The L3 throughput can be higher than expected when packed access
is not enabled.  The ratio depends on the number of bytes in a
transaction and the EMIF interface width.

The throughput was measured for the following settings/cases:

* Case 1: Burst size of 64 bytes, packed access disabled
* Case 2: Burst size of 64 bytes, packed access enabled
* Case 3: Burst disabled, packed access disabled

Throughput measurements were done during McASP-based audio
playback on the Jacinto6 EVM using the omapconf tool [1]:
$ omapconf trace bw -m sdma_rd

 ---------------------------------------------------------
                                  Throughput (MB/s)
  Audio parameters            Case 1    Case 2    Case 3
 ---------------------------------------------------------
  44.1kHz, 16-bits, stereo      1.41      0.18      1.41
  44.1kHz, 32-bits, stereo      1.41      0.35      1.41
  44.1kHz, 16-bits, 4-chan      2.82      0.35      2.82
  44.1kHz, 16-bits, 6-chan      4.23      0.53      4.23
  44.1kHz, 16-bits, 8-chan      5.64      0.71      5.64
 ---------------------------------------------------------

From above measurements, case 2 is the only one that delivers
the expected throughput for the given audio parameters.  For
that reason, the packed accesses are now enabled.

It's worth to mention that packed accesses cannot be enabled
for all addressing modes. In cyclic transfers, it can be
enabled in the source for MEM_TO_DEV and in dest for DEV_TO_MEM,
as they use post-increment mode which supports packed accesses.

Peter Ujfalusi:
From the TRM regarding to this:
"NOTE: Except in the constant addressing mode, the source or
destination must be specified as packed for burst transactions
to occur."

So w/o the packed setting the burst on the MEM side was not
enabled, this explains the numbers.

[1] https://github.com/omapconf/omapconf

Signed-off-by: Misael Lopez Cruz <misael.lopez@ti.com>
Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>

Since the mapping between the hardware request lines and channels has been
removed it no longer make sense to have too many channels.
Set the number of channels to match with the number of logical channels
supported by sDMA.

Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>

Do not direct map the virtual channels to sDMA request number. When the
sDMA is behind of a crossbar this direct mapping can cause situations when
certain channel can not be requested since the crossbar request number
will no longer match with the sDMA request line.
The direct mapping for virtual channels with HW request lines will make it
harder to implement MEM_TO_MEM mode for the driver.

Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>

Use the dma-requests property from DT to get the number of DMA requests.
In case of legacy boot or failure to find the property, use the default
127 as number of requests.

Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>

Instead of magic numbers in the code, use define for number of logical DMA
channels and DMA requests.

Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>

The sDMA controller is capable of performing memory copy operation. It need
to be configured to software triggered mode and without HW synchronization.
The sDMA can copy data which is aligned to 8, 16 or 32 bits.

Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>

In omap_dma_start_desc the vdesc->node is removed from the virt-dma
framework managed lists (to be precise from the desc_issued list).
If a terminate_all comes before the transfer finishes the omap_desc will
not be freed up because it is not in any of the lists and we stopped the
DMA channel so the transfer will not going to complete.
There is no special sequence for leaking memory when using cyclic (audio)
transfer: with every start and stop of a cyclic transfer the driver leaks
struct omap_desc worth of memory.

Free up the allocated memory directly in omap_dma_terminate_all() since the
framework will not going to do that for us.

Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
CC: <stable@vger.kernel.org>
CC: <linux-omap@vger.kernel.org>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>

Signed-off-by: Vinod Koul <vinod.koul@intel.com>

Now that the generic slave caps code can make use of the device assigned
capabilities, instead of relying on a callback to be implemented.

Make use of this code.

Signed-off-by: Maxime Ripard <maxime.ripard@free-electrons.com>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>