RE: [PATCH] ASoC: DAPM: Add support for multi register mux

From: Songhee Baek
Date: Wed Apr 02 2014 - 03:06:22 EST


> -----Original Message-----
> From: Lars-Peter Clausen [mailto:lars@xxxxxxxxxx]
> Sent: Wednesday, April 02, 2014 12:02 AM
> To: Songhee Baek
> Cc: Arun Shamanna Lakshmi; lgirdwood@xxxxxxxxx; broonie@xxxxxxxxxx;
> swarren@xxxxxxxxxxxxx; perex@xxxxxxxx; tiwai@xxxxxxx; alsa-
> devel@xxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH] ASoC: DAPM: Add support for multi register mux
>
> On 04/02/2014 08:56 AM, Songhee Baek wrote:
> >
> >
> >> -----Original Message-----
> >> From: Lars-Peter Clausen [mailto:lars@xxxxxxxxxx]
> >> Sent: Tuesday, April 01, 2014 11:47 PM
> >> To: Songhee Baek
> >> Cc: Arun Shamanna Lakshmi; lgirdwood@xxxxxxxxx; broonie@xxxxxxxxxx;
> >> swarren@xxxxxxxxxxxxx; perex@xxxxxxxx; tiwai@xxxxxxx; alsa-
> >> devel@xxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> >> Subject: Re: [PATCH] ASoC: DAPM: Add support for multi register mux
> >>
> >> On 04/02/2014 08:17 AM, Songhee Baek wrote:
> >>>> -----Original Message-----
> >>>> From: Lars-Peter Clausen [mailto:lars@xxxxxxxxxx]
> >>>> Sent: Tuesday, April 01, 2014 11:00 PM
> >>>> To: Arun Shamanna Lakshmi
> >>>> Cc: lgirdwood@xxxxxxxxx; broonie@xxxxxxxxxx;
> >> swarren@xxxxxxxxxxxxx;
> >>>> perex@xxxxxxxx; tiwai@xxxxxxx; alsa-devel@xxxxxxxxxxxxxxxx; linux-
> >>>> kernel@xxxxxxxxxxxxxxx; Songhee Baek
> >>>> Subject: Re: [PATCH] ASoC: DAPM: Add support for multi register mux
> >>>>
> >>>> On 04/01/2014 08:26 PM, Arun Shamanna Lakshmi wrote:
> >>>> [...]
> >>>>>>> diff --git a/sound/soc/soc-dapm.c b/sound/soc/soc-dapm.c index
> >>>>>>> c8a780d..4d2b35c 100644
> >>>>>>> --- a/sound/soc/soc-dapm.c
> >>>>>>> +++ b/sound/soc/soc-dapm.c
> >>>>>>> @@ -514,9 +514,9 @@ static int dapm_connect_mux(struct
> >>>>>> snd_soc_dapm_context *dapm,
> >>>>>>> unsigned int val, item;
> >>>>>>> int i;
> >>>>>>>
> >>>>>>> - if (e->reg != SND_SOC_NOPM) {
> >>>>>>> - soc_widget_read(dest, e->reg, &val);
> >>>>>>> - val = (val >> e->shift_l) & e->mask;
> >>>>>>> + if (e->reg[0] != SND_SOC_NOPM) {
> >>>>>>> + soc_widget_read(dest, e->reg[0], &val);
> >>>>>>> + val = (val >> e->shift_l) & e->mask[0];
> >>>>>>> item = snd_soc_enum_val_to_item(e, val);
> >>>>>>
> >>>>>> This probably should handle the new enum type as well. You'll
> >>>>>> probably need some kind of flag in the struct to distinguish
> >>>>>> between the two enum types.
> >>>>>
> >>>>> Any suggestion on the flag name ?
> >>>>>
> >>>>
> >>>> How about 'onehot'?
> >>>>
> >>>> [...]
> >>>>>>> + reg_val = BIT(bit_pos);
> >>>>>>> + }
> >>>>>>> +
> >>>>>>> + for (i = 0; i < e->num_regs; i++) {
> >>>>>>> + if (i == reg_idx) {
> >>>>>>> + change = snd_soc_test_bits(codec, e->reg[i],
> >>>>>>> + e->mask[i],
> >>>>>> reg_val);
> >>>>>>> +
> >>>>>>> + } else {
> >>>>>>> + /* accumulate the change to update the
> >> DAPM
> >>>>>> path
> >>>>>>> + when none is selected */
> >>>>>>> + change += snd_soc_test_bits(codec, e-
> >>> reg[i],
> >>>>>>> + e->mask[i], 0);
> >>>>>>
> >>>>>> change |=
> >>>>>>
> >>>>>>> +
> >>>>>>> + /* clear the register when not selected */
> >>>>>>> + snd_soc_write(codec, e->reg[i], 0);
> >>>>>>
> >>>>>> I think this should happen as part of the DAPM update sequence
