On Fri, Sep 01, 2017 at 02:42:47PM +0000, Brüns, Stefan wrote:

On Freitag, 1. September 2017 15:35:49 CEST Maxime Ripard wrote:

quoted

On Fri, Sep 01, 2017 at 05:04:54AM +0200, Stefan Bruens wrote:

quoted

On Donnerstag, 31. August 2017 16:51:35 CEST Maxime Ripard wrote:

quoted

Hi,

On Thu, Aug 31, 2017 at 01:36:07AM +0200, Stefan Brüns wrote:

quoted

+/* Between SoC generations, there are some significant differences:
+ * - A23 added a clock gate register
+ * - the H3 burst length field has a different offset
+ */

This is not the proper comment style.

quoted

+enum dmac_variant {
+	DMAC_VARIANT_A31,
+	DMAC_VARIANT_A23,
+	DMAC_VARIANT_H3,
+};
+

And this is redundant with what we already have in our structures.

Actually, its not. For H3, there are currently at least 3 register
compatible SoCs: H5 is identical, R40 has 16 dma channels, A64 has 8
channels. So if the current config structure is kept, we need 3 different
compatible strings. Same for the A23, which is register compatible to
e.g. A83t and V3s, but with different numbers of DMA channels.

So either you decorate the code with a cascade of

if ((of_is_compatible(..A23..) || of_is_compatible(..A83T..) || ...) {
} else if ((of_is_compatible(..H3..) || of_is_compatible(..A64..) || ...)
{
} else { /* A31 */
}

in a number of places, or you do it just once.

That's not how you retrieve the structures. They are already
associated to the compatible, and you need to do a single lookup to
get them. So that's nowhere near what you're suggesting. You can have
a look at the of_match_device in the probe function.

Please have a look at the current implementation of how the clock autogating 
in the probe function is done - it matches with the compatible string.

Yeah, and we should get rid of that as well.

Of course we can replace this with a match between sdev->config and the 
various sun6i_dma_config instances, but we would still have to do 3 matches 
for the A23 register configuration (A23 || A83T || V3s) and 3 matches for the 
H3 register configuration (H3 || R40 || A64). There are currently *7* 
different configs (V3s, R40 and A64 taken into account), but only 3 different 
register variants.

This is the same rationale as the "gate_needed" boolean property proposed by 
Icenowy Zheng in the "Allwinner V3s DMA support" patch series. Obviously we 
don't need a boolean, but a ternary option to cater for the gate_needed 
differences - "NO_GATE", "A23_STYLE_GATE", "H3_STYLE_GATE".

Or we can just have an extra field in sun6i_dma_config that would set
the gate register offset? If it's non-zero, use whatever you have set
there, and then you just have to care about two cases, no matter how
many offsets we'll have in the wild.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com