end of a dma|pio data transfer is synced with the
DATAEND irq. This will prevent the mmci driver from ending
the request before the dma|pio job is completely done.

For dma we use DMA_PREP_INTERRUPT to register a callback
function which is called when dma driver is done with job.

To also make sure we prevent hanging forever, waiting for
DATAEND irq or a dma|pio transfer to be done, we setup a timer
when either a DATAEND or dma|pio job is done. Once both
conditions have occured, the timer is cancelled and the data
transfer is completed.

If a timeout occurs, the data transfer is terminated in a
controlled manner and EAGAIN is returned to the framework.
A timeout value of 50 ms has been found to work well for
our usecases.

What is the framework supposed to do with that error code?  Magic the
driver into disabling DMA by some non-existant callback?

Please, stop introducing magic new error codes which have no meaning to
the upper levels unless you're also going to add some handling of those
error conditions as well.

The idea was to trigger a resend from the mmc framework, but I noticed 
that the mmc_blk_issue_rw_rq has been patched quite a lot lately. I will 
revisit this again to see if it still makes sense.