Am 30.06.2021 um 14:13 schrieb Mark Brown [off-list ref]:

On Tue, Jun 29, 2021 at 10:21:45PM +0200, H. Nikolaus Schaller wrote:

quoted

There seems to be a deadlock in probing:

quoted

	e.g. ldo3_reg depends on vdds_1v8_main supply
	vdds_1v8_main depends on smps7_reg supply
	smps7_reg depends on vsys_cobra supply
	vsys_cobra depends on nothing

quoted

This would normally lead to a simple chain as you described above. But
ldo3_reg and smps7_reg share the same driver and can probe successfully or
fail only in common.

I don't see any deadlock there?  Just a straightforward set of
dependencies.  Anything circular would clearly be a driver bug.

I think it could be indirectly circular since ldo3_reg does not find smps3
registered. But I have to run more tests with printk inserted.

So I am not sure if that is the issue but the best hypothesis I currently know of.

quoted

Now if ldo3_reg defers probe through the new rule, smps7_reg is never
probed successfully because it is the same driver. Hence vdds_1v8_main can
not become available. And the Palmas continues to run in boot initialization
until some driver (MMC) wants to switch voltages or whatever.

The driver should just register all the DCDCs before the LDOs, then
everything will sort itself out.

Basically the driver code does it that way. But fails. Probing seems to defer
until deferral limits (AFAIR there is a timer or counter in the probe deferral
queue) an does not succeed.

It only works if I disable the new condition - but apparently it suffices
to do so for the LDOs because I noticed that the smps initialization also
defines supply_name and there it is no problem to have it before
devm_regulator_register(). Every simple problem becomes complex if you look
deeper :)

Which I interpret (without additional tests) that the smps7_reg is registered
but does not pass the (new) test when asked by ldo3_reg.

 It's *possible* you might see a system
trying to link things together regulators of the same type but that's
very unusual as it makes the system less efficient.

I can check that - and there may be other LDOs in the Palmas
chained (as described by device tree).

It will need some test time to get more factual information about this
issue besides that reverting the new rule or adding a trick to make
the Palmas driver initialize properly again.

BR and thanks,
Nikolaus