Re: [RFC][PATCH 1/3] drm/bridge: adv7511: Rework adv7511_power_on/off() so they can be reused internally

[John Stultz]

On Tue, Nov 22, 2016 at 9:38 AM, Laurent Pinchart
<laurent.pinchart@xxxxxxxxxxxxxxxx> wrote:
> Hi John,
>
> On Tuesday 22 Nov 2016 09:25:22 John Stultz wrote:
>> On Tue, Nov 22, 2016 at 12:14 AM, Laurent Pinchart wrote:
>> > On Monday 21 Nov 2016 16:37:30 John Stultz wrote:
>>  @@ -545,24 +554,13 @@ static int adv7511_get_modes(struct adv7511 *adv7511,
>> >>       unsigned int count;
>> >>
>> >>       /* Reading the EDID only works if the device is powered */
>> >>
>> >> -     if (!adv7511->powered) {
>> >> -             regmap_update_bits(adv7511->regmap, ADV7511_REG_POWER,
>> >> -                                ADV7511_POWER_POWER_DOWN, 0);
>> >> -             if (adv7511->i2c_main->irq) {
>> >> -                     regmap_write(adv7511->regmap,
>> >> ADV7511_REG_INT_ENABLE(0),
>> >> -                                  ADV7511_INT0_EDID_READY);
>> >> -                     regmap_write(adv7511->regmap,
>> >> ADV7511_REG_INT_ENABLE(1),
>> >> -                                  ADV7511_INT1_DDC_ERROR);
>> >> -             }
>> >> -             adv7511->current_edid_segment = -1;
>> >> -     }
>> >> +     if (!adv7511->powered)
>> >> +             __adv7511_power_on(adv7511);
>> >
>> > The __adv7511_power_on() function does more than the above, in particular
>> > it performs an expensive regcache_sync() and calls adv7533_dsi_power_on()
>> > for the ADV7533. Don't those operations have side effects that are either
>> > not wanted or not needed here ? In any case this patch modifies the
>> > behaviour of the driver, which needs to be documented in the kernel
>> > message.
>>
>> So yes, while the adv7533 bits aren't needed in the internal function,
>> I'm finding the logic to pulse the HPD and the regcache_sync call seem
>> to be needed side effects, as without that logic, I get i2c_transfer()
>> errors in adv7511_get_edid_block().
>
> Does this patch fix the problem without requiring the 200ms delay ?
>
>> I'll try to rework this patch to split the two changes of reworking
>> the power_on/off function to be re-used (with no logic chage), and the
>> patch to reuse it in get_modes() which resolves a bug.
>
> Have you identified which register write fixes your problem here ?

So I basically used regmap_sync_region() to bisect through the
registers to try to figure out which one calling sync on helps avoid
the issue, and I've narrowed it down to 0x43
(ADV7511_REG_EDID_I2C_ADDR).

If instead of the proposed patch here, I use the following patch
(copy/paste whitespace damage, apologies) seems to make things work
reliably:

diff --git a/drivers/gpu/drm/bridge/adv7511/adv7511_drv.c
b/drivers/gpu/drm/bridge/adv7511/adv7511_drv.c
index 8dba729..87403d7 100644
--- a/drivers/gpu/drm/bridge/adv7511/adv7511_drv.c
+++ b/drivers/gpu/drm/bridge/adv7511/adv7511_drv.c

@@ -555,14 +557,18 @@ static int adv7511_get_modes(struct adv7511 *adv7511,
                                     ADV7511_INT1_DDC_ERROR);
                }
                adv7511->current_edid_segment = -1;
+               regcache_sync_region(adv7511->regmap, 0x43, 0x43);
+
        }

        edid = drm_do_get_edid(connector, adv7511_get_edid_block, adv7511);

-       if (!adv7511->powered)
+       if (!adv7511->powered) {
                regmap_update_bits(adv7511->regmap, ADV7511_REG_POWER,
                                   ADV7511_POWER_POWER_DOWN,
                                   ADV7511_POWER_POWER_DOWN);
+               regcache_mark_dirty(adv7511->regmap);
+       }

        kfree(adv7511->edid);
        adv7511->edid = edid;


But I suspect this isn't a proper fix. I tried adding
ADV7511_REG_EDID_I2C_ADDR to the volatile register list, but that
didn't seem to effect anything (and I still see problematic behavior
if I'm not explictly syncing it as above).

Thoughts on what the right thing is to do?
-john