On Sat, 2012-01-07 at 19:21 +0000, Russell King - ARM Linux wrote:

On Sat, Jan 07, 2012 at 10:54:43AM -0800, David Brown wrote:

quoted

On Fri, Jan 06, 2012 at 05:59:29PM -0800, Daniel Walker wrote:

quoted

quoted

diff --git a/drivers/dma/msm-dma.c b/drivers/dma/msm-dma.c
new file mode 100644
index 0000000..51d9a2b
--- /dev/null
+++ b/drivers/dma/msm-dma.c

...
+static void msm_chan_desc_cleanup(struct msm_dma_chan *chan)
+{
+	struct msm_dma_desc_sw *desc, *_desc;
+	unsigned long flags;
+
+	dev_dbg(chan->dev, "Cleaning completed descriptor of channel %d\n",
+							chan->chan_id);
+	spin_lock_irqsave(&chan->lock, flags);
+
+	list_for_each_entry_safe(desc, _desc, &chan->active_list, node) {
+		dma_async_tx_callback callback;
+		void *callback_param;
+
+		if (msm_dma_desc_status(chan, desc) == DMA_IN_PROGRESS)
+			break;
+
+		/* Remove from the list of running transactions */
+		list_del(&desc->node);
+
+		/* Run the link descriptor callback function */
+		callback = desc->async_tx.callback;
+		callback_param = desc->async_tx.callback_param;
+		if (callback) {
+			spin_unlock_irqrestore(&chan->lock, flags);
+			callback(callback_param);

Are you sure unlocking here is safe? at_hdmac.c holds the lock the
entire time, and fsldma.c deletes the entire list, then runs the
operations.

Good catch.

According to a comment in at_hdmac.c, it is safe to hold the lock
while calling the callbacks, so that's probably the easiest solution.
I suspect that you've got something in another driver expecting the
lock to be released, and that might have to be changed.

It is _not_ safe to hold the lock while calling callbacks.

Please refer to the DMA engine documentation, found here:

Documentation/dmaengine.txt

section 3:

   Note:
        Although the async_tx API specifies that completion callback
        routines cannot submit any new operations, this is not the
        case for slave/cyclic DMA.

        For slave DMA, the subsequent transaction may not be available
        for submission prior to callback function being invoked, so
        slave DMA callbacks are permitted to prepare and submit a new
        transaction.

        For cyclic DMA, a callback function may wish to terminate the
        DMA via dmaengine_terminate_all().

*       Therefore, it is important that DMA engine drivers drop any
*       locks before calling the callback function which may cause a
*       deadlock.

Here's the comment from at_hdmac.c .

                /*
                 * The API requires that no submissions are done from a
                 * callback, so we don't need to drop the lock here
                 */
                if (callback)
                        callback(param);

I don't know much about the DMA engine, but maybe there is some special
case here that makes this ok.. (CC'ed Nicolas Ferre)

Daniel