> >>>>>> like you had earlier. Some special care should probably be take
> >>>>>> to make sure that you de-select the previous mux input before
> >>>>>> selecting the new one if the new one is in a different register than
> the previous one.
> >>>>>
> >>>>> I am not sure I follow this part. We are clearing the 'not selected'
> >>>>> registers before we set the one we want. Do you want us to loop
> >>>>> the logic of soc_dapm_mux_update_power for each register ? or do
> >>>>> you want to change the dapm_update structure so that it takes all
> >>>>> the regs, masks, and values together ?
> >>>>
> >>>> The idea with the dapm_update struct is that the register updates
> >>>> are done in the middle of the power-down and power-up sequence.
> So
> >>>> yes, change the dapm_update struct to be able to hold all register
> >>>> updates and do all register updates in dapm_widget_update. I think
> >>>> an earlier version of your patch already had this.
> >>>
> >>> Is the change similar to as shown below?
> >>>
> >>> for (reg_idx = 0; reg_idx < e->num_regs; reg_idx++) {
> >>> val = e->values[item * e->num_regs + reg_idx];
> >>> ret = snd_soc_update_bits_locked(codec, e->reg[reg_idx],
> >>> e->mask[reg_idx], val);
> >>> if (ret)
> >>> return ret;
> >>> }
> >>>
> >>> During updating of the register's value, the above change can create
> >>> non-zero value in two different registers (very short transition) as
> >>> Mark mentioned for that change so we need to clear register first
> >>> before writing the desired value in the register.
> >>>
> >>> Should we add the clearing all registers and write the mux value in
> >>> desired register in the update function?
> >>>
> >>
> >> In dapm_update_widget() you have this line:
> >>
> >> ret = soc_widget_update_bits(w, update->reg, update->mask, update-
> >>> val);
> >>
> >> That needs to be done for every register update. When you setup the
> >> update struct you need to make sure that the register clears come
> >> before the register set.
> >>
> >> E.g. if you have register 0x3, 0x4, 0x5 and you select a bit in
> >> register 0x4 it should look like this.
> >>
> >> update->reg[0] = 0x3;
> >> update->val[0] = 0x0;
> >> update->reg[1] = 0x5;
> >> update->val[1] = 0x0;
> >> update->reg[2] = 0x4;
> >> update->val[2] = 0x8;
> >>
> >> When you set a bit in register 0x3 it should look like this:
> >>
> >> update->reg[0] = 0x4;
> >> update->val[0] = 0x0;
> >> update->reg[1] = 0x5;
> >> update->val[1] = 0x0;
> >> update->reg[2] = 0x3;
> >> update->val[2] = 0x1;
> >>
> >> So basically the write operation goes into update->reg[e->num_regs-1]
> >> the clear operations go into the other slots before that.
> >
> > Does update reg/val array have the writing sequence, is it correct?
> > And can I assume that update struct has reg/val/mask arrays not pointers?
>
> Right now the update struct does not have support for multiple register
> writes. That's up to you to implement this. I guess making it an array for now
> should be fine. But you need to add some safety checks to make sure that
> num_regs is not larger or equal to the array size.

Thank you for the clarification. We will add this in dapm update struct.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/