Re: [PATCH] soundwire: stream: fix programming slave ports for non-continous port maps

[Pierre-Louis Bossart]

On 7/30/24 10:39, Krzysztof Kozlowski wrote:
> On 30/07/2024 10:23, Krzysztof Kozlowski wrote:
>> On 29/07/2024 16:25, Pierre-Louis Bossart wrote:
>>>
>>>
>>> On 7/29/24 16:01, Krzysztof Kozlowski wrote:
>>>> Two bitmasks in 'struct sdw_slave_prop' - 'source_ports' and
>>>> 'sink_ports' - define which ports to program in
>>>> sdw_program_slave_port_params().  The masks are used to get the
>>>> appropriate data port properties ('struct sdw_get_slave_dpn_prop') from
>>>> an array.
>>>>
>>>> Bitmasks can be non-continuous or can start from index different than 0,
>>>> thus when looking for matching port property for given port, we must
>>>> iterate over mask bits, not from 0 up to number of ports.
>>>>
>>>> This fixes allocation and programming slave ports, when a source or sink
>>>> masks start from further index.
>>>>
>>>> Fixes: f8101c74aa54 ("soundwire: Add Master and Slave port programming")
>>>> Cc: <stable@xxxxxxxxxxxxxxx>
>>>> Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@xxxxxxxxxx>
>>>
>>> This is a valid change to optimize how the port are accessed.
>>>
>>> But the commit message is not completely clear, the allocation in
>>> mipi_disco.c is not modified and I don't think there's anything that
>>> would crash. If there are non-contiguous ports, we will still allocate
>>> space that will not be initialized/used.
>>>
>>> 	/* Allocate memory for set bits in port lists */
>>> 	nval = hweight32(prop->source_ports);
>>> 	prop->src_dpn_prop = devm_kcalloc(&slave->dev, nval,
>>> 					  sizeof(*prop->src_dpn_prop),
>>> 					  GFP_KERNEL);
>>> 	if (!prop->src_dpn_prop)
>>> 		return -ENOMEM;
>>>
>>> 	/* Read dpn properties for source port(s) */
>>> 	sdw_slave_read_dpn(slave, prop->src_dpn_prop, nval,
>>> 			   prop->source_ports, "source");
>>>
>>> IOW, this is a valid change, but it's an optimization, not a fix in the
>>> usual sense of 'kernel oops otherwise'.
>>>
>>> Am I missing something?
>>>
>>> BTW, the notion of DPn is that n > 0. DP0 is a special case with
>>> different properties, BIT(0) cannot be set for either of the sink/source
>>> port bitmask.
>>
>> I think we speak about two different things. port num > 1, that's
>> correct. But index for src_dpn_prop array is something different. Look
>> at mipi-disco sdw_slave_read_dpn():
>>
>> 173         u32 bit, i = 0;
>> ...
>> 178         addr = ports;
>> 179         /* valid ports are 1 to 14 so apply mask */
>> 180         addr &= GENMASK(14, 1);
>> 181
>> 182         for_each_set_bit(bit, &addr, 32) {
>> ...
>> 186                 dpn[i].num = bit;
>>
>>
>> so dpn[0..i] = 1..n
>> where i is also the bit in the mask.

yes, agreed on the indexing.

But are we in agreement that the case of non-contiguous ports would not
create any issues? the existing code is not efficient but it wouldn't
crash, would it?

There are multiple cases of non-contiguous ports, I am not aware of any
issues...

rt700-sdw.c:    prop->source_ports = 0x14; /* BITMAP: 00010100 */
rt711-sdca-sdw.c:       prop->source_ports = 0x14; /* BITMAP: 00010100
rt712-sdca-sdw.c:       prop->source_ports = BIT(8) | BIT(4);
rt715-sdca-sdw.c:       prop->source_ports = 0x50;/* BITMAP: 01010000 */
rt722-sdca-sdw.c:       prop->source_ports = BIT(6) | BIT(2); /* BITMAP:
01000100 */

same for sinks:

rt712-sdca-sdw.c:       prop->sink_ports = BIT(3) | BIT(1); /* BITMAP:
00001010 */
rt722-sdca-sdw.c:       prop->sink_ports = BIT(3) | BIT(1); /* BITMAP:
00001010 */

>> Similar implementation was done in Qualcomm wsa and wcd codecs like:
>> array indexed from 0:
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/sound/soc/codecs/wcd938x-sdw.c?h=v6.11-rc1#n51
>>
>> genmask from 0, with a mistake:
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/sound/soc/codecs/wcd938x-sdw.c?h=v6.11-rc1#n1255
>>
>> The mistake I corrected here:
>> https://lore.kernel.org/all/20240726-asoc-wcd-wsa-swr-ports-genmask-v1-0-d4d7a8b56f05@xxxxxxxxxx/
>>
>> To summarize, the mask does not denote port numbers (1...14) but indices
>> of the dpn array which are from 0..whatever (usually -1 from port number).
>>
> 
> Let me also complete this with a real life example of my work in
> progress. I want to use same dpn_prop array for sink and source ports
> and use different masks. The code in progress is:
> 
> https://git.codelinaro.org/krzysztof.kozlowski/linux/-/commit/ef709a0e8ab2498751305367e945df18d7a05c78#6f965d7b74e712a5cfcbc1cca407b85443a66bac_2147_2157
> 
> Without this patch, I get -EINVAL from sdw_get_slave_dpn_prop():
>   soundwire sdw-master-1-0: Program transport params failed: -2

Not following, sorry. The sink and source masks are separate on purpose,
to allow for bi-directional ports. The SoundWire spec allows a port to
be configured at run-time either as source or sink. In practice I've
never seen this happen, all existing hardware relies on ports where the
direction is hard-coded/fixed, but still we want to follow the spec.

So if ports can be either source or sink, I am not sure how the
properties could be shared with a single array?

Those two lines aren't clear to me at all:

	pdev->prop.sink_dpn_prop = wsa884x_sink_dpn_prop;
	pdev->prop.src_dpn_prop = wsa884x_sink_dpn_prop;