Hi Jacopo,
On 10/01/2019 08:58, Jacopo Mondi wrote:
> On Wed, Jan 09, 2019 at 02:15:33PM +0000, Kieran Bingham wrote:
>> On 09/01/2019 00:15, Laurent Pinchart wrote:
>>> Hello,
>>>
>>> On Monday, 7 January 2019 14:36:28 EET Kieran Bingham wrote:
>>>> On 06/01/2019 15:54, Jacopo Mondi wrote:
>>>>> When the adv748x driver is informed about a link being created from HDMI
>>>>> or AFE to a CSI-2 TX output, the 'link_setup()' callback is invoked. Make
>>>>> sure to implement proper routing management at link setup time, to route
>>>>> the selected video stream to the desired TX output.
>>>>
>>>> Overall this looks like the right approach - but I feel like the
>>>> handling of the io10 register might need some consideration, because
>>>> it's value depends on the condition of both CSI2 transmitters, not just
>>>> the currently parsed link.
>>>>
>>>> I had a go at some pseudo - uncompiled/untested code inline as a suggestion.
>>>>
>>>> If you think it's better - feel free to rework it in ... or not as you
>>>> see fit.
>>>>
>>>>> Signed-off-by: Jacopo Mondi <jacopo+renesas@xxxxxxxxxx>
>>>>> ---
>>>>>
>>>>> drivers/media/i2c/adv748x/adv748x-core.c | 57 +++++++++++++++++++++++-
>>>>> drivers/media/i2c/adv748x/adv748x.h | 2 +
>>>>> 2 files changed, 58 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/drivers/media/i2c/adv748x/adv748x-core.c
>>>>> b/drivers/media/i2c/adv748x/adv748x-core.c index
>>>>> 200e00f93546..a586bf393558 100644
>>>>> --- a/drivers/media/i2c/adv748x/adv748x-core.c
>>>>> +++ b/drivers/media/i2c/adv748x/adv748x-core.c
>>>>> @@ -335,6 +335,60 @@ int adv748x_tx_power(struct adv748x_csi2 *tx, bool
>>>>> on)
>>>>> /* ----------------------------------------------------------------------
>>>>> * Media Operations
>>>>> */
>>>>> +static int adv748x_link_setup(struct media_entity *entity,
>>>>> + const struct media_pad *local,
>>>>> + const struct media_pad *remote, u32 flags)
>>>>> +{
>>>>> + struct v4l2_subdev *rsd = media_entity_to_v4l2_subdev(remote->entity);
>>>>> + struct v4l2_subdev *sd = media_entity_to_v4l2_subdev(entity);
>>>>> + struct adv748x_state *state = v4l2_get_subdevdata(sd);
>>>>> + struct adv748x_csi2 *tx = adv748x_sd_to_csi2(sd);
>>>>> + bool enable = flags & MEDIA_LNK_FL_ENABLED;
>>>>> + u8 io10;
>>>>> +
>>>>> + /* Refuse to enable multiple links to the same TX at the same time. */
>>>>> + if (enable && tx->src)
>>>>> + return -EINVAL;
>>>>> +
>>>>> + /* Set or clear the source (HDMI or AFE) and the current TX. */
>>>>> + if (rsd == &state->afe.sd)
>>>>> + state->afe.tx = enable ? tx : NULL;
>>>>> + else
>>>>> + state->hdmi.tx = enable ? tx : NULL;
>>>>> +
>>>>> + tx->src = enable ? rsd : NULL;
>>>>> +
>>>>> + if (!enable)
>>>>> + return 0;
>>>>
>>>> Don't we potentially want to take any action on disable to power down
>>>> links below ?
>>>>
>>>>> +
>>>>> + /* Change video stream routing, according to the newly enabled link. */
>>>>> + io10 = io_read(state, ADV748X_IO_10);
>>>>> + if (rsd == &state->afe.sd) {
>>>>> + /*
>>>>> + * Set AFE->TXA routing and power off TXB if AFE goes to TXA.
>>>>> + * if AFE goes to TXB, we need both TXA and TXB powered on.
>>>>> + */
>>>>> + io10 &= ~ADV748X_IO_10_CSI1_EN;
>>>>> + io10 &= ~ADV748X_IO_10_CSI4_IN_SEL_AFE;
>>>>> + if (is_txa(tx))
>>>>> + io10 |= ADV748X_IO_10_CSI4_IN_SEL_AFE;
>>>>
>>>> Shouldn't the CSI4 be enabled here too? or are we assuming it's already
>>>> (/always) enabled?
