dmaengine.c: question about device_alloc_chan_resources

[Timur Tabi <timur@freescale.com>]

Dan, Haavard, et al,

I am making some fixes to the Freescale DMA driver (drivers/dma/fsldma.c), and
I've come across a situation I don't understand.  Specifically, I have an issue
with the return values from fsl_dma_alloc_chan_resources(), which is fsldma's
device_alloc_chan_resources() function.

fsldma calls dma_async_device_register() for each DMA controller it finds.  The
problem is that when I use the dmatest driver, this results in
device_alloc_chan_resources() being called multiple times for the same DMA
channel.  When this happens, fsl_dma_alloc_chan_resources() must return -1
otherwise fsldma will hang during unload if the dmatest driver is also loaded.
The hang occurs in dma_async_device_unregister(), when it calls
wait_for_completion().

Here's what happens:

1) fsldma finds one DMA controller with four DMA channels.  It creates all the
channel objects and then calls dma_async_device_register().

2) Via dma_clients_notify_available(), dmatest is notified that there's a new
DMA resource.  It stores those four channels in its dmatest_channels object.

3) fsldma fins another DMA controller with four DMA channels.  Again, it creates
all the channel objects and then calls dma_async_device_register().

4) dma_clients_notify_available() is called again.  This time, however, it tells
dmatest about *eight* channels, not just the four additional ones.  It does this
because fsl_dma_alloc_chan_resources() returns 1 every time it's called.

The result is that the some DMA channel will appear twice in dmatest's channel
list.  For example, I added a bunch of printks to dma_test_add_channnel:

dmatest: Started 1 threads using dma2chan1
dmatest_add_channel:372 chan=df1f2a50
dmatest_add_channel:382 dtc->chan=df1f2898
dmatest_add_channel:382 dtc->chan=df1f2a50
dmatest_add_channel:382 dtc->chan=df1f2898
dmatest_add_channel:382 dtc->chan=df1f2a50

As you can see, df1f2898 appears twice, as does df1f2a50.

So my fix was to add this code to fsl_dma_alloc_chan_resources():

	/* Has this channel already been allocated? */
	if (fsl_chan->desc_pool)
		return -ENOMEM;

This fixes the hang, but I have a suspicion that it's still wrong.  When I look
at the other DMA device drivers, they don't do this.  For instance,
ioat_dma_alloc_chan_resources() does this:

	/* have we already been set up? */
	if (!list_empty(&ioat_chan->free_desc))
		return ioat_chan->desccount;

In other words, it returns the resources that have been allocated already.  But
if I do that, my driver hangs during an unload (if dmatest is loaded).

I think what's happening is that fsldma is the only DMA driver that calls
dma_async_device_register() more than once, and so it's exposing a bug in
dmaengine.c.  Can anyone confirm that?

-- 
Timur Tabi
Linux kernel developer at Freescale