>>>> io10 |= ADV748X_IO_10_CSI4_EN;
>>>>
>>>>> + else
>>>>> + io10 |= ADV748X_IO_10_CSI4_EN |
>>>>> + ADV748X_IO_10_CSI1_EN;
>>>>> + } else {
>>>>> + /* Clear AFE->TXA routing and power up TXA. */
>>>>> + io10 &= ~ADV748X_IO_10_CSI4_IN_SEL_AFE;
>>>>> + io10 |= ADV748X_IO_10_CSI4_EN;
>>>>
>>>> But if we assume it's already enabled ... do we need this?
>>>> Perhaps it might be better to be explicit on this?
>>>>
>>>>> + }
>>>>> + io_write(state, ADV748X_IO_10, io10);
>>>>
>>>> Would it be any cleaner to use io_clrset() here?
>>>>
>>>> Hrm ... also it feels like this register really should be set depending
>>>> upon the complete state of ... &state->...
>>>>
>>>> So perhaps it deserves it's own function which should be called after
>>>> csi_registered() callback and any link change.
>>>>
>>>> /me has a quick go at some psuedo codeishness...:
>>>>
>>>> int adv74x_io_10(struct adv748x_state *state);
>>>> u8 bits = 0;
>>>> u8 mask = ADV748X_IO_10_CSI1_EN
>>>>
>>>> | ADV748X_IO_10_CSI4_EN
>>>> | ADV748X_IO_10_CSI4_IN_SEL_AFE;
>>>>
>>>> if (state->afe.tx) {
>>>> /* AFE Requires TXA enabled, even when output to TXB */
>>>> bits |= ADV748X_IO_10_CSI4_EN;
>>>>
>>>> if (is_txa(state->afe.tx))
>>>> bits |= ADV748X_IO_10_CSI4_IN_SEL_AFE
>>>> else
>>>> bits |= ADV748X_IO_10_CSI1_EN;
>>>> }
>>>>
>>>> if (state->hdmi.tx) {
>>>> bits |= ADV748X_IO_10_CSI4_EN;
>>>> }
>>>>
>>>> return io_clrset(state, ADV748X_IO_10, mask, bits);
>>>> }
>>>>
>>>> How does that look ? (is it even correct first?)
>>>>
>>>>> +
>>>>> + return 0;
>>>>> +}
>>>>> +
>>>>> +static const struct media_entity_operations adv748x_tx_media_ops = {
>>>>> + .link_setup = adv748x_link_setup,
>>>>> + .link_validate = v4l2_subdev_link_validate,
>>>>> +};
>>>>>
>>>>> static const struct media_entity_operations adv748x_media_ops = {
>>>>> .link_validate = v4l2_subdev_link_validate,
>>>>> @@ -516,7 +570,8 @@ void adv748x_subdev_init(struct v4l2_subdev *sd,
>>>>> struct adv748x_state *state,
>>>>> state->client->addr, ident);
>>>>>
>>>>> sd->entity.function = function;
>>>>> - sd->entity.ops = &adv748x_media_ops;
>>>>> + sd->entity.ops = is_tx(adv748x_sd_to_csi2(sd)) ?
>>>>> + &adv748x_tx_media_ops : &adv748x_media_ops;
>>>>
>>>> Aha - yes that's a neat solution to ensure that only the TX links
>>>> generate link_setup calls :)
>>>
>>> Another option would be to bail out from adv748x_link_setup() if the entity is
>>> not a TX*.
>>>
>>
>> I suggested this in v1 - but Jacopo objected with the following:
>>
>>> Checking for is_txa() and is_txb() would require to call
>>> 'adv_sd_to_csi2(sd)' before having made sure the 'sd' actually
>>> represent a csi2_tx. I would keep it as it is.
>>
>
> That was at the time where the .link_setup() callback was called for
> TXs and non-TXs. What you proposed was to call:
>
> #define adv748x_sd_to_csi2(sd) container_of(sd, struct adv748x_csi2, sd)
>
> on variables that we don't have any guarantee that are of type 'struct
> adv748x_csi2'. I still think it is dangerous and should be avoided and
> I worked it around in v1 as:
>
> + if ((sd != &state->txa.sd && sd != &state->txb.sd) ||
>
>> Now I look at the implementation here, I see this is precisely what it
>> is doing anyway .... still converting through adv748x_sd_to_csi2(sd) on
>> an unknown pointer type
>> (which I still believe is a valid thing to do in this instance)
>
> It's not unknown, .link_setup() is only registered for TXs. If it gets
> called, we know we're dealing with a TX.
>
>>
>> So yes, I think this would be simpler having the check at the top of the
>> adv748x_link_setup() call, and thus then there is no need to add a
>> second adv_media_ops structure.
>
> That was what I did in v1, didn't I ?
>
> The current implementation looks better imho, but if the both of you
> prefer something similar to v1 I will consider that.
Given the extra clarification above, I'll not object to keeping it this
way. I still think it's fine to use container of and then check the
pointers for failure. The is_tx() would perform the type-validation :) -
but lets stick with the one that you prefer. Its your patch :)
--
Kieran
